AMD 核心软件组织的 GitHub 全景:1000+ 仓库与 ROCm / NPU 技术路线

92 minute read

Published:

AMD 的 GitHub 不是一个组织,而是一张“核心计算 + 芯片平台(含 NPU)+ AI 模型 + 图形生态”的分层地图。本文以 2026-10-27 的 GitHub 公开数据、ROCm 10.1.0 与 Ryzen AI 官方文档为准:先把组织和仓库盘清,再用图拆开 ROCm 与 Ryzen AI NPU 的调用关系、软件栈和技术路线。(由 AI 辅助整理,数据和链接以官方来源为准)

图示已统一画布尺寸;部分复杂图拆成左右/三列/四列面板,便于对照阅读。


0. 统计口径

本文覆盖 AMD 的 9 个主组织:ROCm、amd、AMD-AGI、AMD-Ecosystem、AMDResearch、GPUOpen-LibrariesAndSDKs、GPUOpen-Effects、GPUOpen-Tools、GPUOpen-Drivers。

这里的“全景”指 AMD 核心软件组织的全景,不是所有 AMD GitHub 组织的穷举;Xilinx 等只在 NPU 上游处引用。

共 1035 个公开仓库:596 个活跃非 fork、366 个 fork(含 17 个已归档 fork)、90 个归档仓库。三列相加为 1052,比 1035 多 17,因为归档 fork 被同时计入 fork 和归档。另外附录还收录 9 个外围官方组织(AMD-HPC、AMDESE、pensando、AMDEPYC、AMDComputeLibraries、ROCm-LS、ROCm-DS、amd-enterprise-ai、AMD-AIM),共 144 个仓库;另有 3 个历史组织的 14 个归档仓库。

不包含:

  • AMD-OSX:社区的黑苹果项目,不是 AMD 官方软件栈。
  • AMDlab:名称相近,但内容是与半导体无关的建筑设计工具。
  • 历史组织名 ROCmSoftwarePlatform、RadeonOpenCompute、ROCm-Developer-Tools:共 14 个归档仓库,已迁移到 ROCm,见附录 11.3。
  • AMD-IT、AMD-DC-GPU、AMD-OpenSIL、AMD-AI 等空组织:当前没有公开仓库,暂不展开。
  • Xilinx:AMD 收购后的独立 GitHub 组织;mlir-aie、mlir-air 是 NPU 编译器上游,只在第 8 节引用,不计入 1035。

说明:仓库数量、Star 数和简介会随时间变化;本文是 2026-10-27 的快照。完整清单保留 GitHub 官方英文简介,避免二次翻译失真;核心仓库在正文中给出中文解读。


1. 一张图看组织:AMD 的 GitHub 是“三层结构”

flowchart TB
  AMD["AMD GitHub 生态<br/>1000+ 公开仓库"]

  subgraph CORE["核心 GPU 计算"]
    ROCM["ROCm<br/>414 个仓库"]
  end

  subgraph PLATFORM["芯片、平台与 AI"]
    AMDORG["amd<br/>126 个仓库"]
    AGI["AMD-AGI<br/>72 个仓库"]
    ECO["AMD-Ecosystem<br/>213 个仓库"]
    RES["AMDResearch<br/>23 个仓库"]
  end

  subgraph GRAPHICS["图形、游戏与可视化"]
    LIB["GPUOpen-Libraries<br/>85 个仓库"]
    FX["GPUOpen-Effects<br/>22 个仓库"]
    TOOLS["GPUOpen-Tools<br/>66 个仓库"]
    DRV["GPUOpen-Drivers<br/>14 个仓库"]
  end

  AMD --> CORE
  AMD --> PLATFORM
  AMD --> GRAPHICS

图 1:AMD GitHub 组织地图。 图里是三层:AMD GitHub 生态 → 3 类组织 → 9 个组织;仓库数量标在组织节点上,不再单独画仓库层。ROCm 负责 GPU 通用计算;amd / AMD-AGI / AMD-Ecosystem / AMDResearch 负责芯片平台、AI 模型与生态移植;四个 GPUOpen 组织负责图形、游戏和可视化。

1.1 数字先看

组织官方定位仓库数活跃非 forkfork(含归档)归档
ROCmAMD ROCm 软件栈:编译器、运行时、数学库、通信库、工具41422512481
amdAMD 公司级组织:Ryzen AI、AOCL、EPYC、FPGA、Quark12697263
AMD-AGIAMD AI Group:开源模型、训练框架、推理优化、Agent7262100
AMD-Ecosystem开源生态软件的统一入口:LLM、数据科学、生命科学、仿真、金融213231900
AMDResearchAMD 研究院:NPU 教学、性能工具、HPC 研究原型232120
GPUOpen-LibrariesAndSDKs图形/游戏 SDK 与库858311
GPUOpen-Effects图形特效库:FSR、TressFX、FEMFX 等222110
GPUOpen-Tools图形/计算性能分析与调试工具665880
GPUOpen-Drivers开源图形驱动与组件14645

fork 列含已归档 fork;归档 列含其中 17 个(ROCm 16、GPUOpen-Drivers 1),所以这两行三列相加会多于仓库数。这 9 个主组织在附录中的 fork 标签数为 349。

1.2 三个关键结论

第一,ROCm 正在“收拢”。

2025—2026 年 ROCm 最大的变化不是新增某个库,而是仓库结构重组:

  • 自 ROCm 7.14 起,ROCm 改用 TheRock 构建和发布;
  • 数学/算子库集中到 ROCm/rocm-libraries 超级仓库;
  • 运行时、通信、性能工具等集中到 ROCm/rocm-systems 超级仓库;
  • 旧的单组件仓库大量标记为 [DEPRECATED] Moved to ...,部分已归档。

所以今天看 ROCm,不能只看 ROCm/rocBLAS、ROCm/rccl 这些旧仓库,而要先看 TheRock、rocm-libraries、rocm-systems 三个入口。

第二,AMD 的 GitHub 远不止 ROCm。

  • amd 组织覆盖 Ryzen AI NPU、AOCL CPU 数学库、EPYC 系统管理、FPGA/Vitis;
  • AMD-AGI 是 AMD 的 AI 模型与训练/推理优化组织;
  • AMD-Ecosystem 是 CUDA 项目向 ROCm 移植的“生态池”;
  • NPU 仓库分散在 amd(RyzenAI-SW、xdna-driver、IRON、Triton-XDNA)、ROCm(FastFlowLM、FastFlow-SDK、HRX)和 AMDResearch(Riallto、NPUEval、Ryzers),详见第 8 节;
  • 四个 GPUOpen 组织构成一条完整的 图形/游戏技术栈,与 ROCm 计算栈相对独立。

第三,AMD-Ecosystem 的 213 个仓库里约 190 个是 fork。

这说明它的角色不是“核心 SDK”,而是 生态移植与适配层:把 llama.cpp、SGLang、FlashInfer、LMCache、MONAI、OpenFold、MuJoCo、gsplat 等上游项目移植到 AMD GPU。真正的核心实现仍然在 ROCm、amd、AMD-AGI 三个组织里。


2. 组织逐个看

2.1 ROCm:核心 GPU 计算栈

ROCm 是 AMD 的开放 GPU 计算平台,覆盖编译器、运行时、数学库、通信库、性能工具和调试工具。官方文档对它的定义是:

ROCm is AMD’s open-source GPU computing platform: an end-to-end ecosystem of compilers, runtimes, and libraries for AI, HPC, and domain-specific workloads.

ROCm 组织有 414 个仓库,但其中大量是历史仓库、fork 和镜像。建议按“新入口 / 仍活跃 / 已迁移”三组看:

新入口(先看这三个)

仓库一句话
TheRockROCm 的统一构建与发布系统,也是 HIP/ROCm 的顶层入口;自 ROCm 7.14 起取代旧的单体发布流程。
rocm-libraries数学与算子库超级仓库:CK、hipBLAS/Lt、rocBLAS、MIOpen、rocFFT、rocPRIM、rocWMMA 等都在这里。
rocm-systems系统项目超级仓库:HIP/CLR、ROCr、RCCL、rocSHMEM、rocprofiler、AMD SMI、hipFile 等都在这里。

仍活跃的组件/工具

仓库一句话
hipHIP 的 C++ 运行时 API 与内核语言,CUDA 风格、可移植到 AMD GPU。
HIPIFY把 CUDA 源码翻译成可移植 HIP C++ 的工具(hipify-clang / hipify-perl)。
llvm-projectAMD 维护的 LLVM 分支,包含 amdclang、AMDGPU 后端、COMGR、device-libs。
clrAMD Compute Language Runtime,包含 hipamd、rocclr 和 OpenCL runtime。
aiterAMD 高性能 AI 算子库,Triton/CK/ASM 多后端,是 vLLM/SGLang/ATOM 的默认 kernel 后端。
ATOM基于 AITER 的轻量 vLLM 风格 LLM 推理引擎。
rocSHMEMGPU 发起的 intra-kernel 网络库,OpenSHMEM 风格对称堆。
ROCgdb基于 GDB 的 ROCm 源码级调试器,支持 CPU/GPU 异构调试。
amdsmiAMD SMI 系统管理接口,控制/监控 GPU,已取代 ROCm SMI。
FastFlowLM在 Ryzen AI NPU 上运行 LLM 的优化推理栈。
FastFlow-SDKFastFlowLM + HRX + iron-tools + iron-ops 的 NPU SDK,目标 XDNA2。
hrx-systemHRX:HIP Runtime Extended,加载/调度 NPU 程序。

已迁移/旧仓库(只作参考)

仓库一句话
ROCR-RuntimeAMD 的 HSA 运行时实现;已迁入 rocm-systems。
rocBLASAMD 原生 BLAS 实现;已迁入 rocm-libraries。
hipBLASLtAMD 高性能灵活 GEMM 实现;已迁入 rocm-libraries,不是 CUDA/rocBLASLt 封装。
MIOpenAMD 开源深度学习 primitives 库;已迁入 rocm-libraries。
composable_kernel用 HIP C++ 组合高性能 ML kernel 的编程模型;已迁入 rocm-libraries。
RCCL多 GPU/多节点集合通信库;已迁入 rocm-systems。
rocprofiler-sdkROCm profiling/tracing 基础设施;已归档/迁入 rocm-systems。

“仍活跃”只表示 GitHub 仓库未归档/仍在使用;其中部分组件已在 rocm-libraries / rocm-systems 里有镜像。写新代码优先从新入口开始。

ROCm 仓库状态提示:

  • 看到 [DEPRECATED] Moved to ROCm/rocm-libraries:说明该库已迁入 rocm-libraries;
  • 看到 [DEPRECATED] Moved to ROCm/rocm-systems:说明该库已迁入 rocm-systems;
  • 看到 ARCHIVED:历史仓库,通常只保留只读镜像或旧版发布;
  • 看到 FORK:上游项目的 AMD 分支(如 PyTorch、LLVM、vLLM)。

注意:FastFlowLM / FastFlow-SDK / hrx-system 是 NPU 项目,托管在 ROCm 组织不等于它们属于 ROCm GPU 栈。详见第 8 节。

2.2 amd:芯片平台与 CPU/NPU 软件

amd 组织不是 ROCm 的重复,而是 AMD 公司层面的“芯片平台软件”:

  • Ryzen AI / XDNA NPU:RyzenAI-SW、xdna-driver、IRON、Triton-XDNA、GAIA、Quark;
  • AOCL / Zen CPU 数学库:BLIS、libFLAME、FFTW、LibM、Sparse、Crypto、Compression、ZenDNN;
  • EPYC 系统管理与安全:ESMI、HSMP、APML、SEV、perfspect、node-scraper;
  • FPGA / 嵌入式 / ISP:Vitis-AI、Kria-RoboticsAI、mini-isp、Linux ISP;
  • 视频分析 / VP AI:VAIP、XIR、VART。
仓库一句话
RyzenAI-SWRyzen AI Software 的示例/教程子集;完整运行时和 Vitis AI EP 在官方安装包中。
xdna-driverAMD XDNA NPU 的 Linux 内核驱动(amdxdna.ko)与 XRT SHIM。
IRON面向 AMD NPU 的 close-to-metal 编程库。
Triton-XDNA在 Ryzen AI NPU 上运行 Triton kernel 的后端。
gaia在 Ryzen AI PC 上构建本地 AI agent 的开源项目。
QuarkAMD 深度学习模型量化工具链,支持 PyTorch/ONNX 和多种低精度格式。
ZenDNN面向 Zen CPU 的 DNN 推理加速库。
ZenDNN-pytorch-pluginZenDNN 的 PyTorch 插件(zentorch)。
aoclAMD Optimizing CPU Libraries 元包。
aocl-sparseAMD 优化稀疏线性代数库。
AMD-PACE面向 AMD CPU 的 PyTorch LLM 推理优化扩展。
amd_smi_exporter把 EPYC/GPU 指标导出到 Prometheus。
sev-utilsAMD SEV 内存加密工具。
Vitis-AIVersal AI Edge 的 AI 推理栈。
vaipVitis AI 的 AI 推理运行时(VAIP)。

2.3 AMD-AGI:开源模型、训练与推理优化

AMD-AGI 是 AMD 的 AI Group。它的仓库可以分成四类:

  1. 开源基础模型:Instella、Instella-MoE、Instella-Math、Instella-T2I、InstellaVL、AMD-LLM、Nitro、Hummingbird;
  2. 训练框架:Primus、Primus-Turbo、Primus-SaFE、torchtitan-amd、maxtext-slurm;
  3. 推理/算子优化:Hyperloom、GEAK、Apex、AgentKernelArena、AMDKernelVault、Infera;
  4. 算法研究:PARD、Gumiho、FLy、DUET-VLM、ReasonLite、Tyr-the-Pruner、LoKRA、ALTO。

注意:AMD 官方模型页有些链接仍指向旧组织 AMD-AIG-AIMA,但 GitHub 会把旧仓库地址重定向到 AMD-AGI。可以认为 AMD-AGI 是当前组织名。

仓库一句话
InstellaAMD 全开源语言模型系列,权重、数据、训练代码开放。
Instella-MoE全开源 MoE 语言模型,端到端在 AMD Instinct GPU 上训练。
Instella-MathAMD 全开源推理/数学模型,用长链式思维 RL 训练。
AMD-LLMAMD-135M 语言模型的训练代码与资源。
Nitro-E高效训练的文生图扩散模型系列。
Nitro-T一天内从零训练的文生图扩散模型。
Hummingbird0.9B 高效文生视频扩散模型。
PrimusYAML 驱动的大规模基础模型训练框架,支持 Megatron、TorchTitan、MaxText。
Primus-Turbo面向 AMD GPU 的大规模训练高性能加速库。
Hyperloom自动优化 AMD GPU 上 LLM 推理负载的 agentic 系统。
GEAK多 agent GPU kernel 性能优化器,面向 CDNA。
Apex用 LLM agent 优化 GPU kernel 的 agent/RL 环境与基准。
AgentKernelArena评估和训练 GPU kernel 优化 agent 的平台。
AMDKernelVaultHIP/Triton kernel 生成、优化与 LLM 训练的 playground。

另外还有 AMD-AIM(AMD AI Group, AIM)的 21 个仓库,聚焦 Radeon、Radeon Cloud 和 GPU MODE 参考 kernel,详见附录 11.2。

2.4 AMD-Ecosystem:CUDA → ROCm 的生态移植池

AMD-Ecosystem 的官方定位是:

The unified public home for open-source ecosystem software libraries, covering simulation, finance, LLM extension, life science, data science, and standalone utilities.

它最重要的特征不是“仓库多”,而是 fork 比例极高:213 个仓库中约 190 个是 fork。这说明它是一个 移植/适配层,把上游 CUDA 项目逐个搬到 ROCm/HIP。

仓库一句话
llama.cppAMD 维护的 llama.cpp,ROCm 上的 LLM 推理。
sglangSGLang 的 ROCm 分支,高性能 LLM 服务框架。
flashinferFlashInfer 的 ROCm 移植。
LMCacheLLM KV cache 层,ROCm 版。
verl字节 verl 强化学习框架的 ROCm 版。
gsplat-amd3D Gaussian Splatting 的 AMD 移植。
hipDFGPU DataFrame 库。
hipMMHIP Memory Manager(ROCm-DS)。
hipGRAPH可移植 GPU 图算法库。
hipRaftRAPIDS RAFT 的 HIP 移植。
hipCIMcuCIM 的 HIP 移植。
moat把 CUDA 项目逐个移植到 ROCm/HIP 的项目。
MONAI医学影像 AI 工具包(ROCm 移植)。
OpenFold3AlphaFold3 风格的开源生物分子结构预测(ROCm 移植)。
mujocoMuJoCo 物理仿真(ROCm 移植)。
IsaacLab机器人学习框架(ROCm 移植)。

2.5 AMDResearch:研究与教学原型

AMDResearch 更像“前孵化器”:NPU 教学、GPU 性能工具、HPC 研究原型、AI4Science 配方。它的仓库不一定进入正式 ROCm 发行版,但能提前看到 AMD 在探索什么。

仓库一句话
RyzersRyzen AI 的模块化 Dockerfile 集合,覆盖 LLM/VLM/VLA、机器人、视觉、NPU。
Riallto面向 Ryzen AI NPU 的教学与实验环境。
AUP Learning Cloud定制 JupyterHub 的 AMD AI 学习云。
DAGEECPU/GPU 任务图执行引擎。
IntelliKitagent-first 的 AMD GPU 性能与验证工具集(Kerncap/Metrix/Linex/Nexus)。
NPUEval面向 Ryzen AI NPU 的 LLM 代码生成评测数据集。
intelliperf自动瓶颈检测与优化编排。
omniprobe记录所有 kernel 时长的轻量工具。
AMDGeniePIMPIM 架构 GEMV 分析性能模型。
AMD-DMA-Latte基于 DMA 引擎的 ML 集合通信实现与微基准。

2.6 GPUOpen:图形、游戏与可视化

四个 GPUOpen 组织构成一条独立的图形栈,和 ROCm 计算栈并行:

  • GPUOpen-LibrariesAndSDKs:SDK 与库(FidelityFX SDK、AMF、Vulkan/D3D12 内存分配器、HIP RT、Radeon ProRender);
  • GPUOpen-Effects:特效库(FSR、TressFX、FEMFX、SSSR、CACAO、Denoiser);
  • GPUOpen-Tools:工具(RGP、RGA、RMV、RRA、OCAT、Compressonator、GPU Reshape);
  • GPUOpen-Drivers:开源驱动(AMDVLK、PAL、XGL、LLPC、GPURT)。
仓库一句话
VulkanMemoryAllocator易集成的 Vulkan 内存分配库。
D3D12MemoryAllocator易集成的 Direct3D 12 内存分配库。
FidelityFX-SDKFidelityFX 效果 SDK 主仓库。
AMFAMD 高级媒体框架 SDK,访问硬件多媒体处理。
HIPRTHIP 光线追踪库。
RadeonRays_SDKCPU/GPU 光线求交加速库。
FidelityFX-FSR2FidelityFX Super Resolution 2。
TressFX毛发/毛皮渲染库。
FEMFX有限元软体/形变效果库。
radeon_gpu_profilerRGP,深度 GPU workload 分析。
radeon_gpu_analyzerRGA,离线编译器与代码分析工具。
compressonator纹理/3D 模型压缩、优化与分析工具。
AMDVLKAMD 开源 Vulkan 驱动(Linux)。
pal平台抽象库,封装 AMD 3D/计算驱动的硬件与 OS 细节。
llpcLLVM-Based Pipeline Compiler,把 shader 编译成 PAL 兼容 code object。
gpurtGPU 光线追踪库。

3. ROCm 软件栈:从应用到 GPU

3.1 分层视图

flowchart TB
  A["应用与框架<br/>PyTorch / JAX / TensorFlow<br/>vLLM / SGLang / ATOM<br/>ONNX Runtime / MIGraphX"]
  B["AI 算子与加速库<br/>AITER / CK / hipBLASLt<br/>MIOpen / AOTriton / tritonBLAS"]
  C["ROCm Core SDK<br/>BLAS / FFT / RAND / SPARSE / SOLVER<br/>rocPRIM / rocWMMA / RCCL / rocSHMEM"]
  A --> B --> C
flowchart TB
  D["编程模型与编译器<br/>HIP / HIPIFY / OpenCL / OpenMP<br/>Triton / Gluon / FlyDSL<br/>amdclang / LLVM AMDGPU / COMGR"]
  E["运行时与驱动<br/>CLR:hipamd + rocclr<br/>ROCr(Linux)/ PAL(Windows)<br/>amdkfd + amdgpu"]
  F["硬件<br/>CDNA / RDNA"]
  D --> E --> F

图 2:ROCm 软件栈分层。 左列:应用→算子→Core SDK;右列:编程模型/编译器→运行时/驱动→GPU。

XDNA NPU 不在 ROCm 支持范围;图 2 只画 GPU 栈,NPU 见第 8 节。

3.2 仓库级调用关系

flowchart TB
  APP["应用<br/>PyTorch / vLLM / SGLang / ATOM"] --> AITER["AITER<br/>CK / Triton / ASM / hipBLASLt"]
  AITER --> MATH["数学与 DNN<br/>rocBLAS / rocFFT / MIOpen / rocWMMA"]
flowchart TB
  HIP["HIP Runtime<br/>CLR:hipamd + rocclr<br/>编译:LLVM AMDGPU / COMGR"] --> RT["ROCr(Linux)/ PAL(Windows)<br/>amdkfd + amdgpu"]
  RT --> HW["CDNA / RDNA GPU"]

图 3:ROCm 仓库级调用关系。 左列:应用→AITER→数学库;右列:HIP Runtime→ROCr/PAL→GPU。MIGraphX/ONNX Runtime 走独立图执行路径,未画入这张 AITER 主链。

3.3 编译与运行调用链

flowchart TB
  CUDA["CUDA 源码"] --> IFY["HIPIFY"] --> S["HIP C++ / OpenCL C"]
  S --> F["Clang 前端<br/>amdclang++"] --> O["LLVM AMDGPU<br/>+ ROCm device-libs"]
  TRI["Triton / Gluon / FlyDSL"] --> MLIR["Triton MLIR / Fly MLIR"] --> O
  O --> CM["COMGR<br/>链接 / 检查 code object"] --> C["HSACO code object"]
flowchart TB
  C["HSACO code object"] --> R["HIP Runtime API<br/>CLR: hipamd + rocclr"]
  R --> H["ROCr(HSA)<br/>Linux"]
  R --> P["PAL<br/>Windows"]
  H --> T["libhsakmt / ROCt thunk"] --> K["amdkfd + amdgpu 内核驱动"]
  P --> K
  K --> G["AMD GPU"]

图 4:编译链(左)与运行链(右)。 HIP 源码由 amdclang 编译,经 LLVM AMDGPU 后端生成 HSACO;运行时 CLR 在 Linux 上走 ROCr,在 Windows 上走 PAL,最终通过 amdkfd/amdgpu 驱动提交到 GPU。

几个容易混淆的点:

  • HIP 不是“另一个 CUDA”:HIP 是 C++ 运行时 API + kernel 语言,API 与 CUDA 高度相似,但后端是 AMD 的 CLR/ROCr。
  • CLR 不是一个库:CLR 包含 hipamd、rocclr、opencl 三部分;rocclr 是虚拟设备接口,Linux 对接 ROCr,Windows 对接 PAL。
  • ROCT thunk 已并入 ROCR-Runtime:旧的 ROCT-Thunk-Interface 仓库已标记废弃,libhsakmt/ROCT 的工作并入 ROCR-Runtime。
  • COMGR 是编译器侧的“胶水”:负责 code object 的链接、编译和检查,不是运行时调度器。

3.4 roc* 和 hip* 的区别

官方文档说得很清楚:

  • roc* 前缀:AMD 原生、用 HIP 写的高性能实现,例如 rocBLAS、rocFFT、rocSPARSE、rocSOLVER、rocPRIM;
  • hip* 前缀:通常是可移植的封装层,实现与 CUDA 对应的 API,让 CUDA 应用尽量少改代码就能跑在 AMD GPU 上,例如 hipBLAS、hipFFT、hipRAND、hipSPARSE、hipSOLVER、hipCUB;例外是 hipBLASLt,它本身是 AMD 的高性能灵活 GEMM 实现。

对应关系可以记成:

功能AMD 实现CUDA 兼容封装
BLASrocBLAShipBLAS
灵活 GEMMhipBLASLt(AMD 实现)—
FFTrocFFThipFFT
随机数rocRANDhipRAND
稀疏rocSPARSEhipSPARSE
求解器rocSOLVERhipSOLVER
并行原语rocPRIMhipCUB
并行算法rocThrust—
DNNMIOpenhipDNN(beta)
张量hipTensor—
线程hipThreads—

例外:hipBLASLt 不是 CUDA 封装,而是 AMD 自己的高性能灵活 GEMM 库;hip* 不应一概理解为 wrapper。

选择原则:

  • 新项目、追求峰值性能:优先用 roc* 或 AITER/CK;
  • 从 CUDA 迁移、想少改代码:先用 hip* 封装层;
  • 需要灵活 GEMM / 融合:用 hipBLASLt + CK/AITER;
  • 需要图推理:用 MIGraphX 或 ONNX Runtime。

4. 技术路线:AMD GPU 编程的五条主路

flowchart TB
  subgraph ROUTES["五条主要编程路线"]
    H["HIP C++<br/>原生高性能路线"]
    C["CUDA → HIPIFY → HIP<br/>迁移路线"]
    OCL["OpenCL C<br/>跨厂商标准路线"]
    OMP["OpenMP / Fortran<br/>HPC 路线"]
    TRI["Triton / Gluon / FlyDSL<br/>AI Kernel DSL 路线"]
  end

  H --> LLVM["amdclang / LLVM AMDGPU"]
  C --> LLVM
  OCL --> LLVM
  OMP --> LLVM
  TRI --> MLIR["Triton MLIR / rocMLIR"]
  MLIR --> LLVM
  LLVM --> HSACO["HSACO / code object"]
  HSACO --> RT["CLR / ROCr / PAL"]
  RT --> GPU["AMD GPU"]

  PY["Python / hip-python"] --> H
  FORT["Fortran / hipfort"] --> OMP

图 5:五条编程路线。 所有路线最终都会落到 LLVM AMDGPU 后端、HSACO 和 HIP/ROCr 运行时。

4.1 路线一:CUDA → HIP 迁移

适合:已有 CUDA 代码、想快速跑在 AMD GPU 上。

路径:

CUDA 源码 → HIPIFY(hipify-clang / hipify-perl)→ HIP C++ → amdclang → HSACO → CLR/ROCr → GPU

关键仓库:

  • HIPIFY:CUDA 到 HIP 的源码翻译;
  • hipify_torch:PyTorch 生态的 CUDA→HIP 翻译;
  • hipother:在 NVIDIA 平台上把 HIP API 映射回 CUDA 的兼容层;
  • hipBLAS/hipFFT/hipRAND/hipSPARSE/hipSOLVER/hipCUB:CUDA 对应库的封装。

现实提醒:HIPIFY 不能自动翻译所有 CUDA 代码。没有 HIP 等价库的 CUDA 库、以及针对 NVIDIA 架构做的性能优化,仍然需要人工重写和调优。官方 HIPIFY 文档明确建议:迁移后要做代码审查、替换不支持的库、重新调试和性能优化。

4.2 路线二:原生 HIP 高性能

适合:新写 kernel、追求峰值性能、需要精细控制。

路径:

HIP C++ → amdclang → LLVM AMDGPU → HSACO → CLR/ROCr → GPU

关键仓库:

  • hip:HIP 运行时 API 与 kernel 语言;
  • llvm-project:amdclang、AMDGPU 后端、device-libs;
  • rocPRIM/rocThrust/rocWMMA/hipTensor:并行原语与张量;
  • rocprofiler-sdk/rocprofiler-compute/ROCgdb:性能分析与调试。

优化工具链:

  • rocprofiler-compute:kernel 级指标;
  • rocprofiler-systems:CPU/GPU 系统级 trace;
  • roc-optiq:profiler 可视化;
  • ROCgdb + ROCdbgapi:源码级调试;
  • rocprof-trace-decoder:thread trace 解码。

4.3 路线三:OpenCL 标准路线

适合:已有 OpenCL 代码、需要跨厂商可移植性。

路径:

OpenCL C → clang/SPIR-V → CLR OpenCL runtime → ROCr/PAL → GPU

关键仓库:

  • clr 中的 opencl 实现;
  • clang-ocl(历史仓库);
  • ROCm-OpenCL-Runtime(已归档,旧版);
  • OpenCL-CTS、OpenCL-Headers(Khronos 规范/测试)。

ROCm 10.x 仍提供 amdrocm-opencl 包(OpenCL runtime + ICD loader),但 OpenCL 不是 AMD 当前 AI/HPC 的主推路线。新项目通常优先 HIP 或 Triton。

4.4 路线四:OpenMP / Fortran HPC

适合:Fortran/C/C++ HPC 应用、已有 OpenMP 代码。

路径:

Fortran/C/C++ + OpenMP target → amdclang / flang + LLVM OpenMP → GPU

关键仓库:

  • aomp:AMD 的 OpenMP offload 编译器发行版;
  • flang:Fortran 前端;
  • hipfort:Fortran 到 HIP 的绑定;
  • ROCm-Device-Libs、llvm-project:device 端运行时和编译器。

典型应用:GROMACS、OpenFOAM、SPECFEM3D、rocHPL、rocHPCG。

4.5 路线五:Triton / Gluon / FlyDSL

适合:AI kernel、想用 Python 写高性能 kernel、需要比 HIP 更高的抽象。

路径:

Python Triton / Gluon / FlyDSL → Triton MLIR / Fly MLIR → LLVM AMDGPU → HSACO → GPU

关键仓库:

  • triton:AMD 维护的 Triton 分支;
  • aotriton:Triton 算子的 ahead-of-time 编译与分发,首个 kernel 是 FlashAttention;
  • tritonBLAS:基于 Triton 的分析模型驱动 GEMM 库,不需要 autotune;
  • iris:AMD RAD 的 Triton 多 GPU RMA/SHMEM 框架;
  • FlyDSL:Python 前端 + Fly MLIR dialect,显式表达 layout/tiling/data movement;
  • gfx950-gluon-tutorials:在 MI350/MI355 上用 Gluon 写高性能 GEMM/Attention 的教程。

三条子路线怎么选:

子路线抽象层次适合
Triton高可移植、开发快、大多数 AI kernel
Gluon中低需要显式控制 layout、pipeline、register
FlyDSL低需要 MLIR 级 layout 代数和深度定制

Gluon 的官方教程说得很直白:Triton 强在可移植和生产力,但最后 10%—20% 的性能需要显式控制;Gluon 是 Triton 的 dialect,用 block-level 编程模型补上这一段。


5. AI 推理与训练路线

5.1 AI 推理:vLLM / SGLang / ATOM + AITER

flowchart TB
  MODEL["模型<br/>Llama / DeepSeek / Qwen / Kimi"] --> SERVE["推理框架<br/>vLLM / SGLang / ATOM"]
  SERVE --> AITER["AITER<br/>Attention / MoE / GEMM / Norm"]
  AITER --> BACKENDS["算子后端<br/>CK / Triton / ASM / hipBLASLt"]
  BACKENDS --> HIP["HIP Runtime"]
  LLAMA["llama.cpp"] --> HIP
  HIP --> RT["ROCr / PAL + amdgpu"]
  RT --> GPU["Instinct MI300 / MI350"]
  SERVE --> DIST["分布式<br/>RCCL / MORI / rocSHMEM"]
  DIST --> NIC["RDMA NIC<br/>Pensando AINIC / Thor2 / ConnectX-7"]
  ONNX["ONNX 模型"] --> MIGX["MIGraphX"] --> HIP

图 6:AI 推理路线。 AITER 是 AMD 当前 LLM 推理的核心算子层;vLLM、SGLang、ATOM 都通过 AITER 调用 CK/Triton/ASM/hipBLASLt;llama.cpp 直接走 HIP/rocBLAS,不经过 AITER。

AITER 的定位(官方 README):

  • AMD 高性能 AI 算子库;
  • 多后端:Triton、Composable Kernel、手写 ASM;
  • 覆盖推理和训练,不只是 serving kernel;
  • 框架无关,可集成到 vLLM、SGLang 或自定义框架;
  • 官方称其为 “default kernel backend for LLM inference on AMD GPUs”。

ATOM 是基于 AITER 的轻量 vLLM 风格实现,支持 OpenAI 兼容 API、TP/DP/EP、P/D 分离、KV cache 传输(Mooncake/MORI-IO)、FP8/MXFP4/INT8/INT4 量化、speculative decoding。

MIGraphX 是另一条路线:把 ONNX 模型编译成 AMD GPU 上的图执行引擎,适合非 Transformer 或需要图优化的场景。

5.2 AI 训练:Primus + Megatron/TorchTitan/MaxText

flowchart TB
  S["Primus-SaFE<br/>稳定性 / 故障容错"] -.-> P["Primus<br/>YAML 驱动训练编排"]
  P --> B["训练后端<br/>Megatron-LM / TorchTitan<br/>JAX MaxText / Megatron Bridge"]
  B --> PT["PyTorch / JAX"]
flowchart TB
  PT["PyTorch / JAX"] --> T["Primus-Turbo / AITER / hipBLASLt"]
  T --> C["RCCL / MORI / rocSHMEM"]
  C --> H["HIP Runtime → ROCr → amdgpu"]
  H --> G["MI300X / MI325X / MI355X 集群"]

图 7:训练编排(左)与算子/运行(右)。 Primus 是编排层,下面是 Megatron/TorchTitan/MaxText,再下面是 Primus-Turbo/AITER/hipBLASLt,最后是 RCCL/MORI 和 GPU。

Primus 的官方文档把它定义为“YAML 驱动的、面向 AMD GPU 的大规模基础模型训练框架”,支持:

  • Megatron-LM;
  • TorchTitan;
  • JAX MaxText;
  • Megatron Bridge;
  • HummingbirdXT。

它的三层生态是:

Primus-SaFE(稳定性/集群管理)
        ↓
Primus-LM(训练编排)
        ↓
Primus-Turbo(算子/kernel)

5.3 HPC 路线

flowchart TB
  APP["GROMACS / OpenFOAM / SPECFEM3D / rocHPL"] --> LANG["MPI + OpenMP + Fortran/C++"]
  LANG --> COMP["amdclang / AOMP / Flang / hipfort"]
flowchart TB
  LIB["rocBLAS / rocFFT / rocSPARSE / rocSOLVER"] --> COMM["RCCL / rocSHMEM"]
  COMM --> RT["HIP Runtime → ROCr → amdgpu"]
  RT --> GPU["MI300A / MI355X / MI250X"]

图 8:应用与编译(左)、数学库与运行(右)。 HPC 应用通常走 MPI + OpenMP/Fortran,再调用 rocBLAS/rocFFT/rocSPARSE/rocSOLVER 和 RCCL/rocSHMEM。

HPC 关键仓库:

  • aomp、flang、hipfort:编译与 Fortran 绑定;
  • rocBLAS、rocFFT、rocSPARSE、rocSOLVER、rocALUTION:数学库;
  • RCCL、rocSHMEM:集合通信;
  • rocHPL、rocHPCG、rocHPL-MxP:HPC 基准;
  • Gromacs、OpenFOAM_HMM、SPECFEM3D:应用移植。

5.4 分布式通信路线

flowchart TB
  APP["训练 / 推理框架"] --> CCL["集合通信"]
  CCL --> RCCL["RCCL<br/>AllReduce / AllGather / ReduceScatter"]
  CCL --> SHMEM["rocSHMEM<br/>GPU 发起、对称堆"]
  CCL --> MORI["MORI<br/>MORI-EP / MORI-IO / MORI-CCL / MORI-UMBP"]
  RCCL --> XGMI["xGMI / PCIe"]
  RCCL --> RDMA["RDMA"]
  SHMEM --> RDMA
  MORI --> RDMA
  RDMA --> NIC["Pensando AINIC / Broadcom Thor2 / ConnectX-7"]
  NIC --> CLUSTER["多节点 GPU 集群"]

图 9:分布式通信路线。 节点内走 xGMI/PCIe,节点间走 RDMA;RCCL 是通用集合通信,rocSHMEM 是 GPU 发起的细粒度通信,MORI 面向 MoE EP 和 KV cache 传输。

三条路线的关系:

  • RCCL:最通用的集合通信库,类似 NCCL 的 AMD 对应物;
  • rocSHMEM:OpenSHMEM 风格,适合 GPU 内核内部直接做远程内存访问;
  • MORI:模块化 RDMA 框架,MORI-EP 做 MoE dispatch/combine,MORI-IO 做 KV cache 点对点传输,MORI-CCL 做定制集合通信。

MORI 官方 README 明确支持 AMD Pensando DSC、Broadcom Thor2、NVIDIA Mellanox ConnectX-7 等 NIC,并已被 SGLang、vLLM、DeepSpeed 等集成。


6. 图形与光追路线

flowchart TB
  APP["游戏 / 渲染 / 内容创作"] --> API["Vulkan / DX12 / OpenGL"]
  APP --> FX["FidelityFX SDK"] --> API
  API --> AMDVLK["AMDVLK"] --> XGL["XGL"]
flowchart TB
  XGL["XGL"] --> PAL["PAL / LLPC"] --> KMD["amdgpu 内核驱动"] --> GPU["Radeon GPU"]
  HIPRT["HIP RT"] --> HIP["HIP / ROCm"] --> KMD

图 10:图形 API/驱动前端(左)与平台/内核/光追(右)。 图形栈走 Vulkan/DX12/OpenGL → AMDVLK/XGL/LLPC/PAL;光追可以走 FidelityFX/HIP RT;HIP RT 把光追带到 HIP/ROCm。

AMDVLK 的架构(官方 README):

  • AMDVLK:开源 Vulkan 驱动;
  • PAL:平台抽象库,封装硬件和 OS 细节;
  • XGL:Vulkan API 层;
  • LLPC:LLVM-Based Pipeline Compiler,把 shader 编译成 PAL 兼容 code object;
  • GPURT:GPU 光线追踪库;
  • LLVM:shader 编译基础设施。

说明:pal、xgl、llpc、gpurt 仓库目前为归档状态,但 AMDVLK 的架构仍由这些组件构成;阅读源码时以 AMDVLK README 引用的组件关系为准。

GPUOpen 工具链:

  • RGP:Radeon GPU Profiler,深度分析 GPU workload;
  • RGA:Radeon GPU Analyzer,离线编译和代码分析;
  • RMV:Radeon Memory Visualizer,显存使用分析;
  • RRA:Radeon Raytracing Analyzer,BVH 可视化;
  • OCAT:FPS overlay 与性能测量;
  • Compressonator:纹理/模型压缩;
  • GPU Reshape:指令级插桩与验证。

7. 硬件路线:CDNA、RDNA、XDNA

flowchart TB
  MI100["MI100<br/>CDNA1 gfx908"] --> MI250["MI250X<br/>CDNA2 gfx90a"]
  MI250 --> MI300["MI300X/A<br/>CDNA3 gfx942"]
  MI300 --> MI355["MI355X<br/>CDNA4 gfx950"]
flowchart TB
  RDNA3["RX 7000<br/>RDNA3 gfx110x"] --> RDNA35["Ryzen AI Max<br/>RDNA3.5 gfx115x"]
  RDNA35 --> RDNA4["RX 9000 / R9700<br/>RDNA4 gfx120x"]
flowchart TB
  XDNA["XDNA / XDNA2<br/>AIE2 / AIE2P"] --> IRON["IRON / Triton-XDNA"]

图 11:Instinct(左)、Radeon/Ryzen(中)、Ryzen AI NPU(右)。 CDNA 面向数据中心 Instinct,RDNA 面向 Radeon/Ryzen,XDNA 面向 Ryzen AI NPU;前两者走 ROCm,XDNA 走 Ryzen AI Software。

7.1 关键规格

产品架构LLVM target显存CU典型用途
MI355XCDNA4gfx950288 GiB HBM3e256大模型训练/推理
MI350XCDNA4gfx950288 GiB HBM3e256大模型训练/推理
MI350PCDNA4gfx950144 GiB128推理/训练
MI325XCDNA3gfx942256 GiB HBM3e304大模型推理/训练
MI300XCDNA3gfx942192 GiB HBM3304大模型推理/训练
MI300ACDNA3gfx942128 GiB HBM3228APU,CPU+GPU 统一内存
MI250XCDNA2gfx90a128 GiB HBM2e220HPC/AI
MI210CDNA2gfx90a64 GiB HBM2e104HPC/AI
MI100CDNA1gfx90832 GiB HBM2120早期 CDNA
RX 9070 XTRDNA4gfx120116 GiB64消费级/工作站
RX 7900 XTXRDNA3gfx110024 GiB96消费级/工作站
Ryzen AI Max+ 395RDNA3.5gfx1151共享内存—APU/iGPU

数据来自 ROCm 10.1.0 官方 GPU specifications 与兼容性矩阵;不同 SKU 的显存/规格可能不同。

NPU(XDNA1/XDNA2)的规格和软件栈见第 8 节,不在 ROCm 兼容性矩阵里。

7.2 ROCm 10.1 的支持范围

官方兼容性矩阵列出的 LLVM target 包括:

  • CDNA4:gfx950(MI355X、MI350X、MI350P)
  • CDNA3:gfx942(MI325X、MI300X、MI300A)
  • CDNA2:gfx90a(MI250X、MI250、MI210)
  • CDNA1:gfx908(MI100)
  • RDNA4:gfx1200、gfx1201(RX 9060/9070、Radeon AI PRO R9700)
  • RDNA3:gfx1100、gfx1101、gfx1102、gfx1103
  • RDNA3.5:gfx1150、gfx1151、gfx1152、gfx1153(Ryzen AI 300/400、Ryzen AI Max)

XDNA NPU 不在 ROCm 的支持列表里。Ryzen AI NPU 走的是第 8 节的独立软件栈(Ryzen AI Software、IRON、Triton-XDNA、FastFlow-SDK)。ROCm 支持的是 Ryzen APU 的 iGPU(RDNA3.5),不是 NPU。


8. NPU:Ryzen AI 的软件栈与路线

这一节讲的是 Ryzen AI PC 里的 XDNA NPU,不是 ROCm 的 GPU,也不是 Versal AI Edge 的 Vitis AI。

8.1 硬件:XDNA1 → XDNA2

AMD 的 NPU 不是 GPU 的“小核”,而是一块独立的空间数据流(spatial dataflow)AI Engine 阵列。每个 AI Engine tile 包含:

  • VLIW/SIMD 向量处理器;
  • RISC 标量处理器;
  • 本地程序存储器和数据存储器;
  • 专用互连与 DMA。

它不靠大缓存,而是靠 tile 本地内存和显式数据流,所以功耗低、时延确定。

代际NPU 架构代表产品AIE 阵列NPU TOPS
XDNA1AIE2Ryzen 7040(Phoenix)、Ryzen 8040(Hawk Point)20 tiles,4×510 / 16
XDNA2AIE2PRyzen AI 300(Strix Point)、Ryzen AI Max(Strix Halo)、Ryzen AI 400(Gorgon Point)、Kraken32 tiles,4×850 / 55 / 60
XDNA2 嵌入式AIE2PRyzen AI Embedded P100 / X100、P121iXDNA2最高 50;P121i 30

XDNA2 相对 XDNA1 的关键变化:每 tile 2× MAC、1.6× on-chip memory、block floating point,AI Engine 性能约 3×,NPU 达到 50 TOPS。不同 SKU 的 TOPS 不同:Ryzen AI 300 / Max 为 50,Ryzen AI PRO 300 为 55,Ryzen AI 400 / PRO 400 最高 60。

数字来自 AMD XDNA 架构页、IRON ISCA 2025 教程、Ryzen AI 400 CES 2026 发布页和 Ryzen AI Embedded 发布页。XDNA2 的 32 tiles/4×8 指计算 tile;不同 SKU/工具显示的拓扑可能不同。

7040/8040 的 10/16 TOPS、PRO 300 的 55 TOPS、P121i 的 30 TOPS 分别来自 AMD 对应产品页/发布资料,链接见第 12 节。

flowchart TB
  P["Phoenix / Hawk Point<br/>Ryzen 7040 / 8040"] --> X1["XDNA1<br/>AIE2 · 20 tiles · 4×5<br/>10 / 16 TOPS"]
flowchart TB
  S["Strix Point / Strix Halo<br/>Ryzen AI 300 / Max"] --> X2["XDNA2<br/>AIE2P · 32 tiles · 4×8<br/>50 TOPS"]
  K["Kraken / Gorgon<br/>Ryzen AI 400"] --> X2
  X2 --> H["最高 60 TOPS<br/>Ryzen AI 400 / PRO 400"]
flowchart TB
  E["P100 / X100"] --> EX["XDNA2<br/>最高 50 TOPS"]
  P2["P121i"] --> EX2["XDNA2<br/>30 TOPS"]

图 12:XDNA1(左)、XDNA2(中)、嵌入式 XDNA2(右)。 XDNA1 是 20 tiles/4×5、10/16 TOPS;XDNA2 是 32 tiles/4×8、50–60 TOPS;嵌入式 P100/X100 最高 50 TOPS,P121i 为 30 TOPS。

8.2 软件栈:两条栈、五层

Ryzen AI NPU 的软件栈可以按“部署栈”和“编程栈”来理解。两条栈共用驱动和硬件,但入口、编译器和运行时不同。

层部署栈(ONNX)编程栈(IRON / Triton / FastFlow)
应用/模型PyTorch / TensorFlow → ONNX;CNN/Transformer;LLMIRON Python、Triton kernel、FastFlowLM
框架/APIONNX Runtime + Vitis AI EP;OGA/llama.cpp;LemonadeIRON API、Triton、FastFlowLM
量化/编译Quark(INT8/BF16);Vitis AI EP 图切分/编译IRON ops、MLIR-AIE / MLIR-AIR;Peano → PDI/xclbin
运行时XRT + XRT SHIM(Linux)/ Windows NPU 驱动栈HRX(FastFlow-SDK)或 XRT(IRON/Triton,Linux)
驱动/硬件amdxdna(Linux)/ Windows NPU 驱动 → XDNA1/XDNA2同上

OGA = ONNX Runtime GenAI;Ryzen AI 文档里 llama.cpp 只支持 iGPU,NPU-only 的 LLM 路线用 OGA/FastFlowLM。Lemonade 是多厂商的本地 LLM SDK,FastFlowLM 已集成到 Lemonade。

flowchart TB
  A1["应用/模型<br/>ONNX:PyTorch / TF → ONNX<br/>IRON:Python / Triton / FastFlowLM"] --> A2["框架/API<br/>ONNX Runtime + Vitis AI EP<br/>IRON API / Triton / FastFlowLM"]
flowchart TB
  B1["量化/编译<br/>Quark / Vitis AI EP<br/>MLIR-AIE / MLIR-AIR + Peano"] --> B2["编译产物<br/>量化模型 / 编译图<br/>PDI / xclbin"]
flowchart TB
  C1["运行时<br/>XRT + XRT SHIM<br/>HRX"] --> C2["驱动/硬件<br/>amdxdna / Windows NPU 驱动<br/>XDNA1 / XDNA2"]

图 13:NPU 软件栈的三段压缩视图。 左:应用/模型 + 框架/API;中:量化/编译 + 产物;右:运行时 + 驱动/硬件。完整五层见上表。

8.3 四条路线

把上面的栈拆开,实际有四个入口:

  1. ONNX / Vitis AI EP:最像“部署”的路线,适合 CNN/Transformer;
  2. IRON:close-to-metal,适合写自定义算子/kernel;
  3. Triton-XDNA:把 Triton kernel 编译到 NPU,适合 Triton 用户/研究;
  4. FastFlowLM / FastFlow-SDK:NPU-first LLM 运行时,适合开箱即用跑 LLM/VLM/ASR。

另外,Vitis-AI / VAIP / XIR / VART 是 Versal AI Edge 的独立路线,不是 Ryzen AI NPU,见 8.6。

flowchart TB
  A1["PyTorch / TF"] --> A2["ONNX"] --> A3["Quark"] --> A4["ORT + VAI-EP"] --> A5["XRT / SHIM"] --> A6["NPU"]
flowchart TB
  B1["IRON"] --> B2["MLIR-AIE"] --> B3["Peano → PDI"] --> B4["HRX / XRT"] --> B5["NPU"]
flowchart TB
  C1["Triton"] --> C2["triton-shared"] --> C3["MLIR Transform"] --> C4["MLIR-AIR/AIE"] --> C5["XRT binary"] --> C6["NPU"]
flowchart TB
  D1["Lemonade / GAIA"] --> D2["FastFlowLM"] --> D3["IRON ops"] --> D4["HRX"] --> D5["NPU"]

图 14:四条 NPU 路线(从左到右)。 A:ONNX/Vitis AI EP;B:IRON;C:Triton-XDNA;D:FastFlowLM/FastFlow-SDK。

路线入口编译/运行时适合成熟度
A. ONNX / Vitis AI EPPyTorch/TF → ONNXQuark → ONNX Runtime + Vitis AI EP → XRT/XRT SHIMCNN/Transformer 部署,改动最小官方主推;Windows/Linux
B. IRONPython IRONMLIR-AIE + Peano → PDI → HRX/XRT自定义算子/kernel、数据流官方开源;XDNA1/2
C. Triton-XDNATriton kerneltriton-shared → MLIR Transform → MLIR-AIR/AIE → XRT binaryTriton 用户/研究实验性;AIE2/AIE2P
D. FastFlowLMLemonade/GAIA/CLIFastFlowLM → IRON ops → HRX本地 LLM/VLM/ASR,开箱即用早期访问;XDNA2
  • A 是 Ryzen AI Software 的主推部署路线:Vitis AI EP 会切分图,把能跑 NPU 的部分交给 NPU;Windows 上不支持的算子可回退 CPU/iGPU,Linux 当前是 NPU-only。Linux 的运行时栈是 XRT + XRT SHIM,Windows 走 NPU 驱动栈。
  • B 的 IRON 本身不提供 kernel,kernel 来自 MLIR-AIE 的 aie_kernels;@iron.jit 把 Python 定义的工作负载编译成 NPU 程序,ObjectFifo 负责声明式数据搬运。
  • C 目前是实验项目,官方 README 说 matmul 可达到手写 NPU 实现的性能,90% 测试配置达到 baseline 的 90% 以上;支持 matmul、elementwise、softmax、layer norm。
  • D 的 FastFlow-SDK 是早期访问,包含 FastFlowLM + HRX + iron-tools + iron-ops,目标 XDNA2;FastFlowLM 支持 256k context,已被 Lemonade 集成。

8.4 NPU 仓库地图

NPU 相关仓库不在一个组织里,而是分散在 amd、ROCm、AMDResearch,以及上游 Xilinx。

核心 NPU 仓库

组织仓库角色一句话
amdRyzenAI-SW示例/教程Ryzen AI Software 的示例与教程子集;完整运行时和 Vitis AI EP 在官方安装包中。
amdxdna-driverLinux 驱动amdxdna.ko + XRT SHIM + 固件打包;需要 Linux 6.10+。
amdIRON编程 APIclose-to-metal Python API,基于 MLIR-AIE 和 Peano。
amdTriton-XDNA编译器后端Triton → MLIR-AIR/AIE 的实验后端,产出 XRT binary。
amdQuark量化PyTorch/ONNX 量化工具链,支持 INT8/BF16 等低精度。
amdgaiaAgent/应用本地 AI agent;NPU 后端走 FastFlowLM。
amdonnxruntime框架 forkAMD 的 ONNX Runtime 分支;Ryzen AI Software 通过 ONNX Runtime + Vitis AI EP 部署模型。
ROCmFastFlowLMLLM 运行时NPU-first LLM/VLM/ASR/Embedding 推理,支持 256k context。
ROCmFastFlow-SDKSDKFastFlowLM + HRX + iron-tools + iron-ops;XDNA2 早期访问。
ROCmhrx-system运行时HRX:HIP Runtime Extended,加载/调度 NPU 程序。
AMDResearchRiallto教学Ryzen AI NPU 的教学与实验环境。
AMDResearchNPUEval评测面向 AIE kernel 代码生成的 LLM 评测集。
AMDResearchRyzers环境Ryzen AI Dockerfiles,覆盖 LLM/机器人/视觉/NPU。

相邻/上游仓库

组织仓库角色一句话
amdVitis-AIVersalVersal AI Edge 的推理栈,不是 Ryzen AI NPU。
amdvaip / xir / vartVersalVitis AI 的运行时、图 IR 和 runtime 组件。
amdLIRA归档Ryzen AI 300 NPU 上的 ASR 示例,已归档。
ROCmhrx-solutions / hrx-loom-library示例/工具HRX 方案与 kernel 语料。
Xilinxmlir-aie上游编译器MLIR-AIE dialect / AIE kernels;iron-tools 的前身。
Xilinxmlir-air上游编译器MLIR-AIR;Triton-XDNA 的 lowering 目标。

RyzenAI-SW 是示例/教程仓库,不是完整安装包;完整 Ryzen AI Software 需要从 AMD 下载页安装。ROCm/FastFlowLM、FastFlow-SDK、HRX 是 NPU 项目,只是托管在 ROCm 组织,不代表 ROCm 支持 XDNA NPU。Xilinx/mlir-aie、mlir-air 不在本文 1035 个仓库的统计内,但它们是 IRON/Triton-XDNA 的上游。

8.5 与 ROCm 的边界

维度ROCm GPURyzen AI NPU
硬件CDNA / RDNAXDNA1 / XDNA2
计算单元CUAIE tiles
编译 targetgfx908 / gfx90a / gfx942 / gfx950 / gfx110x / gfx120xaie2 / aie2p
编程HIP / OpenCL / OpenMP / TritonONNX + Vitis AI EP / IRON / Triton-XDNA
运行时CLR / ROCr / PALXRT / XRT SHIM / HRX
驱动amdgpu + amdkfdamdxdna(Linux)/ Windows NPU 驱动
发行ROCm Core SDK / TheRockRyzen AI Software / FastFlow-SDK
典型硬件MI300X / MI355X / RX 7900Ryzen AI 300 / 400 / Max
组织ROCmamd / ROCm / AMDResearch

一句话:ROCm 管 GPU,Ryzen AI Software 管 NPU。Ryzen AI APU 上两者都有:RDNA3.5 iGPU 走 ROCm,XDNA2 NPU 走 Ryzen AI Software / IRON / FastFlow。

8.6 不要混淆:Vitis AI EP、Vitis AI 和 Versal

  • Vitis AI EP:ONNX Runtime 的 Execution Provider,跑在 Ryzen AI NPU 上;Ryzen AI Software 用它做图切分/编译。
  • amd/Vitis-AI、vaip、xir、vart:Vitis AI 的 Versal AI Edge 推理栈,目标 FPGA/自适应 SoC,不是 Ryzen AI NPU。典型路径是 ONNX → XIR → VAIP → VART → Versal。
  • XDNA NPU vs RDNA iGPU:Ryzen AI APU 同时有两者;ROCm 支持 iGPU,Ryzen AI Software/IRON 支持 NPU。

8.7 一句话总结

Ryzen AI NPU 不是 ROCm 的附属品;它有自己的驱动(amdxdna)、运行时(XRT / HRX)、编译器(Vitis AI EP / MLIR-AIE + Peano / MLIR-AIR)和 LLM 运行时(FastFlowLM)。按目标选:部署走 ONNX/Vitis AI EP,写 kernel 走 IRON,Triton 用户走 Triton-XDNA,跑 LLM 走 FastFlowLM。


9. 仓库重组:为什么旧仓库都写着 “Moved”

flowchart TB
  L["旧版:单仓库 / 单组件发布<br/>ROCm/ROCm、ROCm/MIOpen、ROCm/rocBLAS ..."] --> S["超级仓库<br/>rocm-libraries / rocm-systems"]
  L -.-> A["归档 / 只读镜像"]
flowchart TB
  S["超级仓库"] --> T["TheRock<br/>统一构建、测试、发布"] --> P["ROCm Core SDK 10.x<br/>DEB/RPM / Tarball / Wheel / Runfile"] --> E["ROCm Extras / Standalone ONNX"]

图 15:旧版单仓库(左)与超级仓库/TheRock 发布(右)。 从“每个组件一个仓库”转向“超级仓库 + TheRock 统一发布”。

9.1 三个新入口

TheRock:

  • 官方定义:The HIP Environment and ROCm Kit;
  • 角色:ROCm 的顶层构建与发布系统;
  • 自 ROCm 7.14 起,ROCm 通过 TheRock 构建和发布;
  • 提供 nightly ROCm、PyTorch、JAX 构建;
  • 支持 Linux 多发行版和原生 Windows;
  • 构建从源码需要约 200 GB 磁盘空间。

rocm-libraries:

  • 数学/算子库超级仓库;
  • projects/ 下包括:composablekernel、hipblas、hipblas-common、hipblaslt、hipcub、hipdnn、hipfft、hiprand、hipsolver、hipsparse、hipsparselt、hiptensor、miopen、rocblas、rocfft、rocprim、rocrand、rocsolver、rocsparse、rocthrust、rocwmma、rpp;
  • shared/ 下包括:rocroller、tensile、mxdatagenerator、origami。

rocm-systems:

  • 系统项目超级仓库;
  • projects/ 下包括:amdsmi、aqlprofile、clr、hip、hipfile、hipother、hip-tests、rccl、rdc、rocdbgapi、rocdecode、rocjpeg、rocm-core、rocminfo、rocmsmilib、rocprofiler、rocprofiler-compute、rocprofiler-register、rocprofiler-sdk、rocprofiler-systems、rocr-debug-agent、rocrruntime、rocshmem、roctracer。

9.2 包结构变化

ROCm 10.x 的包名从 rocm-* / roc* / hip* 统一改为 amdrocm-*:

新包内容
amdrocm-llvmamdclang++、hipcc、flang
amdrocm-runtimeHIP、ROCR、runtime compilation、COMGR
amdrocm-blasrocBLAS、hipBLAS、hipBLASLt、hipSPARSELt
amdrocm-dnnhipDNN、MIOpen
amdrocm-fftrocFFT、hipFFT、hipFFTW
amdrocm-sparserocSPARSE、hipSPARSE
amdrocm-solverrocSOLVER、hipSOLVER、rocALUTION
amdrocm-profilerrocprofiler-systems、rocprofiler-compute、rocprofiler-sdk、roctracer
amdrocm-debuggerrocgdb、ROCdbgapi、ROCR Debug Agent
amdrocm-openclOpenCL runtime 和 ICD loader
amdrocm-rcclRCCL
amdrocm-rocshmemrocSHMEM
amdrocm-hipfilehipFile
amdrocm-ckComposable Kernel

安装目录也从 /opt/rocm-7.2/ 改为 /opt/rocm/core-10.1,并通过 update-alternatives 提供 /opt/rocm 兼容符号链接。

9.3 对开发者的实际影响

  1. 不要再从旧仓库起步:新项目看 rocm-libraries / rocm-systems / TheRock;
  2. 旧博客/脚本可能失效:包名、安装路径、组件归属都变了;
  3. 用 TheRock 构建更可控:适合需要最新功能或自定义组合的开发者;
  4. 发行版安装优先用 amdrocm-* 元包:避免单独安装组件导致依赖冲突。

10. 怎么选:按目标找路线

你的目标首选路线关键仓库
跑 PyTorch 训练/推理PyTorch ROCm wheels 或 TheRockpytorch、TheRock、rocm-libraries
部署 LLM 推理vLLM / SGLang / ATOM + AITERvllm、sglang、ATOM、aiter
做 LLM 训练Primus + Megatron/TorchTitan/MaxTextPrimus、Primus-Turbo、Megatron-LM
写高性能 kernelHIP 或 Triton/Gluon/FlyDSLhip、triton、aotriton、FlyDSL
从 CUDA 迁移HIPIFY → HIPHIPIFY、rocm-libraries(hipBLAS/hipCUB)
HPC / FortranAOMP/hipfort + rocm-librariesaomp、hipfort、rocm-libraries
多机多卡RCCL / rocSHMEM / MORIrocm-systems(RCCL)、rocSHMEM、mori
图推理 / ONNXMIGraphX / ONNX RuntimeAMDMIGraphX、onnxruntime、hip-ep
图形 / 光追GPUOpen / AMDVLK / HIP RTAMDVLK、FidelityFX-SDK、HIPRT
Ryzen AI NPUONNX/Vitis AI EP、IRON、Triton-XDNA、FastFlowLMRyzenAI-SW、xdna-driver、IRON、FastFlow-SDK

标 rocm-libraries / rocm-systems 的条目表示组件已迁入超级仓库,旧单仓库只作参考。

10.1 三条最常用的组合

组合 A:LLM 推理

模型 → vLLM/SGLang/ATOM → AITER → CK/Triton/ASM/hipBLASLt → HIP → ROCr → MI300X/MI355X

组合 B:LLM 训练

Primus → Megatron/TorchTitan/MaxText → PyTorch/JAX → RCCL/MORI → HIP → MI300X/MI355X 集群

组合 C:HPC

GROMACS/OpenFOAM → MPI + OpenMP/Fortran → AOMP/hipfort → rocBLAS/rocFFT/rocSPARSE → RCCL → HIP → MI250X/MI300A

11. 附录:本文统计范围内的全部仓库清单

以下清单按组织分组,按 Star 数从高到低排列。状态 列含义:活跃 = GitHub 上非 fork、非归档;fork = 上游项目的 AMD 分支;归档 = 只读历史仓库。简介保留 GitHub 官方英文原文。

活跃 只表示 GitHub 状态,不代表项目仍推荐使用;简介含 [DEPRECATED] Moved to ... 的仓库,请以 rocm-libraries / rocm-systems 为准。

fork 列含已归档的 fork;附录状态列只标一个主状态,所以 9 个主组织在附录中的 fork 标签数为 349,小于正文的 fork 总数 366。

说明:ROCm/legacy-rocm-build 是旧 ROCm/ROCm 主页仓库的归档名;今天 ROCm 的入口是 ROCm/TheRock。表中的 — 表示 GitHub 仓库未填写简介,这类仓库多为内部工具、第三方库镜像或历史占位仓库。

11.1 九个主组织

ROCm:414 个仓库(ROCm 核心 GPU 计算栈)

活跃非 fork 225 · fork 124 · 归档 81 · 按 Star 数排序

仓库Star状态官方简介
legacy-rocm-build6,814归档AMD ROCm™ Software - GitHub Home
hip4,421活跃HIP: C++ Heterogeneous-Compute Interface for Portability
FastFlowLM1,995活跃Run LLMs on AMD Ryzen™ AI NPUs in minutes; purpose-built and deeply optimized for the AMD NPUs.
TheRock1,408活跃The HIP Environment and ROCm Kit - A lightweight open source build system for HIP and ROCm
MIOpen1,194活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
HIPIFY734活跃HIPIFY: Convert CUDA to Portable C++ Code
tensorflow-upstream707forkTensorFlow ROCm port
aiter598活跃AI Tensor Engine for ROCm
composable_kernel548活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo. NOTE: develop branch is maintained as a read-only mirror
rocm-systems543活跃super repo for rocm systems projects
ROCm-docker521活跃Dockerfiles for the various software layers defined in the ROCm software platform
amdgpu470活跃This repository is a fork of the upstream AMDGPU Driver, used by the ROCm project. To submit a patch, please follow the guidelines at https://docs.kernel.org/process/submitting-patches.html#submittingpatches . For rep…
rocm-libraries456活跃super repo for rocm libraries
hcc444归档HCC is an Open Source, Optimizing C++ Compiler for Heterogeneous Compute currently for the ROCm GPU Computing Platform
rccl420活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
rocBLAS403活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
k8s-device-plugin399活跃Kubernetes (k8s) device plugin to enable registration of AMD GPU to a container cluster
omnitrace358归档Omnitrace: Application Profiling, Tracing, and Analysis
AMDMIGraphX337活跃AMD's graph optimization engine.
rocm-examples328活跃A collection of examples for the ROCm software stack
FlyDSL304活跃FlyDSL is the Python front‑end of the project: a Flexible Layout Python DSL for expressing tiling, partitioning, data movement, and kernel structure at a high level.
ROCR-Runtime278活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
Tensile262活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
pytorch261forkTensors and Dynamic neural networks in Python with strong GPU acceleration. To use PyTorch with latest ROCm releases, see: https://rocm.docs.amd.com/en/latest/rocm-for-ai/pytorch.html
aomp246活跃AOMP is an open source Clang/LLVM based compiler with added support for the OpenMP® API on Radeon™ GPUs. Use this repository for releases, issues, documentation, packaging, and examples.
llvm-project244forkThis is the AMD-maintained fork of the LLVM git repository. This repository accepts pull requests and issues related to AMD fork-specific topics (amd/*). For all other issues/PRs, please submit upstream at https://git…
flash-attention241forkFast and memory-efficient exact attention
MIVisionX218活跃AMD MIVisionX is a computer vision toolkit built around a highly optimized, conformant open-source implementation of the Khronos OpenVX™ 1.3.2 specification. As of the 4.0.0 release, MIVisionX ships three components: …
HIP-Examples214归档Examples for HIP
iris203活跃AMD RAD's multi-GPU Triton-based framework for seamless multi-GPU programming
rocFFT203活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
ATOM201活跃AiTer Optimized Model
rocMLIR195活跃—
mori188活跃Modular RDMA Interface
clr185活跃—
ROC-smi183活跃ROC System Management Interface
rocPRIM177活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
ROCm-OpenCL-Runtime169归档ROCm OpenOpenCL Runtime
gpufort168归档GPUFORT: S2S translation tool for CUDA Fortran and Fortran+X in the spirit of hipify
rocprofiler-compute166归档[DEPRECATED] Moved to ROCm/rocm-systems repo
rocm_smi_lib155活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
hipBLAS154活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
rocprofiler153归档[DEPRECATED] Moved to ROCm/rocm-systems repo
triton146forkDevelopment repository for the Triton language and compiler
amd_matrix_instruction_calculator145活跃A tool for generating information about the matrix multiplication instructions in AMD Radeon™ and AMD Instinct™ accelerators
rocSHMEM145活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
rocWMMA143活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
librocdxg142活跃repo for AMD ROCDXG project
amdsmi136活跃AMD SMI
rocSPARSE136活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
spur136活跃spur your compute
HIP-CPU133归档An implementation of HIP that works on CPUs, across OSes.
rocRAND130活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
hipCaffe125归档(Deprecated) hipCaffe: the HIP port of Caffe
vllm124forkA high-throughput and memory-efficient inference and serving engine for LLMs
rocThrust123活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
rocSOLVER118活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
hipBLASLt116活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
LLVM-AMDGPU-Assembler-Extra113归档LLVM AMDGPU Assembler Helper Tools
ROCmValidationSuite110活跃A system validation and diagnostics tool for monitoring, stress testing, detecting, and troubleshooting issues impacting AMD GPUs in high-performance computing environments
gpu-operator108活跃—
gpuaidev106活跃Repository to host ROCm Developer Hub Notebook Tutorials
rocALUTION104活跃Next generation library for iterative sparse solvers for ROCm platform
aotriton102活跃Ahead of Time (AOT) Triton Math Library
hipfort97活跃Fortran interfaces for ROCm libraries
ROCm-Device-Libs97归档ROCm Device Libraries
rccl-tests94活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
ROCT-Thunk-Interface93活跃ROCm's Thunk Interface
rocm_bandwidth_test90活跃Bandwidth test for ROCm
rocm-blogs85活跃—
hipCUB83活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
roctracer83活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
ROCgdb80活跃This is ROCgdb, the ROCm source-level debugger for Linux, based on GDB, the GNU source-level debugger.
TransferBench80活跃TransferBench is a utility capable of benchmarking simultaneous copies between user-specified devices (CPUs/GPUs)
rocprof-compute-viewer79活跃—
rocm-cmake78活跃CMake modules used within the ROCm libraries
TransformerEngine77活跃—
rocHPL75活跃High Performance Linpack for Next-Generation AMD HPC Accelerators
rpp74活跃AMD ROCm Performance Primitives (RPP) library is a comprehensive high-performance computer vision library for AMD processors with HIP/OpenCL/CPU back-ends.
FastFlowLM-internal72归档FLM app mirror
rocm-xio72活跃A ROCm library for GPU-Initiated IO. This provides support for initiating IO from a ROCm-capable GPU against a range of targets including RDMA NICs, NVMe SSDs and AMD GPU DMA engines (SDMAs).
bitsandbytes71fork8-bit CUDA functions for PyTorch
tritonBLAS69活跃A lightweight triton-based General Matrix Multiplication (GEMM) library.
atmi68归档Asynchronous Task and Memory Interface, or ATMI, is a runtime framework and programming model for heterogeneous CPU-GPU systems. It provides a consistent, declarative API to create task graphs on CPUs and GPUs (integr…
device-metrics-exporter68活跃Device Metrics Exporter exports metrics from AMD devices (GPUs) to collectors like Prometheus.
hipSPARSE68活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
hipFFT65活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
hrx-system61活跃HRX: Hip Runtime Extended
MIOpenGEMM61归档MIOpenGEMM is now deprecated
hipDNN57活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
gfx950-gluon-tutorials55活跃A practical guide to high-performance gluon kernel development on AMD GFX9 GPUs.
hipThreads51活跃—
hipTensor50活跃AMD’s C++ library for accelerating tensor primitives
rocm-cli49活跃ROCm CLI is a command-line tool for setting up and running local AI on AMD GPUs, with a full-screen TUI dashboard for GPU telemetry, model serving, and chat.
container-toolkit47活跃Offers tools that streamline the use of AMD GPUs with containers.
FLM_Q4NX_Converter46活跃A converter for transferring gguf Q4_0, Q4_1 to FLM Q4NX
ROCm-CompilerSupport45归档The compiler support repository provides various Lightning Compiler related services.
MAD44活跃MAD (Model Automation and Dashboarding)
Megatron-LM44forkOngoing research training transformer models at scale
ROCclr43归档—
rocminfo43活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
hip-python42活跃HIP Python Low-level Bindings
hip-tests41活跃—
HIPCC41归档HIPCC: HIP compiler driver
rocHPCG41活跃HPCG benchmark based on ROCm platform
ROCm-ComputeABI-Doc41归档ROCm - AMDGPU Compute Application Binary Interface
rocm-install-on-linux40活跃—
rocm-install-on-windows38活跃—
Experimental_ROC36归档Experimental and Intriguing Tools for ROCm
MISA36活跃Machine Intelligence Shader Autogen. AMDGPU ML shader code generator. (previously iGEMMgen)
rocDecode35活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
rocr_debug_agent35活跃The ROCdebug-agent is a library that can be loaded by ROCm Platform Runtime to provide some debugging functionality.
hipamd34归档—
xformers34forkHackable and optimized Transformers building blocks, supporting a composable construction.
roc-optiq33活跃A visualizer for the ROCm Profiler Tools
half32活跃—
jax32forkComposable transformations of Python+NumPy programs: differentiate, vectorize, JIT to GPU/TPU, and more
OpenFOAM_HMM32活跃Refactoring OpenFOAM with OpenMP target offloading and use of HMM to offload work onto GPUs
ROCm.github.io32归档ROCm Website
rocprofiler-sdk31归档[DEPRECATED] Moved to ROCm/rocm-systems repo
k8s-gpu-dra-driver30活跃K8s DRA driver for AMD GPUs
libhipcxx30活跃The C++ Standard Library for your entire system.
rdc30活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
rocmProfileData30活跃—
rocprofiler-systems30归档[DEPRECATED] Moved to ROCm/rocm-systems repo
mxnet29forkLightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more
rocm-ernic29活跃A vfio-user based emulated AMD Pensando RDMA NIC.
rocm-jax29活跃—
hipify_torch28活跃—
Thrust28归档HIP back-end for Thrust that has been replaced by rocThrust
clang-ocl27归档OpenCL compilation with clang compiler.
hipRAND27活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
hipSOLVER27活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
omnistat27活跃Scale-out system monitoring
hipFile26活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
rocAL26活跃The AMD rocAL is designed to efficiently decode and process images and videos from a variety of storage formats and modify them through a processing graph programmable by the user.
pyrsmi25活跃python package of rocm-smi-lib
apex24forkA PyTorch Extension: Tools for easy mixed precision and distributed training in Pytorch
ROCm-OpenCL-Driver24归档ROCm OpenCL Compiler Tool Driver
clARMOR23归档OpenCL tool to detect buffer overflows in GPU kernels
pytorch-micro-benchmarking23活跃—
numba-hip22归档HIP backend patch for Numba, the NumPy aware dynamic Python compiler using LLVM.
StatusMonitor22归档OpenCL multiGPU sample monitoring system health
torch_migraphx22活跃Torch-MIGraphX integrates AMD's graph inference engine with the PyTorch ecosystem.
fleet-chiplet-megakernel20活跃—
roc-stdpar20归档—
ROCdbgapi20活跃The AMD Debugger API is a library that provides all the support necessary for a debugger and other tools to perform low level control of the execution and inspection of execution state of AMD's commercially available …
nnvm-rocm19归档NNVM for ROCm Examples
rocm-aic19活跃ROCm AMD Infinity Context - A distributed, shared, low-latency KV Cache solution for AI Inference..
rocprof-trace-decoder19活跃—
aws-ofi-rccl18活跃—
cvs18活跃—
hip-ep18活跃ROCm Execution Provider for ONNX Runtime on AMD GPUs. Compiles ONNX models through MLIR to LLVM bitcode, JIT-loads them at session creation, and dispatches to hipDNN, MIOpen, hipBLASLt, and custom HIP kernels
tensorcast18活跃—
rocBLAS-Examples17归档Examples illustrating usage of the rocBLAS library
rocm-docs-core17活跃ROCm Documentation Python package for ReadTheDocs build standardization
DeepEP15活跃—
hipCOMP-core15活跃hipCOMP is a library for fast lossless compression/decompression on the GPU. This repository contains the algorithms.
hipObject15活跃A ROCm library for RDMA-based object storage with Direct-to-VRAM DMA support on AMD GPUs.
RIXL15活跃DEPRECATED REPOSITORY. ROCm Inference Transfer Library (RIXL) is a port of the NIXL library for AMD GPUs. See README_rocm.md for AMD specific instructions for compiling RIXL
AMD_Hackathon14活跃AMD_Robotics_Hackathon_2025
FastFlow-SDK14活跃FastFlow-SDK is a software stack for running and developing applications on AMD XDNA Neural Processing Engines (NPUs).
DirectGMA_CL13归档Simple example showing how to use DGMA in OpenCL
HCC-Example-Application13归档HCC Sample Applications
hipSPARSELt13活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
AITemplate12forkAITemplate is a Python framework which renders neural network into high performance CUDA/HIP C++ code. Specialized for FP16 TensorCore (NVIDIA GPU) and MatrixCore (AMD GPU) inference.
hcc-clang-upgrade12归档stage the upgrade of hcc-clang to clang ToT
hipCollections12活跃Header-only library of GPU-accelerated, concurrent data structures.
llvm12归档 / fork—
xla12forkA machine learning compiler for GPUs, CPUs, and ML accelerators
aorta11活跃—
gpu-cluster-networking11活跃Cluster networking documentation for AMD Instinct accelerators
hipBench11活跃HIP Kernel Benchmarking Library
rocm-amdgpu-bench11归档—
hipDF10归档hipDF - GPU DataFrame Library
madengine10活跃madengine is a streamlined CLI tool for running and benchmarking AI models on ROCm GPUs, offering a production‑ready workflow for local single node or remote multi node execution with integrated performance monitoring.
onnxruntime10forkONNX Runtime: cross-platform, high performance scoring engine for ML models
rocm-core10活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
fio9forkFlexible I/O Tester
Paddle9forkPArallel Distributed Deep LEarning
rocmlirTriton9活跃MLIR kernel generator for AMD GPUs (GEMM, conv, attention, fused), built on a Triton compilation backend.
criu8forkCheckpoint/Restore tool
formal-models8活跃—
hipother8活跃—
instinct-docs8活跃AMD Instinct Data Center GPU docs
MITuna8活跃—
nccl-rccl-parser8forkTool to run rccl-tests/nccl-tests based on from an application
network-operator8活跃Network Operator simplifies usage of AMD AINICs in a Kubernetes Environment
OSU_Microbenchmarks8归档ROCm - UCX enabled OSU_Benchmarks
rbuild8活跃Rocm build tool
rocJPEG8活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
rocm-kpack8归档Prototype for a kernel packing library
rocPyDecode8活跃rocPyDecode is a set of Python bindings to rocDecode C++ library which provides full HW acceleration for video decoding on AMD GPUs.
spur-cloud8活跃GPU as a Service platform built on Spur
amdgpu-docs7活跃AMD GPU Driver Documentation
axis7活跃AXIS: Agent eXecution Isolation Substrate — secure sandbox runtime for AI agents
cupy7forkA NumPy-compatible array library accelerated by CUDA
DeepSpeed7forkDeepSpeed is a deep learning optimization library that makes distributed training easy, efficient, and effective.
Gromacs7活跃ROCm's implementation of Gromacs
rockrel7活跃—
rtg_tracer7活跃—
timemory7活跃Modular C++ Toolkit for Performance Analysis and Logging. Profiling API and Tools for C, C++, CUDA, Fortran, and Python. The C++ template API is essentially a framework to creating tools: it is designed to provide a u…
amd-nhc6活跃—
AMDiffusionBenchmark6活跃Benchmarking training throughput of contemporary diffusion models on AMD's GPU hardware.
CTranslate26forkFast inference engine for Transformer models
embers6活跃—
flang6活跃Mirror of flang repo: The source repo is https://github.com/flang-compiler/flang . Once a day the master branch is updated from the upstream source repo and then locked. AOMP or ROCm developers may commit or create P…
hipBLAS-common6活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
jitify6活跃A single-header C++ library for simplifying the use of HIP Runtime Compilation (HIPRTC).
repo-digest6活跃Automated commit digests using Claude Code agents
rocHPL-MxP6活跃—
rocm-llvm-python6归档Low-level Cython and Python bindings to the (ROCm) LLVM and AMD COMGR C API. Also ships the official LLVM Clang bindings.
rocm-spack-pkgs6归档Repository to host spack recipes for ROCm
transformers6fork🤗Transformers: State-of-the-art Natural Language Processing for Pytorch and TensorFlow 2.0.
ucx6forkUnified Communication Library X
clang5归档 / fork—
DeepBench5forkBenchmarking Deep Learning operations on different hardware
gpu-agent5活跃GPU Agent provides programmable APIs to configure & monitor AMD Instinct GPUs
hip-testsuite5活跃—
hipMM5归档HIP Memory Manager (ROCm-DS)
nvtop5forkGPUs process monitoring for AMD, Intel and NVIDIA
rccl-rdma-sharp-plugins5活跃—
rdma-perftest5活跃—
rocm-docs5活跃Source for the ROCm documentation website (rocm.docs.amd.com)
rocRoller5活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
rocsift5活跃C99 AMDGPU Debug Library wrapping KFD sysfs, debugfs, and more.
SPIRV-LLVM-Translator5活跃This is the AMD-maintained fork of the SPIRV-LLVM_Translator git repository. This repository accepts issues related to AMD fork-specific topics. For all other issues/PRs, please submit upstream at https://github.com/K…
amd-anp4活跃AINIC Network Plugin
aomp-extras4活跃hostcall services library, math library, and utilities
aqlprofile4活跃[DEPRECATED] Moved to ROCm/rocm-systems repo
builder4forkContinuous builder and binary build scripts for pytorch
cm4活跃Frontend for configuring/building/testing CMake projects
dask-cuda-rocm4forkUtilities for Dask and CUDA interactions
DeepLearningExamples4forkState-of-the-Art Deep Learning scripts organized by models - easy to train and deploy with reproducible accuracy and performance on enterprise-grade infrastructure.
horovod4forkDistributed training framework for TensorFlow, Keras, and PyTorch.
jax-aiter4活跃—
mblas-bench4活跃Basic BLAS benchmark for GEMMs
MIFin4活跃Tuna centric MIOpen client
mxDataGenerator4活跃[DEPRECATED] Moved to ROCm/rocm-libraries repo
rocCV4活跃rocCV is an efficient GPU-accelerated library for image pre- and post-processing, powered by AMD's HIP platform
rocm-spack4forkA flexible package manager that supports multiple versions, configurations, platforms, and compilers.
rocm-tests4活跃A pytest based system test framework that validates the full AMD ROCm stack across diverse environments.
rocprofiler-systems-skills4活跃A collection of skills used by the rocprof-sys team
AMD-ROCm-User-Experience-3归档AMD ROCm User Experience, including Training, Support, and Technical Documentation.
amdgpu-windows-interop3归档Support files for interfacing ROCm to the AMDGPU Windows Driver
bert3forkTensorFlow code and pre-trained models for BERT
hip-fpsan3活跃—
hrx-solutions3活跃Solutions built on the HRX system
maxtext3forkA simple, performant and scalable Jax LLM!
miopen_cudnn_ops3归档 / fork—
ompi3forkCopy of the Open MPI repository
rocjitsu-test-corpus3活跃—
rocm-papi-component3归档PAPI integration in ROCm profiling and tracking tools
rocm-timeline-generator3fork—
ROCmDS-cmake3归档This is a collection of CMake modules that are useful for all ROCm-DS projects. By sharing the code in a single place it makes rolling out CMake fixes easier.
ucc3forkUnified Communication Collectives Library
whisper3forkRobust Speech Recognition via Large-Scale Weak Supervision
clang-tools-extra2归档 / forkMirror of official clang-tools-extra git repository located at http://llvm.org/git/clang-tools-extra. Updated every five minutes.
cluster-networking2活跃—
detectron22forkDetectron2 is FAIR's next-generation research platform for object detection and segmentation.
device-config-manager2活跃Device Config Manager allows configuring AMD devices (GPUs for now) using simple config JSON files
distributed_inference2活跃InferenceX Distributed Inference MI355X Recipe
eigen2归档This is the ROCmSoftwarePlatform fork of the Eigen library: https://gitlab.com/libeigen/eigen
FAMBench2forkBenchmarks to capture important workloads.
gputt2forkgpuTT: GPU Tensor Transpose library
hrx-loom-library2活跃Example kernel corpus for the HRX loom compiler
lld2归档 / fork—
mesa-fork2活跃A fork of https://gitlab.freedesktop.org/mesa/mesa - Mesa 3D graphics library
MIOpenExamples2归档MIOpen examples
mshadow2forkMatrix Shadow:Lightweight CPU/GPU Matrix and Tensor Template Library in C++/CUDA for (Deep) Machine Learning
pytorch-lightning2forkThe lightweight PyTorch wrapper for high-performance AI research. Scale your models, not the boilerplate.
rocGemmDriver2活跃rocGemmDriver
rocm-installer-runfile2活跃Code for the ROCm runfile installer which is one method users can install ROCm and the AMDGPU driver on Linux
rocprofiler-register2归档[DEPRECATED] Moved to ROCm/rocm-systems repo
spur-toolkit2活跃Tools to deploy and debug Spur.
tinygrad2forkYou like pytorch? You like micrograd? You love tinygrad! ❤️
triton-inference-server-onnxruntime_backend-legacy2活跃—
triton-inference-server-vllm_backend-legacy2活跃—
vision2forkDatasets, Transforms and Models specific to Computer Vision
.github1活跃AMD ROCm™ Platform - GitHub Home
AFLplusplus1forkThe fuzzer afl++ is afl with community patches, qemu 5.1 upgrade, collision-free coverage, enhanced laf-intel & redqueen, AFLfast++ power schedules, MOpt mutators, unicorn_mode, and a lot more!
amd-rocm-optimization-guide1活跃Performance optimization techniques and tutorials for AMD GPUs using HIP and ROCm
amd-rocm-programming-guide1活跃AMD ROCm Programming Guide
benchmarks1forkBenchmark code
CHROMA1归档ROCm implementation of CHROMA
CNTK-11forkMicrosoft Cognitive Toolkit (CNTK), an open source deep-learning toolkit
compiler-rt1归档 / forkMirror of official compiler-rt git repository located at http://llvm.org/git/compiler-rt. Updated every five minutes.
cublasgemm-benchmark1归档 / forkcode for benchmarking GPU performance based on cublasSgemm and cublasHgemm
DeepSpeed-MII1forkMII makes low-latency and high-throughput inference possible, powered by DeepSpeed.
DeepSpeedExamples1forkExample models using DeepSpeed
dlrm1forkAn implementation of a deep learning recommendation model (DLRM)
dyninst1forkDyninstAPI: Tools for binary instrumentation, analysis, and modification.
fastai1forkThe fastai deep learning library, plus lessons and tutorials
FBGEMM1forkFB (Facebook) + GEMM (General Matrix-Matrix Multiplication) - https://code.fb.com/ml-applications/fbgemm/
firdev-llvm-project1归档THIS REPOSITORY IS DEPRECATED AND WILL BE DELETED SOON. The new repo is f18-llvm-project
fuzz-fill1活跃Fuzzing to fill test suite gaps
ggml-staging-automation1活跃Release management repo for AMD staging ggml project branches (prior to upstreaming).
hip-mpi-testsuite1活跃—
hip-rdma-perftest1归档Verbs/RDMA benchmarks with ROCM support
hip-verbs-acceptance1活跃Acceptance suite for the HIP runtime and RDMA Verbs: functional and performance tests for GPU, NIC and GPU+NIC paths, gated by per-platform policies.
hipOMB1归档OSU MPI benchmarks with ROCm support
InferenceMAX1forkOpen Source Continuous Inference Benchmarking - GB200 NVL72 vs MI355X vs B200 vs H200 vs MI325X & soon™ TPUv6e/v7/Trainium2/3/GB300 NVL72 - DeepSeek 670B MoE, GPTOSS
IRFuzzer1归档 / fork—
k8s-network-device-plugin1活跃Network device plugin for Kubernetes
k8s-network-node-labeller1活跃Discovers AMD AINIC's characteristics and labels Kubernetes nodes with the corresponding information
kineto1forkA CPU+GPU Profiling library that provides access to timeline traces and hardware performance counters.
migraphx-benchmark1活跃—
miopen-benchmark1归档 / forkbenchmarking miopen
models1forkModels and examples built with TensorFlow
mosaicml_benchmarks_internal1forkFast and flexible reference benchmarks
nixl1forkAMD maintained fork of the NIXL repository
og1归档OpenMP GCC support
OpenCL-CTS1forkThe OpenCL Conformance Tests
pyro1forkDeep universal probabilistic programming with Python and PyTorch
pytest-abort1活跃Pytest plugin + helpers for attributing hard crashes (SIGSEGV/SIGABRT) to the last running test and keeping HTML reports mergeable.
pytorch-examples1forkA set of examples around pytorch in Vision, Text, Reinforcement Learning, etc.
Quartz1活跃Central CI/CD Data Hub for the ROCm Ecosystem
rocgputreeshap1活跃ROCm support for GPUTreeShap
rocm-blogs-sphinx1活跃—
rocm-recipes1活跃Recipes for rocm
SPECFEM3D1归档ROCm implementation of SPECFEM3D
superbenchmark1forkProvide hardware and software benchmarks for AI systems
system-acceptance-docs1活跃—
taichi_examples1活跃—
therock-ci-config1活跃Shared CI runner and infrastructure config for ROCm projects
triton-inference-server-server-legacy1活跃—
aomp-fork-loop0归档 / forkAOMP is an open source Clang/LLVM based compiler with added support for the OpenMP® API on Radeon™ GPUs. Use this repository for releases, issues, documentation, packaging, and examples,.
audio0forkData manipulation and transformation for audio signal processing, powered by PyTorch
Benchmark0归档 / fork—
blis0forkBLAS-like Library Instantiation Software Framework
broadcom-roce-exporter0活跃—
ByteMLPerf0forkAI Accelerator Benchmark focuses on evaluating AI Accelerators from a practical production perspective, including the ease of use and versatility of software and hardware.
camunda-jenkins-shared-library0forkCamunda community Jenkins Shared Library
ck-jit0活跃—
ClassyVision0forkAn end-to-end PyTorch framework for image and video classification
ClientEnablement0活跃—
common-infra-operator0活跃Shared CI/CD workflows, pytest test framework, and utilities for ROCm GPU Operator components
D108850-lld-bug-reproduction0归档Reproduction wrapper for the bug with LLVM D108850
dask0forkParallel computing with task scheduling
dask-hip0活跃Utilities for Dask and HIP interactions
distributed0归档 / forkA distributed task scheduler for Dask
dnn-benchmarking0活跃Benchmarking and validation tool for hipDNN graphs.
f18-llvm-project0归档Mirror of flang-compiler repo: The source repo is https://github.com/flang-compiler/f18-llvm-project . Once a day the main and fir-dev branches are updated from the upstream source repo and then locked. ROCm developer…
faiss0forkA library for efficient similarity search and clustering of dense vectors.
fastai20forkTemporary home for fastai v2 while it's being developed
fastbook0forkThe fastai book, published as Jupyter Notebooks
frugally-deep0forkHeader-only library for using Keras (TensorFlow) models in C++.
FunctionalPlus0forkFunctional Programming Library for C++. Write concise and readable C++ code.
gloo0forkCollective communications library with various primitives for multi-machine training.
GOTCHA0forkGOTCHA is a library for wrapping function calls in shared libraries
grover0归档 / forkCode for Defending Against Neural Fake News, https://rowanzellers.com/grover/
hsa-class0归档—
instinct-ai-examples0活跃—
jax-lab0活跃A JAX benchmarking framework for reproducible and comparable evaluation of configurable workloads on AMD GPUs.
jax-triton0forkDevelopment fork of `jax-triton`: integrations between JAX and OpenAI Triton
jax-triton.old0活跃ROCm fork of jax-triton
jaxpp0活跃—
kernel-module-management0活跃—
kernl0forkKernl lets you run PyTorch transformer models several times faster on GPU with a single line of code, and is designed to be easily hackable.
libfabric0forkOpen Fabric Interfaces
libflame0forkHigh-performance object-based library for DLA computations
llama-cpp-temp-fork0forkThis is a temporary fork of llama.cpp to support AMD development efforts while we are migrating GitHub organizations. Final location: https://github.com/AMD-Ecosystem/llama.cpp
migraphx_userguide0归档—
miopenDockerBuilder0归档Scripts and
MONAI-Core0归档 / forkAI Toolkit for Healthcare Imaging
multi-gpu-programming-models0forkExamples demonstrating available options to program multiple GPUs in a single node or a cluster
onnxruntime-training-examples0forkExamples for using ONNX Runtime for model training.
OpenCL-Headers0forkKhronos OpenCL-Headers
OpenLLM0forkOperating LLMs in production
openwebtext0forkAn open clone of the GPT-2 WebText dataset by OpenAI. Still WIP.
pollara-roce-exporter0活跃—
pytorch.github.io0forkThe website for PyTorch
pytorch_scatter0forkPyTorch Extension Library of Optimized Scatter Operations
QoLA0活跃Quality-of-Life AITER -- a wrapper library that sits atop AITER to provide a disentangled build solution for AITER consumers.
recommenders-addons0forkAdditional utils and helpers to extend TensorFlow when build recommendation systems, contributed and maintained by SIG Recommenders.
res_embed0forkEmbed resources into CMake targets in a cross-platform way
rocALmodels0归档 / forkTensorFlow models repository with RALI support
rocComposer0归档AMD composer for High Performance Deep Learning Kernels and Libraries
ROCgdb-docs0活跃This repository contains the documentation for AMD ROCm Debugger (ROCgdb) on Linux, AMD’s source-level debugger based on the GNU Debugger (GDB). ROCgdb enables heterogeneous debugging across ROCm software stacks that …
rocGemmDriver-10归档 / forkrocGemmDriver
rockbuilder0活跃—
rocLibUtils0活跃Shared code for the math and communication libraries
rocm-cmake-build-tools0活跃A collection of CMake modules/functions used internally within the ROCm libraries
rocm-docs-style0活跃—
rocm-fortran-generator0活跃Generates the Fortran bindings of HIP, ROCTx and the ROCm math libraries from their C headers
rocm-framework-tests0活跃—
rocm-handbook0活跃ROCm handbooks target developers who require a unified, book-style reference for ROCm and HIP
rocm-security-gh0活跃This Repo provides Security baseline files, scan tools, branch protection rules etc., to use by the ROCm org repositories
rocm-spack-packages0forkSpack's community package recipes
rocprofiler-systems-report0归档Workspace for generating reports assisting maintenance and engineering oversight
sccache0forksccache with ROCm and CUDA compilation (NVCC) support
shap0forkA game theoretic approach to explain the output of any machine learning model.
SPIRV-Headers0forkSPIRV-Headers
sukha-tools0forkProfiling tools intended for Nvidia or AMD GPUs
tensorflow-addons0forkUseful extra functionality for TensorFlow 2.x maintained by SIG-addons
tensorflow-build0forkBuild-related tools for TensorFlow
tensorflow-runtime0forkA performant and modular runtime for TensorFlow
tensorflow-serving0活跃TensorFlow Serving dockerfiles for ROCM
tensorflow_fashionmnist0fork—
tensorpipe0forkA tensor-aware point-to-point communication primitive for machine learning
text-generation-inference0forkLarge Language Model Text Generation Inference
TheRockMigrationTest0forkThe HIP Environment and ROCm Kit - A lightweight open source build system for HIP and ROCm
tiny-rocm-cxx0归档—
torchrec0forkPytorch domain library for recommendation systems
torchtitan0forkA PyTorch native platform for training generative AI models
translate0forkTranslate - a PyTorch Language Library
triton-inference-server-backend-legacy0活跃—
triton-inference-server-core-legacy0活跃—
triton-inference-server-pytorch_backend-legacy0归档—
triton-inference-server-third_party-legacy0活跃—
triton-kernels0fork—
ucx-py-rocm0forkPython bindings for UCX
utk-magma0fork—
vcpkg0forkC++ Library Manager for Windows, Linux, and MacOS
amd:126 个仓库(AMD 芯片平台、Ryzen AI、AOCL、EPYC、FPGA)

活跃非 fork 97 · fork 26 · 归档 3 · 按 Star 数排序

仓库Star状态官方简介
gaia1,598活跃Build AI agents for your PC
RyzenAI-SW894活跃AMD Ryzen™ AI Software includes the tools and runtime libraries for optimizing and deploying AI inference on AMD Ryzen™ AI powered PCs.
xdna-driver623活跃AMD XDNA Driver (amdxdna.ko) for Linux and XRT SHIM library
OpenCL-caffe529活跃This is a Experimental version of OpenCL by AMD Research, we now recommend you to use The official BVLC Caffe OpenCL branch is over at Caffe branch now at https://github.com/BVLC/caffe/tree/opencl
skills466活跃Official AMD catalog of AI agent skills. Empower your AI agents with AMD's optimized SW stack.
HPCTrainingExamples220活跃—
MxGPU-Virtualization210活跃—
Quark186活跃—
blis171forkBLAS-like Library Instantiation Software Framework
ZenDNN160活跃—
IRON151活跃Close-to-metal programming for AMD NPUs
mini-isp146活跃A minimal, open-source Image Signal Processor (ISP) for AMD FPGA, implemented in Verilog.
aocl-libm-ose136活跃AOCL-LibM
playbooks119活跃Official repository of AMD Playbooks
amd-lab-notes118活跃AMD lab notes with code examples to demonstrate use of AMD GPUs
AMD-ASPFW114活跃—
ryzen-ai-documentation108活跃Onboarding documentation source for the AMD Ryzen™ AI Software Platform. The AMD Ryzen™ AI Software Platform enables developers to take pretrained machine learning models in popular frameworks and run them on laptop…
Chromium-WebCL95活跃WebCL implementation for Chromium
furious.js82活跃scientific computing package for JavaScript - inspired by NumPy
amd_smi_exporter67活跃The AMD SMI Exporter exports AMD EPYC CPU & Datacenter GPU metrics to the Prometheus server.
firmware_binaries65活跃—
amd-fftw64活跃FFTW code optimized for AMD based processors
UIF62活跃—
Triton-XDNA58活跃—
amd_energy55活跃—
fuzzyHSA53活跃—
esmi_ib_library51活跃E-SMI: EPYC™ System management Interface In-band Library
libflame50forkHigh-performance object-based library for DLA computations
aocl-compression45活跃A software library of lossless data compression methods tuned and optimized for AMD “Zen”-based CPUs
apml_modules44活跃AMD APML modules, apml_sbtsi and apml_sbrmi registers to hwmon framework provding temperature and power metrics of the socket. They are extended by registering a misc_device, which provides ioctl interface to BMC admins.
amd_hsmp43活跃AMD HSMP module to provide user interface to system management features.
InfinityHub-CI39活跃—
aocl-libmem38活跃A library of AMD optimized memory and string functions
aocl-sparse38活跃AMD optimized Sparse Linear Algebra library
HPCTrainingDock37活跃—
ZenDNN-pytorch-plugin36活跃—
aocl-crypto33活跃—
aocl30活跃—
LIRA29归档This tool helps you easily deploy ASR models on NPUs on AMD's Ryzen AI 300 series laptops
esmi_oob_library28活跃for hosting E-SMI Out-of-band code
win-libm27归档Core Math functions for MS Windows
sd-sandbox26活跃Open source Stable Diffusion examples and workflows for Ryzen AI software
virtio-npu23活跃—
openmm-hip21活跃—
aocl-utils20活跃AOCL-Utils library to get CPU architecture, Cache information and CPU features flags etc.
Kria-RoboticsAI20活跃—
AMD-PACE16活跃AMD PACE is a PyTorch extension for high-performance LLM inference on AMD CPUs. It provides a framework to develop and test novel optimizations, accelerating real-time deployment.
go_amd_smi16活跃—
whisper.cpp16forkPort of OpenAI's Whisper model in C/C++
Vitis-AI15活跃AMD’s stack for AI inference on Versal AI Edge and Versal AI Edge Gen 2. Host tools, examples, and reference designs organized per device series.
ZenDNN-onnxruntime14活跃—
aocl-fftz12活跃AMD's High-performance Fast Fourier Transform (FFT) library optimized for Zen-based processors
node-scraper12活跃Automated data collection and analysis for the purposes of system debug.
sev-utils12活跃Utilities and tools for AMD SEV memory encryption technologies
ama-sdk11活跃—
aocl-dlp11活跃A library of AMD Optimized Deep Learning Primitives
ZenDNN-tensorflow-plugin11活跃—
ADA10活跃AMD Deployment Assistant
AOCL.jl10活跃—
Open-Field-Health-Check10活跃—
scalapack10活跃DEPRECATED. This Scalapck repository is deprecated. The last version in this repository is 3.0. Refer to "aocl-scalapack" repository under the same "amd" organization for AOCL Scalapack 3.1 release onwards. https://gi…
aocl-data-analytics9活跃AMD Optimized Data Analytics Library
resource-optimizer9活跃—
ZenDNN-pytorch9活跃—
GenAI-contest8活跃—
mumps-build8活跃—
spack8forkA flexible package manager that supports multiple versions, configurations, platforms, and compilers.
onnxruntime7forkONNX Runtime: cross-platform, high performance ML inferencing and training accelerator
skillscope7活跃AMD's testing harness for AI Agent Skills
ImageTilingNPU6活跃—
quark-documentation6归档Quark (new quantizer) documentation. https://quark.docs.amd.com/latest/index.html#
aocl-scalapack5forkAMD optimized ScaLAPACK development repository. Please refer build and test documentation from AOCL User Guide at AMD Developer portal https://developer.amd.com/amd-aocl/
dotnet-runtime5fork.NET is a cross-platform runtime for cloud, mobile, desktop, and IoT apps.
vaip5活跃vaip open source
Linux_ISP_Kernel4活跃—
pensando-elk4活跃—
xir4活跃vaip open source
agentsim3活跃Developer tool for visualizing, debugging, and understanding agent traces
aocl-spack3活跃DEPRECATED. This Spack repo is deprecated. Refer to "spack" repository under the same "amd" organization for AOCL Spack packages. https://github.com/amd/spack
convnet-benchmarks3forkEasy benchmarking of all public open-source implementations of convnets
DynamicDispatch3活跃vaip open source
embedded-x86-ai3活跃—
Linux_ISP_libcamera3活跃—
logslop3活跃LogsLop is a log summarization tool that removes repeated messages from log files, enabling readers and agents to home in quickly on key message events.
node-unzip3forknode.js cross-platform unzip using streams
rocALUTION3forkNext generation library for iterative sparse solvers for ROCm platform
ryzenai-mcp-server3活跃Sharing code that allows for a fully managed MCP server that provides up-to-date access to AMD Ryzen AI documentation, code samples, and official examples (docs, code samples, and examples are all pulled from public i…
ZenDNN-tensorflow3活跃—
apollo2forkAn open autonomous driving platform
Gyanam-GPU-Observability-Assistant2活跃Fleet Intelligence for Realtime GPU observability
lz42forkExtremely Fast Compression algorithm
opencv2forkOpen Source Computer Vision Library
OPTIMA-Architecture2活跃OPTIMA — Data Intelligence Platform AI and IT Analytics. A self-hosted Docker Compose data intelligence platform combining Apache Iceberg, PuppyGraph, Dremio, Dagster, and an AI assistant for natural language queries…
transformers2活跃vaip open source
zstd2forkZstandard - Fast real-time compression algorithm
AOCL_jll.jl1活跃—
existdb-node1forkA node.js client library for the eXist-db Native XML Database.
flashrom-amd1活跃Publish AMD's changes to the opensource flashrom project
git-time-lapse1活跃—
graph_engine1活跃vaip open source
oneDNN-Zen1活跃For upstreaming AMD CPU optimization in OneDNN, which is widely used in all framework like PyTorch, TF, vLLM
tock1forkA secure embedded operating system for microcontrollers
tools-and-sdks1活跃For distributing software via Spack tool
vart1活跃vaip open source
VVAS1活跃For the VAI 6.2 release, we are planning to migrate the VVAS GHES source code to github.com/amd/VVAS
acoll0活跃—
aocl-rng0活跃AOCL-RNG 4.2 Library Documentation. For latest API information, refer to link in README.md
astc-encoder0forkThe Arm ASTC Encoder, a compressor for the Adaptive Scalable Texture Compression data format.
avm0forkAVM (AOM Video Model) is the reference software for AV2 codec from Alliance for Open Media (https://aomedia.org/).
ClickHouse0forkClickHouse® is a real-time analytics database management system
LMCache_zen0活跃—
lzbench0forklzbench is an in-memory benchmark of open-source compressors
lzbench-old0活跃—
mlcflow0forkMLCFlow: Simplifying MLPerf Automations
mlperf-automations0forkThis repository contains automation scripts designed to run MLPerf Inference benchmarks. Originally developed for the Collective Mind (CM) automation framework, these scripts have been adapted to leverage the MLC auto…
opencv_extra0forkOpenCV extra data
openrng0活跃—
openssl0forkTLS/SSL and crypto library
opus0forkModern audio compression for the internet.
perfspect0活跃—
processwatch0活跃—
quark-examples0活跃—
target_factory0活跃vaip open source
trace_logging0活跃vaip open source
unilog0活跃vaip open source
Yggdrasil0forkCollection of builder repositories for BinaryBuilder.jl
AMD-AGI:72 个仓库(AMD AI Group:模型、训练、推理优化)

活跃非 fork 62 · fork 10 · 归档 0 · 按 Star 数排序

仓库Star状态官方简介
Instella334活跃Fully Open Language Models with Stellar Performance
Hyperloom256活跃An agentic system that auto-optimizes LLM workloads on AMD GPUs.
GEAK196活跃Generating Efficient AI-Centric Kernels
AMD-LLM195活跃Training code and resources for AMD-135M language models on AMD GPUs.
Primus136活跃A flexible and high-performance training framework designed for large-scale foundation model training on AMD GPUs
Nitro-E129活跃Nitro-E is a family of text-to-image diffusion models focused on highly efficient training.
Nitro-1126活跃This repo provides a working re-implementation of Latent Adversarial Diffusion Distillation by AMD
AgentKernelArena123活跃AgentKernelArena is a platform for evaluating and training AI agents for GPU kernel optimization, combining realistic tasks, correctness validation, and performance-based rewards for reinforcement learning.
TraceLens100活跃Automating analysis from trace files
Magpie85活跃A lightweight, general-purpose framework for evaluating GPU kernel and benchmark.
Instella-MoE80活跃Instella MoE is a fully open state-of-the-art Mixture-of-Experts language model, trained end-to-end on AMD Instinct™ GPUs.
Apex78活跃Agents, and RL environment, for optimizing GPU kernels on AMD ROCm using LLM agents. Benchmarks LLM serving workloads end-to-end, profiles bottleneck kernels, optimizes them via Claude Code or Codex, and scores on com…
Primus-Turbo71活跃A high-performance acceleration library dedicated to large-scale model training on AMD GPUs
Micro-World70活跃For world model code developing and releasing.
Primus-SaFE59活跃Primus-SaFE(Stability and Fault Endurance)
Hummingbird47活跃AMD 0.9B efficient text to video diffusion model
FLy45活跃FLy: Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match (ICLR 2026)
Nitro-T42活跃Nitro-T is a family of text-to-image diffusion models focused on highly efficient training.
PARD37活跃PARD: Accelerating LLM Inference with Low-Cost PARallel Draft Model Adaptation (ICLR 26)
Gumiho33活跃Official Implementation of "Gumiho: A Hybrid Architecture to Prioritize Early Tokens in Speculative Decoding" (ICML'25)
DoSSR30活跃[NeurIPS 2024]Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs
maxtext-slurm30活跃Toolkit for launching and observing MaxText training on Slurm-managed GPU clusters
DUET-VLM26活跃DUET-VLM: Dual stage Unified Efficient Token reduction for VLM Training and Inference
Infera26活跃More token goodput from frontier models. A distributed, SLA-aware serving mesh — disaggregated prefill/decode, KV-aware routing, and cache offload, tuned to your production SLA.
Instella-T2I24活跃Instella-T2I: Pushing the Limits of 1D Discrete Latent Space Image Generation
InstellaVL24活跃Official code repo for instellavl-1b
ReNeg19活跃ReNeg: Learning Negative Embedding with Reward Guidance
HummingbirdXT17活跃This repository presents an efficient acceleration pipeline for Diffusion Transformer (DiT) based video generation models, optimized for AMD client-grade GPUs, including Navi48 dGPUs and Strix Halo iGPUs.
torchtitan-amd17forkA PyTorch native platform for training generative AI models
GEAK-eval15活跃—
Neurips2025-GPU-kernels-Tutorial15活跃Repo containing artifacts for Neurips 2025 tutorial- How to Build Agents to Generate Kernels for Faster LLMs (and Other Models!)
AMD-Hybrid-Models14活跃Official repo for AMD hybrid models training and inference workflow
ReasonLite12活跃Efficient Reasoning Model
AMDKernelVault10活跃AMDKernelVault A research and engineering playground for HIP and Triton kernel generation, optimization, and LLM training, including continued pretraining, SFT, and RL (RLHF/RLVR) pipelines.
Instella-Math9活跃Training code for Instella-Math model
pytorch-training-benchmark8活跃DEPRECATED: This is a benchmark script only for archiving purpose
gpt-fast7活跃The GPT-Fast for Multimodal Models on AMD GPUs
AMD_QTViT6活跃QT-ViT: Improving Linear Attention in ViT with Quadratic Taylor Expansion
FarSkip-Collective6活跃Training and inference implementation of FarSkip-Collective models enabling communication-computation overlap
Tyr-the-Pruner6活跃Týr-the-Pruner: Structural Pruning LLMs via Global Sparsity Distribution Optimization (NeurIPS25)
AMD_FDViT5活跃FDViT: Improve the Hierarchical Architecture of Vision Transformer (ICCV 2023)
AMD_IFFN5活跃Enhancing Vision Transformer: Amplifying Non-Linearity in Feedforward Network Module (ICML 2024)
awesome-rocm-autodrive5活跃Examples of training autodrive models in ROCm
PrimusClaw5活跃LLM agent orchestration on Kubernetes: autonomous coding-agent sessions in per-session sandboxes.
DynamicChunkingDiT3活跃Official code for DC-DiT: Adaptive Compute and Elastic Inference for Visual Generation via Dynamic Chunking
m3d_rocm3活跃This project is an optimized version of Matrix3D. It has better compatibility with ROCm ecosystem.
AMD-Spark2活跃SparK: Query-Aware Unstructured Sparsity with Recoverable KV Cache Channel Pruning
Athena-PRM2活跃Athena-PRM: Enhancing Large Reasoning Model with Data-efficient Process Reward Models
HF_PEFT_GPT_OSS2活跃Hugging Face PEFT for GPT_OSS
Primus-DLRM2活跃DLRM implementation for Primus
sand-pipeline2活跃Synthetic data generation pipeline, finetuning and evaluation scripts.
AMD-EfficientVLM1活跃Introducing the AMD-EfficientVLM (Vision-Language Model)
axlearn1forkAn Extensible Deep Learning Library
nixl1forkNVIDIA Inference Xfer Library (NIXL)
ALTO0活跃ALTO: Advanced Low-precision Training and Optimization
AMD_MSWA0活跃MSWA: Refining Local Attention with Multi-ScaleWindow Attention
AMDLongContextServing0活跃Kimi-Linear long-context FP8 benchmark tooling for vLLM on AMD MI355X (AITER MLA head-padding + patched decode/prefill kernels).
composer0forkSupercharge Your Model Training
diffusion-models-inference0活跃Diffusion model inference benchmarking, profiling and optimizing
DLRMBenchmark0活跃Repository for Showcasing DLRM v2 Functionality on AMD
dynamo0forkA Datacenter Scale Distributed Inference Serving Framework
faster-action-models0活跃Faster inference for World Action Models (FastWAM, DiT4DiT) on AMD GPUs: GEAK optimized kernels, fewer denoising steps, and Compact Action Experts.
hipBLASLt0forkhipBLASLt is a library that provides general matrix-matrix operations with a flexible API and extends functionalities beyond a traditional BLAS library
LoKRA0活跃Low Kruskal Rank Adaptation (ICML 2026)
mlcommons-inference0forkReference implementations of MLPerf® inference benchmarks
mlperf-common0活跃mlperf-common
pr_pundit0活跃An agent for OSS contributions
prime_amd0活跃Prime repository with functionality enabled on AMD
Repo2Reproduce0活跃Agentic system for reproducing research paper results from official paper and code on AMD ROCm GPUs.
ssd0forkA lightweight inference engine supporting speculative speculative decoding (SSD).
torchforge0forkPyTorch-native post-training at scale
vllm-20260forkA high-throughput and memory-efficient inference and serving engine for LLMs
AMD-Ecosystem:213 个仓库(AMD 开源生态移植与适配)

活跃非 fork 23 · fork 190 · 归档 0 · 按 Star 数排序

仓库Star状态官方简介
gsplat-amd39活跃—
llama.cpp27forkLLM inference in C/C++
hipDF19活跃hipDF - GPU DataFrame Library
flashinfer12forkFlashInfer+ROCm: ROCm port of FlashInfer
hipMM8活跃HIP Memory Manager (ROCm-DS)
hipVS7活跃—
warp7forkA Python framework for accelerated simulation, data generation and spatial computing.
hipGRAPH6活跃Portable, accelerated graph algorithms for GPUs
ray6forkRay is a unified framework for scaling AI and Python applications. Ray consists of a core distributed runtime and a set of AI Libraries for accelerating ML workloads.
rocGRAPH6活跃Accelerated graph algorithms for AMD GPUs
xgboost6forkScalable, Portable and Distributed Gradient Boosting (GBDT, GBRT or GBM) Library, for Python, R, Java, Scala, C++ and more. Runs on single machine, Hadoop, Spark, Dask, Flink and DataFlow
taichi5forkProductive, portable, and performant GPU programming in Python.
dgl4forkPython package built to ease deep learning on graph, on top of existing DL frameworks.
hipRaft4活跃—
moat4活跃Porting popular CUDA projects to ROCm/HIP, one repository at a time. Driven by Claude.
triton-inference-server-server3forkThe Triton Inference Server provides an optimized cloud and edge inferencing solution.
ComfyUI2forkThe most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.
Genesis2forkA generative world for general-purpose robotics & embodied AI learning on rocm
hipCIM2forkcuCIM - RAPIDS GPU-accelerated image processing library
LichtFeld-Studio2forkTrain, inspect, edit, automate, and export 3D Gaussian Splatting scenes from a single native application.
quadrants2forkquadrants on rocm
rocm-rag2活跃—
Stanford-Megatron-LM2forkOngoing research training transformer models at scale
verl2forkverl: Volcano Engine Reinforcement Learning for LLMs
ai-ecosystem1活跃This repository contains documentation for the ROCm AI ecosystem: framework installation and setup, large-scale model training, LLM and diffusion serving, and performance optimization on AMD GPUs. The ecosystem builds…
datascience1活跃AMD ROCm-DS - GitHub Home
ElasticFusion1forkReal-time dense visual SLAM system
lifescience-examples1活跃—
MASt3R-SLAM1fork[CVPR 2025] MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors
megablocks1fork—
MONAI1forkAI Toolkit for Healthcare Imaging
monai-model-zoo1forkMONAI Model Zoo that hosts models in the MONAI Bundle format.
MONAILabel1forkMONAI Label is an intelligent open source image labeling and learning tool.
ROCmDS-cmake1活跃This is a collection of CMake modules that are useful for all ROCm-DS projects. By sharing the code in a single place it makes rolling out CMake fixes easier.
simulation1活跃—
triton-inference-server-core1forkThe core library and APIs implementing the Triton Inference Server.
.github0活跃AMD Ecosystem Github Repo
3DGS-LM0fork3DGS-LM accelerates Gaussian-Splatting optimization by replacing the ADAM optimizer with Levenberg-Marquardt. (ICCV 2025)
3DUNDERWORLD-SLS-GPU_CPU0forkA structured light scanner
3P-ADMM-PC20fork—
accelerated-scan0forkAccelerated First Order Parallel Associative Scan
aihwkit0forkIBM Analog Hardware Acceleration Kit
alien0forkALIEN is a CUDA-powered artificial life simulation program.
amgcl0forkC++ library for solving large sparse linear systems with algebraic multigrid method
anari-visionaray0fork—
arbor0forkThe Arbor multi-compartment neural network simulation library.
arrayfire0forkArrayFire: a general purpose GPU library.
AutoDock-GPU0forkAutoDock for GPUs and other accelerators
bam0fork—
barney0forkA Scalable (and Optionally, Data-Parallel) ANARI Multi-GPU Path Tracer
baspacho0forkDirect solver for sparse SPD matrices for nonlinear optimization. Implements supernodal Cholesky decomposition algorithm, and supports GPU (CUDA).
bellhopcuda0forkCUDA and C++ port of BELLHOP / BELLHOP3D underwater acoustics simulator
brax0forkMassively parallel rigidbody physics simulation on accelerator hardware.
brian2cuda0forkA brian2 extension to simulate spiking neural networks on GPUs
catboost0forkA fast, scalable, high performance Gradient Boosting on Decision Trees library, used for ranking, classification, regression and other machine learning tasks for Python, R, Java, C++. Supports computation on CPU and GPU.
Clair30forkClair3 - Symphonizing pileup and full-alignment for deep learning-based long-read variant calling
CogVideo0forktext and image to video generation: CogVideoX (2024) and CogVideo (ICLR 2023)
colmap0forkCOLMAP - Structure-from-Motion and Multi-View Stereo
CTranslate20forkFast inference engine for Transformer models
CubbyFlow0forkVoxel-based fluid simulation engine for computer games
cuBQL0forkA CUDA BVH Build And Query Library
cubvh0forkMesh tools.
cuda-efficient-features0forkA CUDA implementation of keypoint detection and descriptor extraction
CUDA-ScanMatcher-ICP0forkA high performance CUDA implementation of Scan Matching via the Iterative Closest Point Algorithm
cudaKDTree0fork—
CudaSift0forkA CUDA implementation of SIFT for NVidia GPUs (1.2 ms on a GTX 1060)
CuMesh0forkCuda mesh utils.
cuPDLP-C0forkCode for solving LP on GPU using first-order methods
cuPDLPx0forkA GPU-Accelerated First-Order LP Solver
cupoch0forkRobotics with GPU computing
CuRast0forkCuda-Based Software Rasterization for Billions of Triangles
cuSZ0forkA GPU accelerated error-bounded lossy compression for scientific data.
CV-CUDA0forkCV-CUDA™ is an open-source, GPU accelerated library for cloud-scale image processing and computer vision.
dart0forkResearch-focused C++23 physics engine for robotics, animation, and machine learning, with Python bindings
dask-hip0活跃First release of rocm based rocm-ds dask-hip
DEM-Engine0forkA dual-GPU DEM solver with complex grain geometry support
dgSPARSE-Lib0forkPyTorch-Based Fast and Efficient Processing for Various Machine Learning Applications with Diverse Sparsity
dietgpu0forkGPU implementation of a fast generalized ANS (asymmetric numeral system) entropy encoder and decoder, with extensions for lossless compression of numerical and other data types in HPC/ML applications.
diff-surfel-rasterizations0forkCollection of differentiable surfel rasterizers with different feature channels.
diff-surfel-tracing0forkA fully differentiable 2D Gaussian ray tracer built on 2DGS and OptiX.
DiffPhysDrone0forkPublished on Nature Machine Intelligence! The first real robot(quadrotor) based on differentiable physics training.
egg.c0forkEGGROLL in C, integer-first training
EnvGS0fork[CVPR 2025] EnvGS: Modeling View-Dependent Appearance with Environment Gaussian
espresso0forkThe ESPResSo package
evo20forkGenome modeling and design across all domains of life
evogp0forkA GPU-accelerated library for Tree-based Genetic Programming, leveraging PyTorch and custom CUDA kernels for high-performance evolutionary computation. It supports symbolic regression, classification, and policy optim…
faiss0forkA library for efficient similarity search and clustering of dense vectors.
FaithC0fork[CVPR 2026 (Oral)] Official Torch/CUDA Implementation of Faithful Contouring
Fast-Poisson-Image-Editing0forkA fast poisson image editing implementation that can utilize multi-core CPU or GPU to handle a high-resolution image input.
faster-gaussian-splatting0forkAn efficient and research-friendly Gaussian Splatting framework described in the CVPR'26 paper "Faster-GS: Analyzing and Improving Gaussian Splatting Optimization"
FastGeodis0forkFast Implementation of Generalised Geodesic Distance Transform for CPU (OpenMP) and GPU (CUDA)
fdtd3d0forkfdtd3d is an open source 1D, 2D, 3D FDTD electromagnetics solver with MPI, OpenMP and CUDA support for x64, ARM, ARM64, RISC-V, PowerPC, Wasm architectures
ffpa-attn0fork🤖FFPA: Extends FlashAttention-2 via Split-D for large headdims, 1.5x~3×↑🎉 vs SDPA, up to 430T🎉 on H200.
finance0活跃—
FLAMEGPU20forkFLAME GPU 2 is a GPU accelerated agent based modelling framework for CUDA C++ and Python
flashinfer-bench0forkBuilding the Virtuous Cycle for AI-driven LLM Systems
flower0forkFlower: A Friendly Federated AI Framework
foldmason0forkMultiple Protein Structure Alignment at Scale with FoldMason
foldseek0forkFoldseek enables fast and sensitive comparisons of large structure sets.
fused-ssim0forkLightning fast differentiable SSIM.
gaussian_splatting0forkUnofficial implementation of 3D Gaussian Splatting in PyTorch + CUDA with MIT license
gdtk0forkThe Gas Dynamics Toolkit (GDTk) is a set of software tools for simulating high speed fluid flow, maintained at The University of Queensland and the University of Southern Queensland, Australia.
GOMC0forkGOMC - GPU Optimized Monte Carlo is a parallel molecular simulation code designed for high-performance simulation of large systems
GooFit0forkCode repository for the massively-parallel framework for maximum-likelihood fits, implemented in CUDA/OpenMP
google-deepvariant0forkDeepVariant is an analysis pipeline that uses a deep neural network to call genetic variants from next-generation DNA sequencing data.
gpu4pyscf0forkA plugin to use Nvidia GPU in PySCF package
GPU_IPC0forkThis is the first fully GPU Optimized IPC framework
Gpufit0forkGPU-accelerated Levenberg-Marquardt curve fitting in CUDA
GPUMD0forkGraphics Processing Units Molecular Dynamics
gpuRIR0forkPython library for Room Impulse Response (RIR) simulation with GPU acceleration
gRASPA0forkGPU Monte Carlo Simulation Code with a taste of RASPA
gsplat0forkCUDA accelerated rasterization of gaussian splatting
gtsam_points0forkA collection of GTSAM factors and optimizers for point cloud SLAM
HEonGPU0forkHEonGPU is a high-performance library that optimizes Fully Homomorphic Encryption (FHE) on GPUs. Leveraging GPU parallelism, it reduces computational load through concurrent execution. Its multi-stream architecture mi…
hip-ucxx0活跃First release of hip-ucxx for the ROCmDS 26.03 release supporting ROCm 7.2
icicle0forkA hardware acceleration library for compute intensive cryptography :ice_cube:
IsaacLab0forkUnified framework for robot learning built on NVIDIA Isaac Sim
k20forkFSA/FST algorithms, differentiable, with PyTorch compatibility.
kaldi0forkkaldi-asr/kaldi is the official location of the Kaldi project.
kaldifeat0forkKaldi-compatible online & offline feature extraction with PyTorch, supporting CUDA, batch processing, chunk processing, and autograd - Provide C++ & Python API
kaolin0forkA PyTorch Library for Accelerating 3D Deep Learning Research
KittenGpuLBVH0forkA high performance and friendly GPU LBVH implementation.
LC-framework0fork—
lc00forkOpen source neural network chess engine with GPU acceleration and broad hardware support.
LEAP0forkcomprehensive library of 3D transmission Computed Tomography (CT) algorithms with Python and C++ APIs, a PyQt GUI, and fully integrated with PyTorch
libSGM0forkStereo Semi Global Matching by cuda
lifescience0活跃AMD Life Science
LightGBM0forkA fast, distributed, high performance gradient boosting (GBT, GBDT, GBRT, GBM or MART) framework based on decision tree algorithms, used for ranking, classification and many other machine learning tasks.
LiteGS0forkA refactored codebase for Gaussian Splatting. Training 3DGS in 50 seconds!
llm-awq0fork[MLSys 2024 Best Paper Award] AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration
llm.c0forkLLM training in simple, raw C/CUDA
llmextension0活跃—
LMCache0forkLMCache: Supercharge Your LLM with the Fastest KV Cache Layer
mahout0forkApache Mahout - an environment for quickly creating scalable, performant machine learning applications.
ManiSkill0forkManipulation Skill Framework, an open source GPU parallelized robotics simulator and benchmark
marian-dev0forkFast Neural Machine Translation in C++ - development repository
mcx0forkMonte Carlo eXtreme (MCX) - Physically accurate and validated GPU ray-tracer
metaeuk0forkMetaEuk - sensitive, high-throughput gene discovery and annotation for large-scale eukaryotic metagenomics
mHC.cu0forkmHC kernels implemented in CUDA
minibwa0forkSuccessor of bwa-mem for short-read alignment
minimap20forkA versatile pairwise aligner for genomic and spliced nucleotide sequences
mjlab0forkIsaac Lab API, powered by MuJoCo-Warp, for RL and robotics research
MMseqs20forkMMseqs2: ultra fast and sensitive search and clustering suite
MPPI-Generic0forkTemplated C++/CUDA implementation of Model Predictive Path Integral Control (MPPI)
mujoco0forkMulti-Joint dynamics with Contact. A general purpose physics simulator.
mujoco_warp0forkGPU-optimized version of the MuJoCo physics simulator.
mumax30forkGPU-accelerated micromagnetic simulator
nerfacc0forkA General NeRF Acceleration Toolbox in PyTorch.
nerfacc-amd0活跃—
newton0forkAn open-source, GPU-accelerated physics simulation engine built upon NVIDIA Warp, specifically targeting roboticists and simulation researchers.
nextflow0forkA DSL for data-driven computational pipelines
ntransformer0forkHigh-efficiency LLM inference engine in C++/CUDA. Run Llama 70B on RTX 3090.
NVFlare0forkNVIDIA Federated Learning Application Runtime Environment
OCTproZ0forkOCTproZ is an open source software for optical coherence tomography processing and visualization.
ohm0forkAn efficient, extensible occupancy map supporting probabilistic occupancy, normal distribution transforms in CPU and GPU.
omp-bowtie2-prime0forkA simplified version of bowtie2 that uses OpenMP based parallelization
oneflow0forkOneFlow is a deep learning framework designed to be user-friendly, scalable and efficient.
op43dgs0fork[ECCV'24] On the Error Analysis of 3D Gaussian Splatting and an Optimal Projection Strategy
Open3D0forkOpen3D: A Modern Library for 3D Data Processing
opencv0forkOpen Source Computer Vision Library
opencv_contrib0forkRepository for OpenCV's extra modules
openfold0forkTrainable, memory-efficient, and GPU-friendly PyTorch reproduction of AlphaFold 2
openfold-30forkA fully open source biomolecular structure prediction model based on AlphaFold3
OpenUSD0forkUniversal Scene Description
openvdb0forkOpenVDB - Sparse volume data structure and tools
Paddle0forkPArallel Distributed Deep LEarning: Machine Learning Framework from Industrial Practice (『飞桨』核心框架,深度学习&机器学习高性能单机、分布式训练和跨平台部署)
PaddleNLP0forkEasy-to-use and powerful LLM and SLM library with awesome model zoo.
PaddleOCR0forkTurn any PDF or image document into structured data for your AI. A powerful, lightweight OCR toolkit that bridges the gap between images/PDFs and LLMs. Supports 100+ languages.
PaddleRec0forkRecommendation Algorithm
PaddleSpeech0forkEasy-to-use Speech Toolkit including Self-Supervised Learning model, SOTA/Streaming ASR with punctuation, Streaming TTS with text frontend, Speaker Verification System, End-to-End Speech Translation and Keyword Spotti…
PhysX0forkNVIDIA PhysX SDK
plumed20forkDevelopment version of plumed 2
plvs0forkPLVS is a real-time SLAM system with points, lines, volumetric mapping and 3D unsupervised incremental segmentation.
Pointcept0forkPointcept: Perceive the world with sparse points, a codebase for point cloud perception research. Latest works: Utonia (ICML'26), Concerto (NeurIPS'25), Sonata (CVPR'25 Highlight), PTv3 (CVPR'24 Oral)
popsift0forkPopSift is an implementation of the SIFT algorithm in CUDA.
prismatic0forkC++/CUDA package for parallelized simulation of image formation in Scanning Transmission Electron Microscopy (STEM) using the PRISM and multislice algorithms
qrack0forkComprehensive, GPU accelerated framework for developing universal virtual quantum processors
Quest0fork[ICML 2024] Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference
QUICK0forkQUICK: A GPU-enabled ab intio quantum chemistry software package
rmagine0forkA robot wants to simulate too
rmcl0forkMobile Robot Localization in 3D Triangle Meshes & Geometric Scene Graphs
RoboticsDiffusionTransformer0forkRDT-1B: a Diffusion Foundation Model for Bimanual Manipulation
rocmds-logger0活跃—
rocOpt0forkGPU accelerated decision optimization
RWKV-CUDA0forkThe CUDA version of the RWKV language model ( https://github.com/BlinkDL/RWKV-LM )
RXMesh0forkGPU-accelerated triangle mesh processing
SCAMP0forkThe fastest way to compute matrix profiles on CPU and GPU!
sglang0forkSGLang is a high-performance serving framework for large language models and multimodal models.
splatad0forkSplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving
sppark0forkZero-knowledge template library
stdgpu0forkstdgpu: Efficient STL-like Data Structures on the GPU
STRUMPACK0forkStructured Matrix Package (LBNL)
symforce0forkFast symbolic computation, code generation, and nonlinear optimization for robotics
thundergbm0forkThunderGBM: Fast GBDTs and Random Forests on GPUs
ThunderGBM-amd0活跃—
TIGRE0forkTIGRE: Tomographic Iterative GPU-based Reconstruction Toolbox
tiny-vllm0forkBuild your own high performance LLM inference engine in C++ and CUDA - a smaller version of vLLM
torch-linear-assignment0forkBatch computation of the linear assignment problem on GPU.
triton-inference-server-backend0forkCommon source, scripts and utilities for creating Triton backends.
triton-inference-server-onnxruntime_backend0forkThe Triton backend for the ONNX Runtime.
triton-inference-server-python_backend0forkTriton backend that enables pre-process, post-processing and other logic to be implemented in Python.
triton-inference-server-pytorch_backend0forkThe Triton backend for the PyTorch TorchScript models.
triton-inference-server-tensorflow_backend0forkThe Triton backend for TensorFlow.
triton-inference-server-third_party0forkThird-party source packages that are modified for use in Triton.
triton-inference-server-vllm_backend0fork—
TTT3R0fork[ICLR 2026] A simple state update rule to enhance length generalization for CUT3R
TurboFNO0forkThe first fully fused FFT–GEMM–iFFT GPU kernel.
unified-cache-management0forkPersist and reuse KV Cache to speedup your LLM.
Velvet0forkA CUDA-accelerated cloth simulation engine based on Extended Position Based Dynamics (XPBD).
visionaray0forkA C++-based, cross platform ray tracing library
yalm0forkYet Another Language Model: LLM inference in C++/CUDA, no libraries except for I/O
YarnBall0forkA massively parallel high performance GPU Cosserat Rods Simulator
ZhiLight0forkA highly optimized LLM inference acceleration engine for Llama and its variants.
AMDResearch:23 个仓库(AMD 研究院与教学原型)

活跃非 fork 21 · fork 2 · 归档 0 · 按 Star 数排序

仓库Star状态官方简介
Ryzers89活跃Ryzen AI Dockerfiles for ML, Robotics and Development Environments
Riallto87活跃The Riallto Open Source Project from AMD
aup-learning-cloud69活跃AUP Learning Cloud is a customized JupyterHub platform that delivers an intuitive, hands‑on AI learning experience with AMD‑accelerated toolkits.
DAGEE50活跃Directed Acyclic Graph Execution Engine (DAGEE) is a C++ library that enables programmers to express computation and data movement, as task graphs that are scheduled concurrently and asynchronously on both CPUs and GPUs.
intellikit37活跃IntelliKit is a collection of intelligent tools designed to make GPU kernel development, profiling, and validation accessible to LLMs and human developers alike. Built for AMD ROCm, these tools provide clean abstracti…
NPUEval34活跃NPUEval is an LLM evaluation dataset written specifically to target AIE kernel code generation on RyzenAI hardware.
intelliperf23活跃Automated bottleneck detection and solution orchestration
hpcfund21活跃AMD HPC Research Fund Cloud
omniprobe18活跃Simply log all kernel durations
aup-ai-tutorials15活跃AUP AI Tutorials
introduction_to_hip14活跃—
AMDGeniePIM8活跃This repository contains AMD GeniePIM, an analytical performance model for modeling GEMV execution on emerging PIM architectures.
atmi6forkAsynchronous Task and Memory Interface, or ATMI, is a runtime framework and programming model for heterogeneous CPU-GPU systems. It provides a consistent, declarative API to create task graphs on CPUs and GPUs (integr…
instrument-amdgpu-kernels6forkLLVM/MLIR based compiler instrumentation of AMD GPU kernels
aup-teaching-labs5活跃Provide a series of AMD learning materials to support teaching and instruction.
ai4science-studio4活跃How-to guides and AMD/ROCm optimization recipes for open AI-for-science models.
dh_comms3活跃Facilities to allow device code to push arbitrary data onto a FIFO queue that can be processed at runtime by host code.
kerneldb3活跃A library for interacting with CDNA code objects to support performance analysis
OmniHub2活跃Tools for AI/ML Workload Analysis and Characterization
AMD-DMA-Latte1活跃A collection of DMA-based ML collective implementations and microbenchmarks which harness advanced DMA engine features on AMD Instinct GPUs.
AdaHOP0活跃—
geometric-nn-ops0活跃—
intelliperf-action0活跃—
GPUOpen-LibrariesAndSDKs:85 个仓库(GPUOpen SDK 与库)

活跃非 fork 83 · fork 1 · 归档 1 · 按 Star 数排序

仓库Star状态官方简介
VulkanMemoryAllocator3,533活跃Easy to integrate Vulkan memory allocation library
FidelityFX-SDK1,829活跃The main repository for the FidelityFX SDK.
Cauldron1,114活跃A simple framework for rapid prototyping on Vulkan or Direct3D 12
RadeonRays_SDK1,101活跃Radeon Rays is ray intersection acceleration library for hardware and software multiplatforms using CPU and GPU
D3D12MemoryAllocator1,049活跃Easy to integrate memory allocation library for Direct3D 12
V-EZ892活跃—
AMF754活跃The Advanced Media Framework (AMF) SDK provides developers with optimal access to AMD devices for multimedia processing
RadeonProRenderBlenderAddon554活跃This hardware-agnostic rendering plug-in for Blender uses accurate ray-tracing technology to produce images and animations of your scenes, and provides real-time interactive rendering and continuous adjustment of effe…
Capsaicin471活跃AMD ARR team rendering framework
BlenderUSDHydraAddon399活跃This add-on allows you to assemble and compose USD data with Blender data and render it all using various renderers via Hydra.
RenderPipelineShaders360活跃Render Pipeline Shaders SDK
RadeonProRender-Baikal333活跃—
AGS_SDK308活跃AMD GPU Services (AGS) library and samples
RadeonProRenderSDK287活跃AMD Radeon™ ProRender is a powerful physically-based path traced rendering engine that enables creative professionals to produce stunningly photorealistic images.
Orochi260活跃—
RadeonProRenderUSD242活跃This plug-in allows GPU or CPU accelerated viewport rendering on all OpenCL 1.2 hardware for the open source USD and Hydra system. You can build this plug-in as a USDView plug-in or a Houdini plug-in.
MxGPU-Virtualization208归档—
brotli_g_sdk183活跃Brotli-G SDK provides an improved lossless compression ratio with GPU decompression support than the standard Brotli compression algorithm maintained by the IETF (also known as RFC7932)
Radeon-ReLive-VR164活跃—
HelloD3D12161活跃Introductory DirectX 12 sample
OCL-SDK154活跃—
TAN149活跃AMD TrueAudio Next is a software development kit for GPU accelerated audio signal processing
ForwardPlus11148活跃AMD Forward+ sample based on DirectX 11
display-library146活跃AMD Display Library SDK
HIPRT146活跃—
WorkGraphPlayground140活跃A DirectX12-based C++-application that allows graphics programmers to learn and experiment with the new Work Graphs feature using HLSL shaders. It runs on Windows 10 and Windows 11 systems.
glTFSample131活跃A simple demo to show off the capabilities of the Cauldron framework
GPUParticles11128活跃AMD GPU particles sample based on DirectX 11
HelloVulkan123活跃Introductory Vulkan sample
LiquidVR108活跃The LiquidVR™ SDK is a platform based on DirectX 11 designed to simplify and optimize VR development
Schola97活跃Schola is a plugin for enabling Reinforcement Learning (RL) in Unreal Engine. It provides tools to help developers create environments, define agents, and connect to python-based RL frameworks such as OpenAI Gym, RLli…
ADLX91活跃AMD Device Library eXtra
RadeonML87活跃—
TiledLighting1178活跃AMD compute-based tiled lighting sample based on DirectX 11
HIPRTSDK70活跃—
DGF-SDK68活跃The main repository for the AMD Dense Geometry Format SDK
Tessellation66活跃OpenGL sample that demonstrates terrain tessellation on the GPU
nBodyD3D1260forkThis repo contains the DirectX Graphics samples that demonstrate how to build graphics intensive applications on Windows.
GameEngineIntegrations55活跃Information about example integrations of GPUOpen technology into game engines
cpu-core-counts54活跃A sample demonstrating how to correctly detect physical core and logical processor counts on AMD processors.
RadeonProRenderMayaPlugin51活跃This hardware-agnostic rendering plug-in for Maya uses accurate ray-tracing technology to produce images and animations of your scenes, and provides real-time interactive rendering and continuous adjustment of effects.
RadeonImageFilter48活跃—
Vulkan-Samples47活跃based on Khronos Vulkan-Samples
SSAA1145活跃AMD supersample anti-aliasing (SSAA) sample based on DirectX 11
AntiLag2-SDK42活跃The main repository for the AMD Radeon Anti-Lag 2 SDK.
RenderStudioKit40活跃—
MLAA1139活跃AMD morphological anti-aliasing (MLAA) sample based on DirectX 11
RapidFire_SDK39活跃RapidFire SDK - interface for cloud gaming and virtualization (VDI)
WorkGraphComputeRasterizer38活跃A compute/workgraph workload running inside the Cauldron framework
SilhouetteTessellation1133活跃AMD tessellation sample based on DirectX 11
WorkGraphsMeshNodeSample32活跃—
DirectStorageSample31活跃Demonstrating the advantages of using DirectStorage over standard file I/O asset loading. Includes the API and changes required to make such a pipeline work.
SPARSEtextures31活跃OpenGL sample demonstrating the GL_AMD_sparse_texture extension introduced by the AMD FirePro W and Radeon HD 7000 series
MiniDXNN30活跃An implementation of MLP using DX12 cooperative vector.
SeparableFilter1127活跃AMD separable filter sample based on DirectX 11
CrossfireAPI1126活跃AMD Crossfire API sample for DirectX 11
RadeonProRenderUE26活跃—
Streaming-SDK26活跃Advanced Interactive Streaming SDK
DepthBoundsTest1125活跃AMD depth bounds test DirectX 11 driver extension sample
Cauldron-Media24活跃A selection of media files needed by Cauldron samples
WorkGraphsIvyGenerationSample24活跃—
RadeonProRenderANARI22活跃—
DirectGMA_P2P21活跃Samples showing FIrePro DirectGMA features in OpenGL and OpenCL
FreesyncPremiumProSample21活跃A simple demo to show Freesync Premium Pro integration into DX12 and Vulkan API.
WorkGraphsHelloWorkGraphs21活跃—
UnityIntegrations20活跃Examples that demonstrate game integration with some of the libraries and SDKs available on GPUOpen.
DOPPEngine18活跃—
Framelock18活跃OpenGL sample that demonstrates how to enable Framelock in an application
ScholaExamples18活跃Schola Examples is an Unreal Engine project containing sample environments developed with the Schola plugin for Unreal Engine. Schola provides tools to help developers create environments, define agents, and connect t…
OutOfOrderRasterization16活跃Vulkan sample built on the Anvil framework
VkMBCNT16活跃Vulkan mbcnt sample that shows how to use the AMD_shader_ballot extension and mbcnt to perform a fast reduction within a wavefront
Barycentrics1113活跃Barycentric coordinates GCN shader extension sample for DirectX 11
DOPP13活跃—
WorkGraphsDirectX-Graphics-Samples13活跃A fork of https://github.com/microsoft/DirectX-Graphics-Samples modified to include a sample for Work Graphs
Barycentrics1212活跃Barycentric coordinates GCN shader extension sample for DirectX 12
VkD3DDeviceMapping11活跃Simple example to show how to map devices between Vulkan and Direct3D
WorkGraphsHelloMeshNodes11活跃—
RadeonProRenderMaxPlugin10活跃—
RadeonProRender-Tests7活跃—
RadeonProRenderSDKKernels4活跃—
RadeonProRenderSharedComponents3活跃—
CapsaicinTestMedia2活跃Capsaicin Framework test media assets
RadeonProRenderMayaUSD2活跃—
SPIRV-Tools2活跃Mirror of https://github.com/KhronosGroup/SPIRV-Tools
DirectXShaderCompiler0活跃Mirror of https://github.com/microsoft/DirectXShaderCompiler
GPUOpen-Effects:22 个仓库(GPUOpen 图形特效)

活跃非 fork 21 · fork 1 · 归档 0 · 按 Star 数排序

仓库Star状态官方简介
FidelityFX-FSR2,187活跃FidelityFX Super Resolution
FidelityFX-FSR22,081活跃FidelityFX Super Resolution 2
TressFX847活跃DirectX 12 and Vulkan libraries that provide convenient access to realistically rendered and simulated hair and fur
FEMFX511活跃—
FidelityFX372活跃A set of AMD provided GPU effects with provided source and sample implementations
FidelityFX-SSSR312活跃Stochastic Screen Space Reflections (SSSR)
FidelityFX-CAS232活跃Contrast Adaptive Sharpening (CAS)
GeometryFX227活跃DirectX 11 library that provides convenient access to compute-based triangle filtering (CTF)
FidelityFX-SPD213活跃Single Pass Downsampler (SPD)
FidelityFX-CLI203活跃FidelityFX CLI
ShadowFX166活跃DirectX 11 and 12 library that provides a scalable and GCN-optimized solution for deferred shadow filtering
FidelityFX-CACAO159活跃Combined Adaptive Compute Ambient Occlusion (CACAO)
FidelityFX-Denoiser155活跃FidelityFX Denoiser
FidelityFX-FSR2-Unity-URP139活跃FidelityFX FSR 2 for the Unity URP
FidelityFX-ParallelSort125活跃FidelityFX Parallel Sort
FidelityFX-FSR-Unity-URP116活跃FidelityFX FSR 1.0 for the Unity URP
FidelityFX-LPM103活跃Luma Preserving Mapper (LPM)
DepthOfFieldFX92活跃DirectX 11 library that provides a GCN-optimized Compute Shader implementation of Depth of Field
Hybrid-Shadows66活跃A sample demonstrating hybrid ray tracing and rasterisation for shadow rendering and use of the FidelityFX Denoiser.
FidelityFX-VariableShading56活跃FidelityFX Variable Shading
Hybrid-Stochastic-Reflections56活跃A sample demonstrating hybrid ray tracing and rasterisation for reflections.
pbrt-v433forkSource code to pbrt, the ray tracer described in the forthcoming 4th edition of the "Physically Based Rendering: From Theory to Implementation" book.
GPUOpen-Tools:66 个仓库(GPUOpen 分析与调试工具)

活跃非 fork 58 · fork 8 · 归档 0 · 按 Star 数排序

仓库Star状态官方简介
compressonator1,463活跃Tool suite for Texture and 3D Model Compression, Optimization and Analysis using CPUs, GPUs and APUs
GPU-Reshape494活跃GPU Reshape (GRS) is an API & vendor agnostic instrumentation framework, with instruction level validation.
radeon_gpu_analyzer483活跃The Radeon GPU Analyzer (RGA) is an offline compiler and code analysis tool for Vulkan, DirectX, OpenGL, and OpenCL.
radeon_gpu_profiler454活跃Radeon GPU Profiler (RGP) is a tool from AMD that allows for deep inspection of GPU workloads.
ocat365活跃The Open Capture and Analytics Tool (OCAT) provides an FPS overlay and performance measurement for D3D11, D3D12, and Vulkan
frame_latency_meter260活跃—
gpu_performance_api260活跃GPU Performance API for AMD GPUs
radeon_raytracing_analyzer256活跃The Radeon Raytracing Analyzer (RRA) is a tool to visualize and inspect Bounding Volume Hierarchies (BVH) for ray tracing applications.
radeon_memory_visualizer141活跃The Radeon Memory Visualizer (RMV) is a software tool that will allow users to analyze video memory usage on AMD Radeon GPUs.
radeon_compute_profiler85活跃The Radeon Compute Profiler (RCP) is a performance analysis tool that gathers data from the API run-time and GPU for OpenCL™ and ROCm/HSA applications. This information can be used by developers to discover bottleneck…
radeon_gpu_detective67活跃Tool for post-mortem analysis of GPU crashes.
isa_spec_manager62活跃Utilities for accessing AMD's Machine-Readable GPU ISA Specifications.
vscode_extensions29活跃Handy extensions for Visual Studio Code to work on AMD GPU technologies
radeon_developer_panel22活跃The Radeon Developer Panel (RDP) is a software tool that allows users to capture RGP profiles, RMV traces, RRA scenes, and RGD crash analysis dumps on Radeon GPUs.
smoldr22活跃A simple scripting tool to run DX12 shaders on HW using a text input file.
dev_driver_tools19活跃Developer Driver Tools components (Radeon Developer Service and Radeon Developer Panel)
device_info19活跃—
adl10活跃—
appsdk10活跃—
qt_common8活跃—
common_lib_ext_zlib_1.2.87活跃—
common_src_amddxext6活跃—
dynamic_library_module5活跃—
Fossilize5forkA serialization format for various persistent Vulkan object types.
llm_tools5活跃Hub for LLM tooling features, such as MCP servers, skills, agents and more.
renderdoc5forkRenderDoc is a stand-alone graphics debugging tool.
common_src_amdtbasetools4活跃Common base code shared by multiple tool projects
common_src_amdtoswrappers4活跃Common OS Abstraction layer shared by multiple tool projects
common_src_amdvkext4活跃—
Lyra4forkA simple to use, composable, command line parser for C++ 11 and beyond
opengl4活跃—
tsingleton4活跃—
adl_util3活跃—
Catch23forkA modern, C++-native, header-only, test framework for unit-tests, TDD and BDD - using C++11, C++14, C++17 and later (or C++03 on the Catch1.x branch)
cmake-modules3forkMy collection of CMake modules
common_lib_amd_acl3活跃—
common_lib_amd_ags_4.0.03活跃—
common_lib_ext_opengles3活跃—
common_lib_ext_utf8cpp3活跃—
common_src_misc3活跃—
googletest3forkGoogle Test
tinyxml_23活跃—
update_check_api3活跃The UpdateCheckAPI repository provides utility source code for AMD Tools to check for new releases available through GitHub's Release API.
windows_kits3活跃—
comgr_utils2活跃—
common_lib_ext_boost_1.592活跃—
common_lib_ext_glew_1.92活跃—
common_lib_ext_opencv_2.492活跃—
common_lib_ext_tinyxml2_5.0.12活跃—
common_lib_ext_zlib_1.2.102活跃—
common_src_acl_module_manager2活跃—
common_src_amdtmutex2活跃—
common_src_celf2活跃—
common_src_cmake_modules2活跃—
common_src_miniz2活跃—
common_src_version_info2活跃—
common_src_vsprops2活跃Common Visual Studio build settings shared across multiple projects
system_info_utils2活跃—
common_lib_ext_openexr_2.21活跃—
common_lib_ext_yaml_cpp1活跃—
cxxopts1forkLightweight C++ command line option parser
volk1forkMeta loader for Vulkan API
devdriver0活跃—
qt_isa_gui0活跃—
sdl_utils0活跃—
spm_db0活跃—
GPUOpen-Drivers:14 个仓库(GPUOpen 开源图形驱动)

活跃非 fork 6 · fork 4 · 归档 5 · 按 Star 数排序

仓库Star状态官方简介
AMDVLK1,977活跃AMD Open Source Driver For Vulkan
pal413归档Platform Abstraction Library
xgl238归档Vulkan API Layer
llpc185归档LLVM-Based Pipeline Compiler
gpurt85归档GPU Ray Tracing Library
spvgen36活跃Library to Generate SPIR-V Binary
llvm26fork—
llvm-dialects24活跃LLVM Dialects Library
AMD-Gfx-Drivers18活跃Forum for AMD OpenGL and Vulkan graphics drivers
amd-vulkan-versions15活跃Machine-readable mapping from public driver version to AMD Vulkan component version
libamdrdf12活跃AMD Radeon Data File library
llvm-project10归档 / forkThis is the canonical git mirror of the LLVM subversion repository. The repository does not accept github pull requests at this moment. Please submit your patches at http://reviews.llvm.org.
MetroHash10forkMetroHash v1: Exceptionally fast and statistically robust hash functions
CWPack2forkAn ANSI C encoder/decoder for the MessagePack serialization format / msgpack.org[C]

11.2 九个外围官方组织

AMD-HPC:14 个仓库(AMD HPC 应用性能团队)

活跃非 fork 10 · fork 4 · 归档 0 · 按 Star 数排序

仓库Star状态官方简介
CoralGemm16活跃—
rocBudAI8活跃This is rocBudAI -- your ROCm AI Buddy!
ShibuyaStream5活跃—
rooflineExtractor4活跃A tool for automated roofline analysis on HPC and AI workloads
amd-sbd3fork—
fp64monitor2活跃—
abinit1forkThe official github mirror of the Abinit repository. We welcome bug fixes and improvements. Note that most of the active developments are hosted on our https://gitlab.abinit.org/ server. Before embarking on making si…
HMMBench1活跃Set of benchmarks to test HMM functionality and performance
AI-GPU-Pauli-tensor-network0活跃A C++/HIP implementation of real-time quantum dynamics for the 2D transverse-field Ising model, using Tensor Network (TN) methods in the Heisenberg picture on AMD GPUs.
benchWarmer0活跃—
CvxCompress0forkChevron Developed Compression library. Part of the COFII framework.
hip-api-wrappers0活跃—
hpcperf0活跃—
specfem3d0forkSPECFEM3D Cartesian simulates acoustic (fluid), elastic (solid), coupled acoustic/elastic, or poroelastic seismic wave propagation in any type of conforming mesh of hexahedra (structured & unstructured).
AMDESE:51 个仓库(AMD 嵌入式/服务器软件工程(SEV、OpenBMC、EPYC Linux))

活跃非 fork 27 · fork 22 · 归档 3 · 按 Star 数排序

仓库Star状态官方简介
AMDSEV381活跃AMD Secure Encrypted Virtualization
linux-svsm147活跃Linux SVSM (Secure VM Service Module) for secure x86 virtualization in Rust
sev-tool86归档AMD SEV Tool
sev-guest59归档Tools, scripts, and configuration files necessary to demonstrate an end-to-end remote attestation example with SEV-SNP.
OpenBMC43活跃OpenBMC for Genoa SP5 socket platforms
linux35forkLinux kernel source tree
amd-perf-tools34活跃Documentation and diagnostic utilities for running Linux perf on AMD hardware
amd-bmc-ras30活跃—
amd_ucode_info27活跃Parse and display information about an amd-ucode (CPU microcode) container file
mem-encryption-tests26活跃Tests related to Memory Encryption
qemu26活跃QEMU fork
sched-scoreboard16活跃Scheduler scoreboard is a single toolkit to capture and report all the data related to the Linux Kernel Scheduler which can help analyze performance issues potentially caused due to modelling or the heuristics adopted…
linux-aspeed11forkOpenBMC Linux kernel source tree
ovmf10活跃EDK2 / OVMF related projects
amd-power-control9活跃OpenBMC power control for AMD CRBs.
linux-sdxi9forkLinux kernel source tree
bmc-ppr6活跃—
libhsmp6活跃The libhsmp repository is in the process of being deprecated. The repo will remain available for the near future though. All new work on HSMP test suites and commands are a part of the AMDESE/hsmputils repository.
linux-kvm6forkLinux kernel source tree
elves5forkAMD EPYC Linux Validation Exerciser Suite (ELVES)
amd-power-cap4活跃—
bmc-cpuinfo4活跃—
vr-firmware-update4活跃—
epyc-feature-doc3活跃AMD EPYC Linux feature Documentation project
workload-insight-tool3活跃Workload Insight Tool (syswit) is a tool to analyze workloads on Linux. It collects global and process related system stats exposed by the kernel (via procfs and sysfs files) in a time-series manner using collector. C…
amd_hsmp2活跃—
qemu-iommu2活跃—
rasdaemon2forkRasdaemon is a RAS (Reliability, Availability and Serviceability) logging tool. It records memory errors, using the EDAC tracing events. EDAC is a Linux kernel subsystem with handles detection of ECC errors from memor…
u-boot2forkOpenBMC "Das U-Boot" Source Tree
amd-bmc-ubm1活跃—
amd-ipmi-oem1活跃—
amd-lcd-lib1活跃—
avocado-misc-tests1forkCommunity maintained Avocado tests repository
bmcweb1forkA do everything Redfish, KVM, GUI, and DBus webserver for OpenBMC
elves_legacy1活跃AMD EPYC Linux Validation Exerciser Suite (ELVES)
ledmon1forkEnclosure LED Utilities
linux-iommu1forkLinux kernel source tree
linux-sdci1forkLinux kernel source tree
openbmc-build-scripts1归档 / fork—
pmic-error-injection-bmc1活跃—
qemu-openbmc1forkOfficial QEMU mirror. Please see https://www.qemu.org/contribute/ for how to submit changes to QEMU. Pull Requests are ignored. Please only use release tarballs from the QEMU website.
sev-apitest-kerneldriver1活跃Example Kernel driver for SEV apitest
avocado0forkAvocado is a set of tools and libraries to help with automated testing. One can call it a test framework with benefits. Native tests are written in Python and they follow the unittest pattern, but any executable can s…
avocado-vt0forkAvocado VT Plugin
freebsd-src0forkThe FreeBSD src tree publish-only repository. Experimenting with 'simple' pull requests....
libpldm0fork—
linux-mm0forkLinux kernel source tree
pldm0fork—
sev-apitest0活跃AMD SEV Apitest
tp-libvirt0forkTest Provider for Libvirt and related virtualization backends
tp-qemu0forkVirt Test Provider for qemu and other related virtualization backends
pensando:24 个仓库(AMD Pensando 网络与 DPU)

活跃非 fork 8 · fork 16 · 归档 0 · 按 Star 数排序

仓库Star状态官方简介
dsc-drivers8活跃—
dsc-tools4活跃—
pensando-elk4活跃ELK based analytics for Pensando Systems
dsc-linux-patches3活跃Patches to pensando/dsc-linux repo
goipmi3forkGo IPMI library
pypi3活跃—
trex-core2forktrex-core site
beats0fork:tropical_fish: Beats - Lightweight shippers for Elasticsearch & Logstash
box0forkA mruby-based Builder for Docker Images
diskimage-builder0forkImage building tools for OpenStack
docker-registry-client0forkA Go API client for the v2 Docker Registry API
dsc-artifacts0活跃—
external-qemu0活跃—
go-cisco-nx-api0forkCisco NX-OS API client library written in Go.
gogstash0forkLogstash like, written in golang
grpc0forkC++ based gRPC (C++, Python, Ruby, Objective-C, PHP, C#)
ipxe0活跃—
overmount0forkOverlayFS layer mounter used for creating OCI/Docker images - not ready for use
ovn0forkOpen Virtual Network
ovn-vif0fork—
ovs0forkOpen vSwitch
progress0forkProgress meter implemented as an io.Reader with channels
rebase-action0forkGitHub Action to automatically rebase PRs
tftp0forkTFTP server and client library for Golang
AMDEPYC:15 个仓库(AMD EPYC 平台软件)

活跃非 fork 9 · fork 2 · 归档 4 · 按 Star 数排序

仓库Star状态官方简介
kubernetes-power-manager5活跃Kubernetes Operator to provide cluster users with a mechanism to dynamically request adjustment of worker node power management settings
CENTOS-MILAN-SUPPORT-PATCHES4归档—
sev-certify3活跃—
sev_component_test3活跃—
dispatch2活跃Tool for dispatching EFI jobs from GitHub over HTTP boot
Host_Processor_Microbenchmarks2活跃—
kata-runtime-2.x-SEV2归档—
5GCore-UPF1活跃5G Core UPF Reference Architecture on AMD EPYC Platforms
linux1归档—
nri-plugins1forkA collection of community maintained NRI plugins
beacon0活跃workload communication client for dispatch orchestration
devpath0活跃UEFI Device Path parsing library
kata-runtime-1.x-SEV0归档—
kubernetes0forkProduction-Grade Container Scheduling and Management
uefi-reset0活跃UEFI module that resets the system when loaded
AMDComputeLibraries:6 个仓库(AMD 历史计算库)

活跃非 fork 5 · fork 1 · 归档 0 · 按 Star 数排序

仓库Star状态官方简介
morton_filter90活跃A compressed, sparse cuckoo filter (see https://www.vldb.org/pvldb/vol11/p1041-breslow.pdf)
ComputeApps25活跃Compute applications.
OpenMPApps13活跃OpenMP GPU Accelerated Applications
MultiGPUBandwidthTest5活跃MultiGPUBandwidthTest
caffe3forkCaffe: a fast open framework for deep learning.
Nekbone1活跃—
ROCm-LS:6 个仓库(ROCm Life Science(归档))

活跃非 fork 0 · fork 4 · 归档 6 · 按 Star 数排序

仓库Star状态官方简介
hipCIM6归档HIP port of the cuCIM library under the RAPIDS ecosystem.
monai3归档First release of ROCm enabled Monai
GPU_biology0归档 / forkAMD MI250x ready container recipes and test cases for biology tools
hipCIMwip0归档 / forkhipCIM - AMD ROCm fork of RAPIDS cucim
model-zoo0归档 / forkMONAI Model Zoo that hosts models in the MONAI Bundle format.
MONAIwip0归档 / forkAI Toolkit for Healthcare Imaging
ROCm-DS:1 个仓库(ROCm Data Science(归档))

活跃非 fork 0 · fork 1 · 归档 1 · 按 Star 数排序

仓库Star状态官方简介
rocOpt0归档 / forkGPU accelerated decision optimization
amd-enterprise-ai:6 个仓库(AMD 企业 AI 解决方案)

活跃非 fork 6 · fork 0 · 归档 0 · 按 Star 数排序

仓库Star状态官方简介
solution-blueprints22活跃—
aim-build18活跃—
amd-eai-apps14活跃—
aim-deploy7活跃—
aim-engine7活跃—
aim-fine-tuning0活跃AIM Fine-tuning offers tools and starting points to make fine-tuning easy to use, fast, and guaranteed to work in the context of AIMs.
AMD-AIM:21 个仓库(AMD AI Group, AIM:Radeon / Cloud / GPU MODE)

活跃非 fork 14 · fork 7 · 归档 0 · 按 Star 数排序

仓库Star状态官方简介
reference-kernels6forkOfficial Problem Sets / Reference Kernels for the GPU MODE Leaderboard!
inference-skill5活跃—
HIPGenerator3活跃—
ai_sprint_shanghai1活跃Radeon-adapted Triton workshop
InferenceX1forkOpen Source Continuous Inference Benchmarking Qwen3.5, DeepSeek, GPTOSS - GB200 NVL72 vs MI355X vs B200 vs GB300 NVL72 vs H100 & soon™ TPUv6e/v7/Trainium2/3
PaddleX1forkAll-in-One Development Tool based on PaddlePaddle
Physical_AI_Challenge1活跃—
radeon-cloud-docs1活跃Radeon Cloud documentation — user guide and API reference (EN / 简体中文)
Relax1forkAn Asynchronous Reinforcement Learning Engine for Omni-Modal Post-Training at Scale
sglang1forkSGLang is a fast serving framework for large language models and vision language models.
TransferQueue1forkThis is the official **live** mirror of https://gitcode.com/Ascend/TransferQueue. Feel free to contribute!
AMD_Developers_Notebooks0活跃Developer documentation and tutorials
Distill-run0活跃—
FlagTree0forkFlagTree is a unified compiler supporting multiple AI chip backends for custom Deep Learning operations, which is forked from triton-lang/triton.
hf-radeon-gpu-notebooks0活跃Validated Hugging Face model notebooks adapted for AMD Radeon GPUs and Radeon Cloud local inference.
hunyuan-models0活跃—
kfd-vgpu0活跃Driver-enforced multi-tenant GPU partitioning for consumer AMD GPUs (RDNA3/RDNA4). One physical card → N kernel-confined virtual cards, native to Kubernetes.
llm-eval-hub0活跃A self-hosted evaluation platform for OpenAI-compatible LLM APIs, providing automated accuracy, F1, exact match, latency, error analysis, and regression testing.
radeon-cloud-notebooks0活跃Hands-on notebooks for AMD Radeon Cloud on a single 48GB W7900D, adapted from AI Academy (https://github.com/AMD-ROCm-Internal/ai-academy-internal).
sglang-radeon0活跃Run SGLang on AMD Radeon RDNA3 (gfx110x) GPUs - an out-of-tree plugin
zhihu_rednote_articles0活跃source file of the articles that will be launched on zhihu and rednote

11.3 三个历史组织(归档)

ROCmSoftwarePlatform:9 个仓库(ROCm 旧组织名,仓库已迁移到 ROCm)

归档 9 · 按 Star 数排序

仓库Star状态官方简介
cub-hip6归档An implementation of CUB on the ROCM stack. This has been replaced by hipCUB.
MIOpenTensile3归档—
Anakin2归档 / fork—
roctracer1归档 / forkROCm Tracer Callback/Activity Library for Performance tracing AMD GPU's
.github0归档ROCm Software Platform organization README
actions0归档 / fork—
rfcs0归档 / forkPyTorch RFCs (experimental)
ROCm0归档 / forkROCm - Open Source Platform for HPC and Ultrascale GPU Computing
torchrec_private0归档—
RadeonOpenCompute:4 个仓库(ROCm 旧组织名,仓库已迁移到 ROCm)

归档 4 · 按 Star 数排序

仓库Star状态官方简介
ROCm_Documentation111归档Legacy ROCm Software Platform Documentation
rdma-perftest11归档—
rfcs2归档RFCs for changes Requests for ROCm
.github0归档ROCm GitHub Introduction
ROCm-Developer-Tools:1 个仓库(ROCm 旧开发者工具组织,仓库已迁移到 ROCm)

归档 1 · 按 Star 数排序

仓库Star状态官方简介
.github0归档—

12. 参考来源

官方文档

官方仓库