https://fengchen321.github.io/tags/architecture/fengchenHugo 0.139.0 & FixIt v0.3.15zh-CNThu, 23 Jan 2025 12:33:12 +0800https://fengchen321.github.io/posts/ai/gpu_%E7%A2%8E%E7%89%87%E7%AC%94%E8%AE%B0/Thu, 23 Jan 2025 12:33:12 +0800https://fengchen321.github.io/posts/ai/gpu_%E7%A2%8E%E7%89%87%E7%AC%94%E8%AE%B0/AI<h2 id="gpu_碎片笔记" class="heading-element"><span>GPU_碎片笔记</span> <a href="#gpu_%e7%a2%8e%e7%89%87%e7%ac%94%e8%ae%b0" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><h2 id="n卡--a卡--opencl-对比" class="heading-element"><span>N卡 / A卡 / OpenCL 对比</span> <a href="#n%e5%8d%a1--a%e5%8d%a1--opencl-%e5%af%b9%e6%af%94" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><table> <thead> <tr> <th style="text-align: left">Nvidia/CUDA</th> <th style="text-align: left">AMD/HIP</th> <th style="text-align: left">OpenCL</th> </tr> </thead> <tbody> <tr> <td style="text-align: left">Streaming Multiprocessor (SM)</td> <td style="text-align: left">Compute Unit (CU)</td> <td style="text-align: left">Compute Unit</td> </tr> <tr> <td style="text-align: left">Thread Block</td> <td style="text-align: left">Workgroup</td> <td style="text-align: left">Work-group</td> </tr> <tr> <td style="text-align: left">Shared Memory</td> <td style="text-align: left">Local Memory</td> <td style="text-align: left">Local Memory</td> </tr> <tr> <td style="text-align: left">Local Memory</td> <td style="text-align: left">Private Memory</td> <td style="text-align: left">Private Memory</td> </tr> <tr> <td style="text-align: left">grid</td> <td style="text-align: left">grid</td> <td style="text-align: left">NDRange</td> </tr> <tr> <td style="text-align: left">block</td> <td style="text-align: left">block</td> <td style="text-align: left">work-group</td> </tr> <tr> <td style="text-align: left">thread</td> <td style="text-align: left">work-item / thread</td> <td style="text-align: left">work-item</td> </tr> <tr> <td style="text-align: left">warp (32)</td> <td style="text-align: left">wavefront (64)</td> <td style="text-align: left">sub-group</td> </tr> </tbody> </table> <h2 id="硬件规格" class="heading-element"><span>硬件规格</span> <a href="#%e7%a1%ac%e4%bb%b6%e8%a7%84%e6%a0%bc" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><h3 id="fermi-n卡老架构" class="heading-element"><span>Fermi (N卡老架构)</span> <a href="#fermi-n%e5%8d%a1%e8%80%81%e6%9e%b6%e6%9e%84" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h3><ul> <li>SM 数：16</li> <li>每 SM 的 CUDA 核心数：32</li> <li>Warp 大小：32</li> <li>每 SM 最大线程：1536</li> </ul> <h3 id="gfx906-amddcu-典型" class="heading-element"><span>gfx906 (AMD/DCU 典型)</span> <a href="#gfx906-amddcu-%e5%85%b8%e5%9e%8b" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h3><ul> <li>CU 数：64</li> <li>SE：4</li> <li>每 CU 的 SIMD：4</li> <li>每 SIMD wavefront 数量：10</li> <li>Wavefront 大小：64</li> <li>单 block 最大线程：1024 (即最多 16 个 wavefront)</li> <li>SGPR(标量寄存器)：每组 16 个，编程时可用的 SGPR 为 0～101，共 102 个</li> <li>VGPR(向量寄存器)：分配粒度为 4 个寄存器一组</li> </ul> <h3 id="bw" class="heading-element"><span>BW</span> <a href="#bw" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h3><ul> <li>CU 数：80</li> <li>SE：8</li> </ul> <p>峰值：$cu数 \times simds数\times频率\times数据布局\times2 /时钟周期$</p>