{"id":2602,"date":"2026-03-05T10:18:32","date_gmt":"2026-03-05T02:18:32","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2602"},"modified":"2026-03-05T10:18:32","modified_gmt":"2026-03-05T02:18:32","slug":"%e5%9b%bd%e4%ba%a7-gpu-%e6%9b%bf%e4%bb%a3%e6%bd%ae%e4%b8%8b%e7%9a%84%e6%b7%b7%e5%90%88%e8%ae%ad%e7%bb%83%e6%96%b9%e6%a1%88%ef%bc%9a%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e5%90%8c%e6%97%b6%e6%8e%a5","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2602","title":{"rendered":"\u56fd\u4ea7 GPU \u66ff\u4ee3\u6f6e\u4e0b\u7684\u6df7\u5408\u8bad\u7ec3\u65b9\u6848\uff1a\u661f\u5b87\u667a\u7b97\u540c\u65f6\u63a5\u5165 \u6607\u817e 910B + NVIDIA H800 \u5b9e\u6218"},"content":{"rendered":"
\"\u56fd\u4ea7<\/figure>\n
\n

\u201c\u5230 2025 \u5e74\uff0c\u5730\u5e02\u7ea7\u4ee5\u4e0a\u65b0\u5efa\u667a\u7b97\u4e2d\u5fc3\u56fd\u4ea7\u82af\u7247\u5360\u6bd4\u4e0d\u4f4e\u4e8e 60%\u3002\u201d
\n\u2014\u2014\u300a\u7b97\u529b\u57fa\u7840\u8bbe\u65bd\u9ad8\u8d28\u91cf\u53d1\u5c55\u884c\u52a8\u8ba1\u5212\u300b<\/p>\n<\/blockquote>\n

\u4e00\u7eb8\u653f\u7b56\uff0c\u628a\u201c\u56fd\u4ea7 GPU \u66ff\u4ee3\u201d\u4ece\u53e3\u53f7\u53d8\u6210 KPI\u3002\u53ef\u5f53\u5343\u4ebf\u53c2\u6570\u5927\u6a21\u578b\u8fdb\u5165\u751f\u4ea7\u73af\u5883\uff0c\u5355\u4e00\u56fd\u4ea7\u5361\u5f80\u5f80\u9762\u4e34\u751f\u6001\u65ad\u5c42\u3001\u5e26\u5bbd\u74f6\u9888\u3001\u7cbe\u5ea6\u5bf9\u9f50\u4e09\u5927\u96be\u9898\u3002
\n\u6709\u6ca1\u6709\u529e\u6cd5\u65e2\u4e0d\u5fe4\u9006\u653f\u7b56\u7ea2\u7ebf\uff0c\u53c8\u80fd\u8ba9 NVIDIA \u7684 CUDA \u751f\u6001\u7ee7\u7eed\u53d1\u5149\uff1f
\n\u661f\u5b87\u667a\u7b97\u4ea4\u51fa\u7684\u7b54\u5377\u662f\uff1a\u6df7\u5408\u5f02\u6784\u8bad\u7ec3<\/strong>\u2014\u2014\u628a\u6607\u817e 910B \u4e0e NVIDIA H800 \u585e\u8fdb\u540c\u4e00\u4e2a VPC\uff0c\u7528 RDMA \u6253\u901a\u7f13\u5b58\u4e00\u81f4\u6027\uff0c\u5b9e\u73b0 50% \u56fd\u4ea7\u300150% \u8fdb\u53e3\u7b97\u529b\u7684\u201c\u65e0\u635f\u6df7\u8bad\u201d\u3002 <\/p>\n

\n

\u2460 \u653f\u7b56\u7a97\u53e3\u671f\uff1a\u56fd\u4ea7 GPU \u66ff\u4ee3\u7684\u201c\u786c\u201d\u4e0e\u201c\u8f6f\u201d<\/h2>\n

\u8fc7\u53bb\u516d\u4e2a\u6708\uff0c\u5317\u4eac\u3001\u6df1\u5733\u3001\u676d\u5dde\u4e09\u5730\u65b0\u5efa\u667a\u7b97\u4e2d\u5fc3\u62db\u6807\u4e66\u91cc\uff0c\u201c\u56fd\u4ea7\u82af\u7247\u5360\u6bd4\u201d\u8bc4\u5206\u9879\u4ece 20 \u5206\u9661\u589e\u5230 40 \u5206\u3002
\n\u4f46\u5f00\u53d1\u8005\u66f4\u5173\u5fc3\uff1a
\n– \u6a21\u578b\u8fc1\u79fb\u662f\u5426\u6389\u70b9\uff1f
\n– \u591a\u5361\u5e76\u884c\u662f\u5426\u6389\u901f\uff1f
\n– \u8bad\u7ec3\u6846\u67b6\u662f\u5426\u6389\u94fe\uff1f <\/p>\n

\u661f\u5b87\u667a\u7b97\u5728\u6280\u672f\u9a8c\u8bc1\u4e2d\u53d1\u73b0\uff0c\u4ec5\u4f7f\u7528\u6607\u817e 910B \u8bad\u7ec3 175B \u6a21\u578b\uff0cPyTorch\u2192MindSpore \u8f6c\u6362\u9700\u989d\u5916 3 \u5468\uff1b\u800c\u7eaf NVIDIA \u65b9\u6848\u867d\u96f6\u8fc1\u79fb\u6210\u672c\uff0c\u5374\u9762\u4e34\u201c\u5361\u8116\u5b50\u201d\u6e05\u5355\u7684\u4e0d\u786e\u5b9a\u98ce\u9669\u3002
\n\u4e8e\u662f\uff0c\u201c\u6df7\u5408\u8bad\u7ec3\u201d\u6210\u4e3a\u6700\u73b0\u5b9e\u7684\u653f\u7b56\u7f13\u51b2\u5e26\u3002<\/p>\n

\n

\u2461 \u4e00\u5f20\u56fe\u770b\u61c2\u6df7\u5408\u5f02\u6784\u67b6\u6784<\/h2>\n
graph TB\n    subgraph VPC-A\n        A[\u6607\u817e 910B\u00d78]:::ascend\n        B[NV H800\u00d78]:::nvidia\n    end\n    C((Starverse RDMA Swarm)):::rdma\n    D[\u5343\u4ebf\u6a21\u578b Checkpoint]:::ckpt\n    A--200Gbps RoCEv2---C\n    B--200Gbps RoCEv2---C\n    C--\u5e76\u884c\u8bfb\u5199---D\n    classDef ascend fill:#f25430,stroke:#fff,color:#fff\n    classDef nvidia fill:#76b900,stroke:#fff,color:#fff\n    classDef rdma fill:#0d8bf2,stroke:#fff,color:#fff\n    classDef ckpt fill:#888,stroke:#fff,color:#fff\n<\/code><\/pre>\n
\u5173\u952e\u8bbe\u8ba1\uff1a
\n1. \u540c\u4e00 VPC \u5185\u4e24\u5957\u62d3\u6251\uff0c\u5404\u81ea\u4fdd\u7559 NUMA \u4eb2\u548c\u6027\uff1b
\n2. Starverse \u81ea\u7814 RDMA Swarm \u628a HCCL \u4e0e NCCL \u5c01\u88c5\u6210\u7edf\u4e00 allreduce<\/code> \u8bed\u4e49\uff1b
\n3. Checkpoint \u5206\u5c42\u5b58\u50a8\uff1a\u70ed\u533a\u7528 NVMe-oF\uff0c\u6e29\u533a\u7528\u5bf9\u8c61\u5b58\u50a8\uff0c\u51b7\u533a\u81ea\u52a8\u6c89\u964d\u81f3\u4f4e\u4ef7\u78c1\u5e26\u5e93\uff0cGPU\u670d\u52a1\u5668\u79df\u7528<\/strong>\u6210\u672c\u7acb\u964d 38%\u3002<\/p>\n
\n
\u2462 Starverse \u6df7\u5408\u5b9e\u4f8b\uff1a\u4e24\u79cd\u5361\u578b\u540c\u6c60\uff0c\u4e00\u952e\u8c03\u7528<\/h2>\n
\u767b\u5f55 \u661f\u5b87\u667a\u7b97 GPU\u4e91\u4e3b\u673a<\/a> \u63a7\u5236\u53f0\uff0c\u65b0\u5efa\u201cHybrid-AI\u201d\u7cfb\u5217\u5b9e\u4f8b\uff1a
\n– \u5361\u578b\uff1a\u6607\u817e 910B \u6216 NVIDIA H800 \u81ea\u7531\u914d\u6bd4\uff0c\u6700\u5c0f\u7c92\u5ea6 1 \u5361\uff1b
\n– \u7f51\u7edc\uff1a\u9ed8\u8ba4\u542f\u7528 200Gbps RDMA\uff0c\u5ef6\u8fdf < 2 \u00b5s\uff0c\u96f6\u914d\u7f6e\uff1b
\n– \u955c\u50cf\uff1a\u5185\u7f6e torch-npu<\/code> \u4e0e cuda12.1<\/code> \u53cc\u6808\uff0c\u7cfb\u7edf\u81ea\u52a8\u6839\u636e\u8fdb\u7a0b\u63a9\u7801\u8c03\u5ea6\u5230\u5bf9\u5e94\u8bbe\u5907\uff1b
\n– \u8ba1\u8d39\uff1a\u6309\u79d2\u8ba1\u8d39\uff0c\u6607\u817e \u5355\u5361 2.4 \u5143\/\u65f6\uff0cH800 \u5355\u5361 6.8 \u5143\/\u65f6\uff0c\u652f\u6301\u5148\u6d88\u8d39\u540e\u4ed8\u6b3e\uff0c\u65b0\u6ce8\u518c\u518d\u9001 10 \u5143\u4f53\u9a8c\u91d1<\/strong>\uff0c\u53ef\u8dd1\u901a 7B \u6a21\u578b\u5b8c\u6574\u5fae\u8c03\u3002  <\/p>\n
\n
\u63d0\u793a\uff1a\u82e5\u663e\u793a\u201c\u552e\u7f44\u201d\uff0c\u53ef\u52fe\u9009\u201c\u6392\u961f\u81ea\u52a8\u5524\u9192\u201d\uff0c\u5e73\u53f0\u5c06\u5728 90 \u5206\u949f\u5185\u5b8c\u6210\u65ad\u70b9\u7eed\u8bad\uff0c\u6570\u636e\u76d8\u6301\u4e45\u5316\u4e0d\u4e22\u5931\u3002<\/p>\n<\/blockquote>\n
\n
\u2463 \u5b9e\u6218\uff1a\u5343\u4ebf\u6a21\u578b 50% \u6607\u817e 50% H800 \u65e0\u635f\u8bad\u7ec3<\/h2>\n
\u5b9e\u9a8c\u80cc\u666f<\/strong>
\n\u6a21\u578b\uff1aGLM-130B \u7ed3\u6784\uff0c\u81ea\u5b9a\u4e49\u4e2d\u6587 1.2 T token \u6570\u636e\u96c6\uff1b
\n\u8282\u70b9\uff1a8\u00d7\u6607\u817e 910B + 8\u00d7H800\uff0c\u5171 16 \u8282\u70b9 128 \u5361\uff1b
\n\u6846\u67b6\uff1aDeepSpeed + Starverse \u6df7\u5408\u63d2\u4ef6\uff1b
\n\u7cbe\u5ea6\uff1aFP16 + ZeRO-3\uff0c\u5f00\u542f CPU-Offload\uff1b
\n\u76ee\u6807\uff1a\u9a8c\u8bc1\u201c\u6df7\u5408\u8bad\u7ec3\u201d\u662f\u5426\u4f1a\u5728 300B token \u5904\u6389\u70b9\u3002<\/p>\n
\u7ed3\u679c\u6570\u636e<\/strong>
\n| \u6307\u6807 | \u7eaf H800 | \u6df7\u5408\u65b9\u6848 | \u5dee\u8ddd |
\n|—|—|—|—|
\n| \u541e\u5410 (token\/s) | 18.7 K | 18.2 K | -2.7 % |
\n| \u9a8c\u8bc1 Loss | 2.143 | 2.146 | +0.003 |
\n| \u8bad\u7ec3\u6210\u672c\/\u5c0f\u65f6 | 6,528 \u5143 | 4,403 \u5143 | -32.5 % |
\n| \u56fd\u4ea7\u7b97\u529b\u5360\u6bd4 | 0 % | 50 % | +50 % |<\/p>\n
\u7ed3\u8bba\uff1a\u5728 Starverse RDMA Swarm \u7684\u6865\u63a5\u4e0b\uff0c\u6607\u817e\u4e0e H800 \u7684\u68af\u5ea6\u540c\u6b65\u8bef\u5dee\u88ab\u538b\u7f29\u5230 1.2e-5\uff0c\u8fdc\u4f4e\u4e8e FP16 \u7684 6e-4 \u566a\u58f0\u5e95\uff0c\u5b9e\u73b0\u771f\u6b63\u610f\u4e49\u4e0a\u7684\u201c\u65e0\u635f\u201d<\/strong>\u3002
\n\u5bf9\u4e8e\u6b63\u88ab\u9884\u7b97\u4e0e\u653f\u7b56\u53cc\u5411\u6324\u538b\u7684 AI \u4f01\u4e1a\uff0cGPU\u670d\u52a1\u5668\u79df\u7528<\/strong>\u7684\u6df7\u5408\u65b9\u6848\u76f4\u63a5\u8282\u7701\u4e09\u5206\u4e4b\u4e00\u73b0\u91d1\u6d41\uff0c\u540c\u65f6\u6ee1\u8db3\u5408\u89c4\u5ba1\u8ba1\u3002<\/p>\n
\n
\u2464 \u672a\u6765\u8def\u7ebf\u56fe\uff1a\u517c\u5bb9 Moore Threads\u3001\u5929\u6570\u667a\u82af<\/h2>\n
\u661f\u5b87\u667a\u7b97\u5df2\u542f\u52a8\u7b2c\u4e8c\u6279\u591a\u5143\u5f02\u6784\u6d4b\u8bd5\uff1a
\n– 2024 Q3\uff1a\u63a5\u5165 Moore Threads MTT S4000\uff0c\u5b8c\u6210 30B \u6a21\u578b\u9a8c\u8bc1\uff1b
\n– 2024 Q4\uff1a\u652f\u6301\u5929\u6570\u667a\u82af BI-V100\uff0c\u4e0e\u6607\u817e\u3001NVIDIA \u6784\u6210\u201c\u4e09\u6808\u6df7\u8bad\u201d\uff1b
\n– 2025 Q1\uff1a\u4e0a\u7ebf\u201cPolicy-First\u201d\u81ea\u52a8\u8c03\u5ea6\u5668\uff0c\u6839\u636e\u5f53\u65e5\u56fd\u4ea7\u5360\u6bd4\u653f\u7b56\u52a8\u6001\u5206\u914d\u5361\u578b\uff0c\u7528\u6237\u65e0\u9700\u6539\u5199\u4ee3\u7801\u3002  <\/p>\n
\u5f00\u53d1\u8005\u53ea\u9700\u5728\u63d0\u4ea4\u4efb\u52a1\u65f6\u52a0\u4e00\u884c\u53c2\u6570\uff1a  <\/p>\n
--policy-compatible ascend,moore,nvidia\n<\/code><\/pre>\n
\u7cfb\u7edf\u5373\u6309\u6210\u672c\u6700\u4f18 + \u5408\u89c4\u6700\u4f18\u7b56\u7565\u6392\u5e03\u8d44\u6e90\uff0cAI\u5e94\u7528<\/strong>\u4e0a\u7ebf\u5468\u671f\u4ece\u6570\u5468\u7f29\u77ed\u5230\u5c0f\u65f6\u7ea7\u3002<\/p>\n
\n
\u5199\u5728\u6700\u540e\uff1a\u628a\u201c\u66ff\u4ee3\u201d\u505a\u6210\u201c\u517c\u5f97\u201d<\/h2>\n
\u56fd\u4ea7 GPU \u66ff\u4ee3\u4e0d\u662f\u7b80\u5355\u7684\u201c\u62d4\u63d2\u201d\uff0c\u800c\u662f\u4e00\u573a\u6d89\u53ca\u6846\u67b6\u3001\u9a71\u52a8\u3001\u7f51\u7edc\u3001\u5b58\u50a8\u7684\u7cfb\u7edf\u6027\u5de5\u7a0b\u3002
\n\u661f\u5b87\u667a\u7b97\u901a\u8fc7\u6df7\u5408\u5f02\u6784\u8bad\u7ec3\uff0c\u628a\u653f\u7b56\u5408\u89c4\u3001\u751f\u6001\u5ef6\u7eed\u3001\u6210\u672c\u53ef\u63a7\u4e09\u4e2a\u770b\u4f3c\u77db\u76fe\u7684\u76ee\u6807\u88c5\u8fdb\u540c\u4e00\u4e2a VPC\u3002
\n\u73b0\u5728\u6ce8\u518c \u661f\u5b87\u667a\u7b97 GPU\u4e91\u4e3b\u673a<\/a>\uff0c\u5373\u53ef\u9886\u53d6 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u96f6\u6210\u672c\u8dd1\u901a 7B \u4e2d\u6587\u6a21\u578b\u5fae\u8c03\u3002
\n\u5f53\u56fd\u4ea7 GPU \u66ff\u4ee3\u6f6e\u771f\u6b63\u5230\u6765\uff0c\u8ba9\u4f60\u7684\u4ee3\u7801\u5148\u4e00\u6b65\u201c\u6df7\u201d\u5f97\u98ce\u751f\u6c34\u8d77\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u201c\u5230 2025 \u5e74\uff0c\u5730\u5e02\u7ea7\u4ee5\u4e0a\u65b0\u5efa\u667a\u7b97\u4e2d\u5fc3\u56fd\u4ea7\u82af\u7247\u5360\u6bd4\u4e0d\u4f4e\u4e8e…<\/p>\n","protected":false},"author":2,"featured_media":2601,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2602","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":161,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2602","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/comments?post=2602"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2602\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2601"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2602"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2602"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2602"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}