{"id":535,"date":"2025-05-16T11:32:11","date_gmt":"2025-05-16T03:32:11","guid":{"rendered":"https:\/\/aitimes.link\/?p=535"},"modified":"2025-05-16T11:32:11","modified_gmt":"2025-05-16T03:32:11","slug":"insights-into-deepseek-v3-scaling-challenges-and-reflections-onhardware-for-ai-architectures","status":"publish","type":"post","link":"https:\/\/aitimes.link\/index.php\/2025\/05\/16\/insights-into-deepseek-v3-scaling-challenges-and-reflections-onhardware-for-ai-architectures\/","title":{"rendered":"Insights into DeepSeek-V3: Scaling Challenges and Reflections onHardware for AI Architectures"},"content":{"rendered":"\n<p><a href=\"https:\/\/www.arxiv.org\/pdf\/2505.09343\">https:\/\/www.arxiv.org\/pdf\/2505.09343<\/a><\/p>\n\n\n\n<p>\u6df1\u5ea6\u89e3\u6790 DeepSeek-V3\uff1a\u6269\u5c55\u6311\u6218\u4e0e\u786c\u4ef6\u67b6\u6784\u53cd\u601d<br>1. \u80cc\u666f\u4e0e\u76ee\u6807<\/p>\n\n\n\n<p><br>DeepSeek-V3 \u662f\u7531&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=DeepSeek-AI&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek-AI<\/a>&nbsp;\u5f00\u53d1\u7684\u8d85\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff0c\u57fa\u4e8e\u5176\u524d\u4ee3 DeepSeek-V2 \u7684\u67b6\u6784\u8fdb\u884c\u4e86\u591a\u65b9\u9762\u7684\u4f18\u5316\u3002\u8bba\u6587\u7684\u4e3b\u8981\u76ee\u6807\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5206\u6790\u5927\u89c4\u6a21 LLM \u5728\u8bad\u7ec3\u548c\u63a8\u7406\u4e2d\u9762\u4e34\u7684<strong>\u786c\u4ef6\u74f6\u9888<\/strong>\u3002<\/li>\n\n\n\n<li>\u63a2\u8ba8\u6a21\u578b\u4e0e<strong>\u786c\u4ef6\u534f\u540c\u8bbe\u8ba1\uff08co-design\uff09<\/strong>\u7684\u5fc5\u8981\u6027\u3002<\/li>\n\n\n\n<li>\u63d0\u51fa\u9762\u5411\u672a\u6765\u7684<strong>AI \u786c\u4ef6\u67b6\u6784\u5efa\u8bae<\/strong>\uff0c\u4ee5\u652f\u6301\u66f4\u9ad8\u6548\u7684 LLM \u6269\u5c55\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8bba\u6587\u4e3b\u8981\u805a\u7126\u4e8e\u5176<strong>\u7cfb\u7edf\u7ea7\u67b6\u6784\u8bbe\u8ba1<\/strong>\u3001<strong>\u901a\u4fe1\u673a\u5236\u4f18\u5316<\/strong>\u3001<strong>\u7f51\u7edc\u62d3\u6251\u7ed3\u6784\u6539\u8fdb<\/strong>\uff0c\u4ee5\u53ca\u8fd9\u4e9b\u56e0\u7d20\u5982\u4f55\u5f71\u54cd\u6a21\u578b\u7684\u6027\u80fd\u548c\u6210\u672c\u6548\u7387\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">2.&nbsp;<strong>DeepSeek-V3 \u7684\u6838\u5fc3\u67b6\u6784\u8bbe\u8ba1<\/strong><\/h2>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1032\" height=\"747\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-6768971795db0886e4b6561d6d0d042f_1440w.jpg\" alt=\"\" class=\"wp-image-542\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-6768971795db0886e4b6561d6d0d042f_1440w.jpg 1032w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-6768971795db0886e4b6561d6d0d042f_1440w-300x217.jpg 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-6768971795db0886e4b6561d6d0d042f_1440w-1024x741.jpg 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-6768971795db0886e4b6561d6d0d042f_1440w-768x556.jpg 768w\" sizes=\"auto, (max-width: 1032px) 100vw, 1032px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">2.1&nbsp;<strong>\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b\u673a\u5236\uff08Multi-head Latent Attention,&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=MLA&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">MLA<\/a>\uff09<\/strong><\/h3>\n\n\n\n<p><br>MLA \u662f\u4e00\u79cd\u65e8\u5728\u63d0\u5347\u5185\u5b58\u6548\u7387\u7684\u6ce8\u610f\u529b\u673a\u5236\u3002\u4f20\u7edf\u7684&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=Transformer+%E6%B3%A8%E6%84%8F%E5%8A%9B%E6%9C%BA%E5%88%B6&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">Transformer \u6ce8\u610f\u529b\u673a\u5236<\/a>\u9700\u8981\u5b58\u50a8\u5927\u91cf\u7684 Key\/Value \u7f13\u5b58\uff08KV Cache\uff09\uff0c\u8fd9\u5728\u5904\u7406\u957f\u5e8f\u5217\u65f6\u4f1a\u9020\u6210\u663e\u8457\u7684\u5185\u5b58\u8d1f\u62c5\u3002MLA \u5f15\u5165\u4e86\u201c\u6f5c\u5728\u8868\u793a\u201d\uff08latent representation\uff09\uff0c\u901a\u8fc7\u538b\u7f29 KV \u7f13\u5b58\u6765\u51cf\u5c11\u5185\u5b58\u5360\u7528\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1174\" height=\"416\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-25a11b8a869ceb729533a9ada0229ecf_1440w.jpg\" alt=\"\" class=\"wp-image-543\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-25a11b8a869ceb729533a9ada0229ecf_1440w.jpg 1174w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-25a11b8a869ceb729533a9ada0229ecf_1440w-300x106.jpg 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-25a11b8a869ceb729533a9ada0229ecf_1440w-1024x363.jpg 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-25a11b8a869ceb729533a9ada0229ecf_1440w-768x272.jpg 768w\" sizes=\"auto, (max-width: 1174px) 100vw, 1174px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">2.2&nbsp;<strong>\u6df7\u5408\u4e13\u5bb6\u67b6\u6784\uff08Mixture of Experts,&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=MoE&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">MoE<\/a>\uff09\u4f18\u5316<\/strong><\/h3>\n\n\n\n<p><br>MoE \u662f DeepSeek-V3 \u7684\u6838\u5fc3\u6280\u672f\u4e4b\u4e00\uff0c\u5b83\u901a\u8fc7\u7a00\u758f\u6fc0\u6d3b\u673a\u5236\uff0c\u5728\u4e0d\u589e\u52a0\u8ba1\u7b97\u8d44\u6e90\u7684\u524d\u63d0\u4e0b\u663e\u8457\u63d0\u5347\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1176\" height=\"432\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-4a3cb672874945ad3ced4b0ee38f8455_1440w.jpg\" alt=\"\" class=\"wp-image-540\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-4a3cb672874945ad3ced4b0ee38f8455_1440w.jpg 1176w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-4a3cb672874945ad3ced4b0ee38f8455_1440w-300x110.jpg 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-4a3cb672874945ad3ced4b0ee38f8455_1440w-1024x376.jpg 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-4a3cb672874945ad3ced4b0ee38f8455_1440w-768x282.jpg 768w\" sizes=\"auto, (max-width: 1176px) 100vw, 1176px\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u53c2\u6570\u89c4\u6a21\u5927\u4f46\u8ba1\u7b97\u6210\u672c\u4f4e<\/strong>\uff1a\u4f8b\u5982\uff0cDeepSeek-V3 \u603b\u53c2\u6570\u8fbe\u5230 671B\uff0c\u4f46\u6bcf\u4e2a token \u53ea\u6fc0\u6d3b\u7ea6 37B \u53c2\u6570\u3002<\/li>\n\n\n\n<li><strong>\u7075\u6d3b\u7684\u8def\u7531\u7b56\u7565<\/strong>\uff1a\u91c7\u7528 Top-K \u8def\u7531\u7b97\u6cd5\u9009\u62e9\u6700\u76f8\u5173\u7684\u4e13\u5bb6\u6a21\u5757\u3002<\/li>\n\n\n\n<li><strong>\u5206\u5e03\u5f0f\u8bad\u7ec3\u4f18\u5316<\/strong>\uff1a\u7ed3\u5408 All-to-All \u901a\u4fe1\u6a21\u5f0f\uff0c\u5b9e\u73b0\u9ad8\u6548\u7684\u4e13\u5bb6\u5e76\u884c\uff08Expert Parallelism, EP\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5b9e\u73b0\u96be\u70b9\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u901a\u4fe1\u5f00\u9500\u5927<\/strong>\uff1aEP \u9700\u8981\u9891\u7e41\u7684 All-to-All \u6570\u636e\u4ea4\u6362\u3002<\/li>\n\n\n\n<li><strong>\u8d1f\u8f7d\u4e0d\u5747\u8861<\/strong>\uff1a\u67d0\u4e9b\u4e13\u5bb6\u53ef\u80fd\u88ab\u9891\u7e41\u8c03\u7528\uff0c\u9020\u6210\u70ed\u70b9\u95ee\u9898\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4e3a\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0cDeepSeek \u56e2\u961f\u5f00\u53d1\u4e86&nbsp;<strong>DeepEP<\/strong>\uff0c\u4e00\u4e2a\u9ad8\u6548\u7684 EP \u5b9e\u73b0\u5e93\uff0c\u5e76\u5f00\u6e90\u4f9b\u793e\u533a\u4f7f\u7528\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><br>2.3&nbsp;<strong>\u591a\u4ee4\u724c\u9884\u6d4b\u6a21\u5757\uff08Multi-Token Prediction Module,&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=MTP&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">MTP<\/a>\uff09<\/strong><\/h3>\n\n\n\n<p><br>MTP \u662f\u4e3a\u4e86\u52a0\u901f\u63a8\u7406\u800c\u8bbe\u8ba1\u7684\u6a21\u5757\uff0c\u5141\u8bb8\u6a21\u578b\u5728\u540c\u4e00\u65f6\u95f4\u6b65\u9884\u6d4b\u591a\u4e2a token\uff0c\u4ece\u800c\u663e\u8457\u63d0\u9ad8\u541e\u5410\u91cf\u3002<br>\u5de5\u4f5c\u539f\u7406\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728\u89e3\u7801\u9636\u6bb5\uff0c\u540c\u65f6\u751f\u6210\u591a\u4e2a\u5019\u9009 token\u3002<\/li>\n\n\n\n<li>\u5229\u7528 speculative sampling \u7b56\u7565\uff0c\u63d0\u524d\u9884\u6d4b\u540e\u7eed token\u3002<\/li>\n\n\n\n<li>\u7ed3\u5408\u786c\u4ef6\u52a0\u901f\u5668\uff08\u5982 GPU \u7684 Tensor Core\uff09\u63d0\u5347\u5e76\u884c\u5904\u7406\u80fd\u529b\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6548\u679c\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u663e\u8457\u51cf\u5c11\u63a8\u7406\u5ef6\u8fdf\u3002<\/li>\n\n\n\n<li>\u5728\u6d88\u8d39\u7ea7 GPU \u4e0a\u4e5f\u53ef\u5b9e\u73b0\u9ad8\u541e\u5410\uff08\u5982\u8fd0\u884c DeepSeek-V3 \u7684 TPS \u8fbe\u5230 ~20\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">3.&nbsp;<strong>\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u4e0e\u901a\u4fe1\u4f18\u5316<\/strong><\/h2>\n\n\n\n<p><br>\u968f\u7740 LLM \u7684\u53c2\u6570\u89c4\u6a21\u4e0d\u65ad\u589e\u957f\uff0c\u4f20\u7edf\u9ad8\u7cbe\u5ea6\u6d6e\u70b9\u8fd0\u7b97\uff08\u5982 FP32 \u6216 BF16\uff09\u5728\u8ba1\u7b97\u8d44\u6e90\u3001\u5185\u5b58\u5e26\u5bbd\u548c\u80fd\u8017\u65b9\u9762\u5e26\u6765\u4e86\u5de8\u5927\u538b\u529b\u3002\u5c24\u5176\u662f\u5728 MoE \u67b6\u6784\u4e2d\uff0c\u4e13\u5bb6\u6a21\u5757\u4e4b\u95f4\u7684\u9891\u7e41\u901a\u4fe1\u8fdb\u4e00\u6b65\u52a0\u5267\u4e86\u5bf9\u5e26\u5bbd\u7684\u9700\u6c42\u3002<br>MoE \u4e2d\u6709\u4e24\u4e2a\u5173\u952e\u9636\u6bb5\u6d89\u53ca\u5927\u91cf\u901a\u4fe1\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Dispatch \u9636\u6bb5<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u8f93\u5165 token \u88ab\u8def\u7531\u5230\u591a\u4e2a\u4e13\u5bb6\u6a21\u5757\u3002<\/li>\n\n\n\n<li>\u4f7f\u7528 FP8 \u6216\u00a0<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=LogFMT&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">LogFMT<\/a>\u00a0\u538b\u7f29\u540e\uff0c\u4f20\u8f93\u6570\u636e\u91cf\u51cf\u5c11\uff0c\u901a\u4fe1\u65f6\u95f4\u663e\u8457\u7f29\u77ed\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>Combine \u9636\u6bb5<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u6765\u81ea\u591a\u4e2a\u4e13\u5bb6\u7684\u8f93\u51fa\u88ab\u805a\u5408\u3002<\/li>\n\n\n\n<li>\u5982\u679c\u5728\u6b64\u9636\u6bb5\u4e5f\u4f7f\u7528\u538b\u7f29\u901a\u4fe1\uff0c\u53ef\u4ee5\u907f\u514d\u5e26\u5bbd\u74f6\u9888\uff0c\u63d0\u5347\u805a\u5408\u6548\u7387\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>DeepSeek \u56e2\u961f\u63d0\u51fa\u4e86&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=FP8+%E6%B7%B7%E5%90%88%E7%B2%BE%E5%BA%A6%E8%AE%AD%E7%BB%83&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">FP8 \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3<\/a>&nbsp;\u548c LogFMT \u901a\u4fe1\u538b\u7f29\u683c\u5f0f \u7b49\u521b\u65b0\u6280\u672f\uff0c\u6784\u5efa\u4e86\u4e00\u4e2a\u4ee5\u4f4e\u7cbe\u5ea6\u4e3a\u6838\u5fc3\u7684\u7cfb\u7edf\u7ea7\u4f18\u5316\u6846\u67b6\uff0c\u4ece\u800c\u5728\u4e0d\u727a\u7272\u6a21\u578b\u6027\u80fd\u7684\u524d\u63d0\u4e0b\uff0c\u663e\u8457\u63d0\u5347\u4e86\u8bad\u7ec3\u548c\u63a8\u7406\u6548\u7387\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><br>3.1&nbsp;<strong>FP8 \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3<\/strong><\/h3>\n\n\n\n<p><br>DeepSeek-V3 \u91c7\u7528\u4e86 FP8 \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u6280\u672f\uff0c\u4ee5\u964d\u4f4e\u8ba1\u7b97\u548c\u901a\u4fe1\u5e26\u5bbd\u9700\u6c42\u3002<br>\u5b9e\u73b0\u65b9\u5f0f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728\u5173\u952e\u8ba1\u7b97\u8def\u5f84\uff08\u5982\u6ce8\u610f\u529b\u77e9\u9635\u8ba1\u7b97\u3001FFN \u5c42\uff09\u4f7f\u7528 FP8\u3002<\/li>\n\n\n\n<li>\u5176\u4ed6\u90e8\u5206\u4fdd\u7559\u66f4\u9ad8\u7cbe\u5ea6\uff08\u5982 BF16\uff09\u4ee5\u4fdd\u6301\u6570\u503c\u7a33\u5b9a\u6027\u3002<\/li>\n\n\n\n<li>\u52a8\u6001\u8c03\u6574\u7cbe\u5ea6\u7b56\u7565\uff0c\u6839\u636e\u5c42\u6216\u64cd\u4f5c\u7c7b\u578b\u5207\u6362\u7cbe\u5ea6\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4f18\u70b9\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u51cf\u5c11\u5185\u5b58\u5360\u7528\u548c\u6570\u636e\u4f20\u8f93\u5e26\u5bbd\u3002<\/li>\n\n\n\n<li>\u52a0\u5feb\u8bad\u7ec3\u901f\u5ea6\uff0c\u5c24\u5176\u5728\u5e26\u5bbd\u53d7\u9650\u7684\u96c6\u7fa4\u73af\u5883\u4e2d\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5c3d\u7ba1 FP8 \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u5e26\u6765\u4e86\u663e\u8457\u6536\u76ca\uff0c\u4f46\u4ecd\u5b58\u5728\u4e00\u4e9b\u9650\u5236\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7d2f\u52a0\u7cbe\u5ea6\u4e0d\u8db3 Hopper GPU \u7684 Tensor Core \u5728\u6267\u884c FP8 GEMM \u65f6\u4ec5\u7ef4\u62a4 13 \u4f4d\u5c3e\u6570\uff08FP22\uff09\uff0c\u5e76\u622a\u65ad\u8d85\u51fa\u90e8\u5206\u3002\u8fd9\u53ef\u80fd\u5bfc\u81f4\uff1a\n<ul class=\"wp-block-list\">\n<li>\u957f\u5e8f\u5217\u6ce8\u610f\u529b\u8ba1\u7b97\u4e2d\u7684\u8bef\u5dee\u7d2f\u79ef\u3002<\/li>\n\n\n\n<li>\u6743\u91cd\u66f4\u65b0\u4e0d\u7a33\u5b9a\uff0c\u5f71\u54cd\u6536\u655b\u901f\u5ea6\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u6570\u503c\u7a33\u5b9a\u6027\u95ee\u9898 FP8 \u7684\u52a8\u6001\u8303\u56f4\u6709\u9650\uff0c\u5c24\u5176\u5728\u67d0\u4e9b\u6781\u7aef\u8f93\u5165\u573a\u666f\u4e0b\u53ef\u80fd\u51fa\u73b0\u6ea2\u51fa\u6216\u4e0b\u6ea2\uff0c\u5f71\u54cd\u6a21\u578b\u7a33\u5b9a\u6027\u3002<\/li>\n\n\n\n<li>\u8f6f\u4ef6\u652f\u6301\u4e0d\u5b8c\u5584 \u76ee\u524d\u4e3b\u6d41\u6846\u67b6\uff08\u5982 PyTorch\uff09\u5bf9 FP8 \u7684\u652f\u6301\u4ecd\u5904\u4e8e\u65e9\u671f\u9636\u6bb5\uff0c\u9700\u5927\u91cf\u81ea\u5b9a\u4e49\u4ee3\u7801\u5b9e\u73b0\uff1a\n<ul class=\"wp-block-list\">\n<li>\u81ea\u5b9a\u4e49\u91cf\u5316\/\u53cd\u91cf\u5316\u51fd\u6570\u3002<\/li>\n\n\n\n<li>\u81ea\u5b9a\u4e49 GEMM \u5185\u6838\uff08\u5982 DeepGEMM \u5f00\u6e90\u9879\u76ee\uff09\u3002<\/li>\n\n\n\n<li>\u6df7\u5408\u7cbe\u5ea6\u8c03\u5ea6\u903b\u8f91\u590d\u6742\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u672a\u6765\u65b9\u5411\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e3a FP8 \u6216\u81ea\u5b9a\u4e49\u7cbe\u5ea6\u683c\u5f0f\u63d0\u4f9b\u539f\u751f\u538b\u7f29\/\u89e3\u538b\u5355\u5143 \uff0c\u662f\u89e3\u51b3\u5f53\u524d AI \u901a\u4fe1\u74f6\u9888\u7684\u91cd\u8981\u65b9\u5411\u3002<\/li>\n\n\n\n<li>\u5b83\u80fd\u663e\u8457\u964d\u4f4e\u901a\u4fe1\u5e26\u5bbd\u9700\u6c42 \uff0c\u5e76\u63d0\u5347\u901a\u4fe1\u6548\u7387 \uff0c\u5c24\u5176\u9002\u7528\u4e8e MoE \u8fd9\u7c7b\u901a\u4fe1\u5bc6\u96c6\u578b\u4efb\u52a1\u3002<\/li>\n\n\n\n<li>\u672a\u6765\u786c\u4ef6\u5e94\u8003\u8651\u5c06\u8fd9\u4e9b\u529f\u80fd\u96c6\u6210\u5230\u7f51\u7edc\u63a5\u53e3\u63a7\u5236\u5668\uff08NIC\uff09\u3001I\/O die \u6216 GPU \u5185\u90e8\u903b\u8f91\u4e2d \uff0c\u4ee5\u5b9e\u73b0\u771f\u6b63\u7684\u9ad8\u6548\u901a\u4fe1<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 LogFMT\u901a\u8baf\u538b\u7f29<\/h3>\n\n\n\n<p><br>LogFMT \u662f DeepSeek-V3 \u63d0\u51fa\u7684\u4e00\u79cd\u57fa\u4e8e\u5bf9\u6570\u7a7a\u95f4\u7684\u65b0\u578b\u91cf\u5316\u683c\u5f0f\uff0c\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u662f\u4e3a\u4e86\u5f25\u8865 FP8 \u7684\u4e0d\u8db3\u800c\u63d0\u51fa\u7684 \u3002\u5176\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u6570\u503c\u4ece\u7ebf\u6027\u7a7a\u95f4\u8f6c\u6362\u5230\u5bf9\u6570\u7a7a\u95f4\uff0c\u4ece\u800c\u4f7f\u5f97\u6570\u636e\u5206\u5e03\u66f4\u5747\u5300\uff0c\u4fbf\u4e8e\u8fdb\u884c\u4f4e\u6bd4\u7279\u91cf\u5316\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"745\" height=\"274\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-312c8f3e1cd3bc8069ed3139c1868135_1440w.jpg\" alt=\"\" class=\"wp-image-536\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-312c8f3e1cd3bc8069ed3139c1868135_1440w.jpg 745w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-312c8f3e1cd3bc8069ed3139c1868135_1440w-300x110.jpg 300w\" sizes=\"auto, (max-width: 745px) 100vw, 745px\" \/><\/figure>\n\n\n\n<p><br>\u6570\u636e\u7ed3\u6784\u5b9a\u4e49\uff08LogFMT-nBit\uff09\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7b2c\u4e00\u4f4d\u4e3a\u7b26\u53f7\u4f4d S\u3002<\/li>\n\n\n\n<li>\u5269\u4f59 n-1 \u4f4d\u7f16\u7801 log(|xi|) \u7684\u503c\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6bcf\u4e2a tile \u72ec\u7acb\u8ba1\u7b97 min\/max\uff0c\u5e76\u636e\u6b64\u5f52\u4e00\u5316\u5176\u4ed6\u503c\uff0c\u5b9e\u73b0\u81ea\u9002\u5e94\u91cf\u5316\u8303\u56f4\u3002<br>\u4e3b\u8981\u53ef\u4ee5\u5e94\u7528\u5728\u4ee5\u4e0b\u51e0\u4e2a\u9636\u6bb5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>MoE \u7684 Dispatch \u9636\u6bb5 \uff1a\u4ee4\u724c\u88ab\u8def\u7531\u524d\u8fdb\u884c\u4f4e\u7cbe\u5ea6\u7f16\u7801\u3002<\/li>\n\n\n\n<li>KV Cache \u538b\u7f29 \uff1a\u7f13\u5b58 Key\/Value \u5411\u91cf\uff0c\u51cf\u5c11\u5185\u5b58\u5360\u7528\u3002<\/li>\n\n\n\n<li>\u4e2d\u95f4\u7ed3\u679c\u901a\u4fe1 \uff1aTransformer \u5c42\u4e4b\u95f4\u4f20\u9012\u6fc0\u6d3b\u503c\u65f6\u4f7f\u7528\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5b58\u5728\u7684\u5c40\u9650:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5f53\u524d GPU \u4e0d\u652f\u6301 LogFMT \u539f\u751f\u8fd0\u7b97\uff0c\u9700\u8f6f\u4ef6\u5c42\u5b9e\u73b0\uff0c\u5e26\u6765\u989d\u5916\u5ef6\u8fdf\u3002<\/li>\n\n\n\n<li>\u7f16\u89e3\u7801\u8fc7\u7a0b\u8f83\u590d\u6742\uff0c\u589e\u52a0\u4e86\u8ba1\u7b97\u8d1f\u62c5\u3002<\/li>\n\n\n\n<li>\u76ee\u524d\u4e3b\u8981\u7528\u4e8e\u901a\u4fe1\u538b\u7f29\uff0c\u5728\u5173\u952e\u8ba1\u7b97\u8def\u5f84\uff08\u5982\u6ce8\u610f\u529b\u77e9\u9635\u3001FFN \u5c42\uff09\u4ecd\u9700\u4f7f\u7528 FP8\/BF16\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u672a\u6765\u65b9\u5411\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u786c\u4ef6\u539f\u751f\u652f\u6301 \uff1a\u5efa\u8bae\u672a\u6765 GPU \u548c\u7f51\u7edc\u82af\u7247\u652f\u6301 LogFMT \u7f16\u89e3\u7801\u529f\u80fd\u3002<\/li>\n\n\n\n<li>\u6df7\u5408\u7cbe\u5ea6\u7b56\u7565 \uff1a\u7ed3\u5408 LogFMT \u4e0e FP8\/BF16\uff0c\u5728\u4e0d\u540c\u9636\u6bb5\u7075\u6d3b\u5207\u6362\u3002<\/li>\n\n\n\n<li>\u81ea\u52a8 rounding \u5230\u7ebf\u6027\u7a7a\u95f4 \uff1a\u786e\u4fdd\u91cf\u5316\u8fc7\u7a0b\u65e0\u504f\uff0c\u907f\u514d\u5728\u5bf9\u6570\u7a7a\u95f4\u76f4\u63a5\u820d\u5165\u9020\u6210\u7684\u8bef\u5dee\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">4.&nbsp;<strong>\u4e92\u8054\u7f51\u7edc\u4f18\u5316\u4e0e Scale-up\/Scale-out \u67b6\u6784\u878d\u5408<\/strong><\/h3>\n\n\n\n<p><br>DeepSeek-V3 \u57fa\u4e8e NVIDIA H800 GPU \u6784\u5efa\uff0c\u5e76\u91c7\u7528\u9ad8\u6027\u80fd\u4e92\u8fde\u6280\u672f\uff08\u5982&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=InfiniBand&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">InfiniBand<\/a>&nbsp;\u548c&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=NVLink&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">NVLink<\/a>\uff09\u8fdb\u884c\u8282\u70b9\u95f4\u901a\u4fe1\uff1b<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"501\" height=\"348\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-24b17965fdeb2f001068e76dbd04fa24_1440w.jpg\" alt=\"\" class=\"wp-image-537\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-24b17965fdeb2f001068e76dbd04fa24_1440w.jpg 501w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-24b17965fdeb2f001068e76dbd04fa24_1440w-300x208.jpg 300w\" sizes=\"auto, (max-width: 501px) 100vw, 501px\" \/><\/figure>\n\n\n\n<p><br>\u7f51\u7edc\u62d3\u6251\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u4e2a H800 \u8282\u70b9\u914d\u5907 8 \u4e2a 400Gbps InfiniBand (IB) NIC \u3002<\/li>\n\n\n\n<li>\u5b9e\u73b0\u4e86\u63a5\u8fd1 50GB\/s \u7684\u6709\u6548\u5e26\u5bbd \uff08\u8003\u8651\u5c0f\u6570\u636e\u5305\u548c\u5ef6\u8fdf\u5f71\u54cd\u540e\u5b9e\u9645\u7ea6\u4e3a 40GB\/s\uff09\u3002<\/li>\n\n\n\n<li>\u76f8\u6bd4\u4e4b\u4e0b\uff0cNVLink \u63d0\u4f9b\u7ea6 160GB\/s \u7684\u5e26\u5bbd \uff08\u5355\u5411\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4e3a\u6ee1\u8db3\u76d1\u7ba1\u8981\u6c42\uff0cFP64 \u6027\u80fd\u548c NVLink \u5e26\u5bbd\u88ab\u524a\u51cf\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>NVLink \u5e26\u5bbd\u4ece 900 GB\/s \u964d\u81f3 400 GB\/s\uff0c\u5bfc\u81f4\u8282\u70b9\u5185\u901a\u4fe1\u80fd\u529b\u4e0b\u964d<\/li>\n<\/ul>\n\n\n\n<p>\u4e3a\u6b64\uff0c \u6bcf\u4e2a\u8282\u70b9\u914d\u5907 8 \u4e2a 400Gbps InfiniBand NIC\uff0c\u589e\u5f3a\u8282\u70b9\u95f4\u901a\u4fe1\u80fd\u529b\uff0c\u540c\u65f6DeepSeek-V3 \u901a\u8fc7\u8f6f\u4ef6\u4f18\u5316\uff08\u5982&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=Node-Limited+Routing&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">Node-Limited Routing<\/a>\u3001LogFMT \u7b49\uff09\u5f25\u8865\u786c\u4ef6\u9650\u5236\uff0c\u5b9e\u73b0\u9ad8\u6548\u8bad\u7ec3\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><br>4.1 Hardware-Aware Parallelism(\u786c\u4ef6\u611f\u77e5\u5e76\u884c)<\/h3>\n\n\n\n<p><br>DeepSeek \u56e2\u961f\u63d0\u51fa\u4e86 Hardware-Aware Parallelism \u7b56\u7565\uff0c\u65e8\u5728\u6839\u636e\u786c\u4ef6\u7279\u6027\u4f18\u5316\u6a21\u578b\u5e76\u884c\u65b9\u5f0f\uff0c\u4ee5\u63d0\u5347\u6574\u4f53\u8bad\u7ec3\u548c\u63a8\u7406\u6548\u7387\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td>\u5e76\u884c\u7c7b\u578b<\/td><td>\u662f\u5426\u4f7f\u7528<\/td><td>\u4f7f\u7528\u573a\u666f<\/td><td>\u8bf4\u660e<\/td><\/tr><tr><td><a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=Tensor+Parallelism&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">Tensor Parallelism<\/a>&nbsp;(TP)<\/td><td>\u5426\uff08\u8bad\u7ec3\u65f6\uff09<br>\u662f\uff08\u63a8\u7406\u65f6\uff09<\/td><td>\u63a8\u7406\u52a0\u901f<\/td><td>\u907f\u514d\u5728\u8bad\u7ec3\u4e2d\u4f7f\u7528\uff0c\u56e0 NVLink \u5e26\u5bbd\u9650\u5236\u5bfc\u81f4\u4f4e\u6548\uff1b\u63a8\u7406\u65f6\u9009\u62e9\u6027\u4f7f\u7528\u4ee5\u964d\u4f4e\u5ef6\u8fdf\u3002<\/td><\/tr><tr><td><a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=Pipeline+Parallelism&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">Pipeline Parallelism<\/a>&nbsp;(PP)<\/td><td>\u662f<\/td><td>\u8bad\u7ec3\u52a0\u901f<\/td><td>\u4f7f\u7528 DualPipe \u6280\u672f\u91cd\u53e0 attention \u4e0e MoE \u8ba1\u7b97\uff0c\u51cf\u5c11 pipeline bubbles\u3002<\/td><\/tr><tr><td>Expert Parallelism (EP)<\/td><td>\u662f<\/td><td>MoE \u5206\u5e03\u5f0f\u5904\u7406<\/td><td>\u4f7f\u7528 DeepEP \u5b9e\u73b0\u9ad8\u6548\u7684 All-to-All \u901a\u4fe1\uff0c\u5f00\u6e90\u5b9e\u73b0\u3002<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\"><br>4.2 \u6a21\u578b\u67b6\u6784\u4e0e\u786c\u4ef6\u534f\u540c\u8bbe\u8ba1\uff1a Node-Limited Routing<\/h3>\n\n\n\n<p><br>\u5728 DeepSeek-V3 \u6240\u4f7f\u7528\u7684 NVIDIA H800 GPU \u67b6\u6784 \u4e2d\uff0c\u5b58\u5728\u660e\u663e\u7684 Scale-Up\uff08\u8282\u70b9\u5185\uff09\u4e0e Scale-Out\uff08\u8282\u70b9\u95f4\uff09\u901a\u4fe1\u5e26\u5bbd\u4e0d\u5339\u914d\u95ee\u9898 \uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td>\u901a\u4fe1\u7c7b\u578b<\/td><td>\u5e26\u5bbd\u7406\u8bba\u503c<\/td><td>\u5b9e\u9645\u6709\u6548\u5e26\u5bbd<\/td><td>\u5e26\u5bbd\u6bd4\u4f8b\uff08\u8fd1\u4f3c\uff09<\/td><\/tr><tr><td>Scale-Up (NVLink)<\/td><td>200 GB\/s<\/td><td>~160 GB\/s<\/td><td>4:1<\/td><\/tr><tr><td>Scale-Out (IB NIC)<\/td><td>50 GB\/s<\/td><td>~40 GB\/s \uff08\u7531\u4e8e\u5c0f\u6570\u636e\u5305\u4f20\u8f93\u548c\u5ef6\u8fdf\u5f71\u54cd\uff09<\/td><td>&#8211;<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p><br>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0cDeepSeek \u56e2\u961f\u63d0\u51fa\u4e86 Node-Limited Routing\uff08\u8282\u70b9\u9650\u5236\u8def\u7531\uff09 \u548c TopK \u4e13\u5bb6\u9009\u62e9\u7b56\u7565 \uff0c\u901a\u8fc7\u5c06\u4e13\u5bb6\u90e8\u7f72\u5728\u6709\u9650\u6570\u91cf\u7684\u8282\u70b9\u4e0a\uff0c\u4ece\u800c\u51cf\u5c11\u8de8\u8282\u70b9\u901a\u4fe1\u91cf\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e13\u5bb6\u5206\u7ec4\u4e0e\u90e8\u7f72\n<ul class=\"wp-block-list\">\n<li>\u5c06\u603b\u5171 256 \u4e2a\u4e13\u5bb6 \u5206\u6210 8 \u7ec4 \u3002<\/li>\n\n\n\n<li>\u6bcf\u7ec4\u5305\u542b 32 \u4e2a\u4e13\u5bb6 \uff0c\u5e76\u90e8\u7f72\u5728 \u540c\u4e00\u4e2a\u8282\u70b9 \u4e0a\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Token \u8def\u7531\u7b56\u7565\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u4e2a token \u6700\u591a\u88ab\u8def\u7531\u5230 4 \u4e2a\u8282\u70b9 \u3002<\/li>\n\n\n\n<li>\u8fd9\u6837\u53ef\u4ee5\u786e\u4fdd\uff1a\n<ul class=\"wp-block-list\">\n<li>\u8de8\u8282\u70b9\u901a\u4fe1\u91cf\u51cf\u5c11\u3002<\/li>\n\n\n\n<li>\u540c\u4e00\u8282\u70b9\u5185\u7684\u4e13\u5bb6\u53ef\u901a\u8fc7\u9ad8\u901f NVLink \u901a\u4fe1\uff0c\u964d\u4f4e\u5ef6\u8fdf\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u5047\u8bbe\u4f7f\u7528 8 \u4e2a\u8282\u70b9 \uff08\u5171 64 \u4e2a GPU\uff09\uff0c\u6bcf\u4e2a GPU \u6709 4 \u4e2a\u4e13\u5bb6\uff0c\u603b\u5171\u6709 256 \u4e2a\u4e13\u5bb6\uff0c\u6bcf\u4e2a token \u88ab\u8def\u7531\u5230 1 \u4e2a\u5171\u4eab\u4e13\u5bb6 + 8 \u4e2a\u76ee\u6807\u4e13\u5bb6\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f20\u7edf\u5b9e\u73b0\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5982\u679c\u8fd9 8 \u4e2a\u76ee\u6807\u4e13\u5bb6\u5206\u5e03\u5728\u6240\u6709 8 \u4e2a\u8282\u70b9\u4e0a\uff0c\u5219\u9700\u8fdb\u884c 8 \u6b21 IB \u901a\u4fe1\u3002<\/li>\n\n\n\n<li>\u603b\u901a\u4fe1\u65f6\u95f4\u4e3a\u00a0<code>8 * t<\/code>\uff08t \u4e3a\u5355\u6b21 IB \u4f20\u8f93\u65f6\u95f4\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Node-Limited Routing \u65b9\u6cd5\n<ul class=\"wp-block-list\">\n<li>\u76ee\u6807\u4e13\u5bb6\u88ab\u9650\u5236\u5728\u6700\u591a 4 \u4e2a\u8282\u70b9\u3002<\/li>\n\n\n\n<li>\u6bcf\u4e2a\u8282\u70b9\u53ea\u9700\u4e00\u6b21 IB \u4f20\u8f93\uff0c\u7136\u540e\u901a\u8fc7 NVLink \u5728\u8282\u70b9\u5185\u5e7f\u64ad\u3002<\/li>\n\n\n\n<li>\u603b\u901a\u4fe1\u65f6\u95f4\u964d\u81f3\u00a0<code>4 * t<\/code>\uff0c\u8282\u7701\u4e86 50% \u7684\u901a\u4fe1\u5f00\u9500\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">4.3&nbsp;<strong>Scale-up \u4e0e Scale-out \u7f51\u7edc\u878d\u5408<\/strong><\/h3>\n\n\n\n<p><br>\u5f53\u524d\u786c\u4ef6\u4e2d\uff0cScale-up\uff08\u8282\u70b9\u5185\uff09\u548c Scale-out\uff08\u8282\u70b9\u95f4\uff09\u7f51\u7edc\u662f\u5206\u79bb\u7684\uff0c\u5bfc\u81f4\u901a\u4fe1\u8def\u5f84\u590d\u6742\u3001\u5ef6\u8fdf\u9ad8\u3002<br>\u5c3d\u7ba1 MPFT \u548c EP \u7b49\u6280\u672f\u53d6\u5f97\u4e86\u826f\u597d\u6548\u679c\uff0c\u4f46\u4ecd\u9762\u4e34\u4e00\u4e9b\u6311\u6218\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5f53\u524d IB ConnectX-7 \u4e0d\u652f\u6301\u591a\u5e73\u9762\u65e0\u7f1d\u901a\u4fe1\uff0c\u9700\u901a\u8fc7\u8282\u70b9\u8f6c\u53d1\uff0c\u589e\u52a0\u5ef6\u8fdf\u3002<\/li>\n\n\n\n<li>\u8f6f\u4ef6\u590d\u6742\u5ea6\u9ad8 \uff1a\n<ul class=\"wp-block-list\">\n<li>Node-Limited Routing \u7b56\u7565\u589e\u52a0\u4e86\u901a\u4fe1\u6d41\u6c34\u7ebf kernel \u7684\u5b9e\u73b0\u590d\u6742\u5ea6\u3002<\/li>\n\n\n\n<li>GPU Streaming Multiprocessors(SM) \u9700\u8981\u540c\u65f6\u5904\u7406\u7f51\u7edc\u6d88\u606f\u586b\u5145\u548c NVLink \u8f6c\u53d1\uff0c\u6d88\u8017\u8ba1\u7b97\u8d44\u6e90\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5185\u5b58\u4e00\u81f4\u6027\u95ee\u9898 \uff1a\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u7528 load\/store \u8bed\u4e49\u8fdb\u884c\u8de8\u8282\u70b9\u901a\u4fe1\u65f6\uff0c\u9700\u663e\u5f0f\u63d2\u5165\u5185\u5b58\u5c4f\u969c\uff08fence\uff09\uff0c\u5f71\u54cd\u541e\u5410\u3002<\/li>\n\n\n\n<li>RDMA \u4e2d\u4e5f\u5b58\u5728 out-of-order synchronization \u95ee\u9898\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u672a\u6765\u65b9\u5411\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7edf\u4e00\u7f51\u7edc\u9002\u914d\u5668\uff08Unified Network Adapter\uff09\u3001\n<ul class=\"wp-block-list\">\n<li>\u8bbe\u8ba1\u652f\u6301\u7edf\u4e00 Scale-Up \u4e0e Scale-Out \u7f51\u7edc\u7684 NIC \u6216 I\/O Die\u3002<\/li>\n\n\n\n<li>\u652f\u6301\u57fa\u672c\u4ea4\u6362\u529f\u80fd\uff0c\u5982\u5c06 Scale-Out \u6570\u636e\u5305\u8f6c\u53d1\u5230\u7279\u5b9a GPU\u3002<\/li>\n\n\n\n<li>\u4f7f\u7528\u5355\u4e00 LID\/IP \u5730\u5740 + \u7b56\u7565\u8def\u7531\uff0c\u7b80\u5316\u5730\u5740\u7ba1\u7406\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u4e13\u7528\u901a\u4fe1\u534f\u5904\u7406\u5668\uff08Dedicated Communication Co-Processor\uff09\n<ul class=\"wp-block-list\">\n<li>\u5f15\u5165\u4e13\u7528\u534f\u5904\u7406\u5668\u6216\u53ef\u7f16\u7a0b\u7ec4\u4ef6\uff08\u5982 I\/O Die\uff09\uff0c\u5904\u7406\u7f51\u7edc\u6d41\u91cf\u3002<\/li>\n\n\n\n<li>\u5378\u8f7d GPU SM \u4e0a\u7684\u901a\u4fe1\u4efb\u52a1\uff0c\u9632\u6b62\u6027\u80fd\u9000\u5316\u3002<\/li>\n\n\n\n<li>\u652f\u6301\u786c\u4ef6\u52a0\u901f\u7684\u5185\u5b58\u62f7\u8d1d\u3001\u5e7f\u64ad\u3001\u8f6c\u53d1\u64cd\u4f5c\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u7075\u6d3b\u7684\u6570\u636e\u8f6c\u53d1\u3001\u5e7f\u64ad\u4e0e\u5f52\u7ea6\u673a\u5236\n<ul class=\"wp-block-list\">\n<li>\u786c\u4ef6\u5e94\u652f\u6301\u7075\u6d3b\u7684\u8f6c\u53d1\u673a\u5236\uff0c\u7528\u4e8e MoE \u5206\u53d1\u548c\u805a\u5408\u3002<\/li>\n\n\n\n<li>\u652f\u6301\u9ad8\u6548\u7684\u5f52\u7ea6\u64cd\u4f5c\uff08reduce\uff09\uff0c\u63d0\u5347\u901a\u4fe1\u6548\u7387\u3002<\/li>\n\n\n\n<li>\u5b9e\u73b0\u7c7b\u4f3c\u4e8e\u5f53\u524d GPU SM-based \u5b9e\u73b0\u7684\u673a\u5236\uff0c\u4f46\u7531\u786c\u4ef6\u76f4\u63a5\u5b8c\u6210\u3002\u3001<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u4f4e\u5ef6\u8fdf\u4e92\u8fde\u4e0e\u540c\u6b65\u539f\u8bed\n<ul class=\"wp-block-list\">\n<li>\u652f\u6301 acquire\/release \u540c\u6b65\u8bed\u4e49\uff0c\u51cf\u5c11\u8f6f\u4ef6\u540c\u6b65\u5f00\u9500\u3002<\/li>\n\n\n\n<li>\u63d0\u4f9b\u786c\u4ef6\u7ea7\u539f\u5b50\u64cd\u4f5c\u3001\u5c4f\u969c\u6307\u4ee4\uff0c\u63d0\u5347\u901a\u4fe1\u6548\u7387\u3002<\/li>\n\n\n\n<li>\u652f\u6301\u4e71\u5e8f\u5305\u5230\u8fbe\u7684\u786c\u4ef6\u6392\u5e8f\uff08RAR \u673a\u5236\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u5b9e\u73b0 Scale-Up \u4e0e Scale-Out \u7f51\u7edc\u878d\u5408\u4e4b\u540e\uff0cAI \u7cfb\u7edf\u5c06\u83b7\u5f97\u4ee5\u4e0b\u663e\u8457\u4f18\u52bf\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td>\u4f18\u52bf<\/td><td>\u63cf\u8ff0<\/td><\/tr><tr><td>\u66f4\u4f4e\u7684\u901a\u4fe1\u5ef6\u8fdf<\/td><td>\u7edf\u4e00\u7f51\u7edc\u63a5\u53e3\u51cf\u5c11\u4e86\u4e2d\u95f4\u8f6c\u53d1\u6b65\u9aa4\uff0c\u63d0\u5347\u901a\u4fe1\u6548\u7387\u3002<\/td><\/tr><tr><td>\u66f4\u9ad8\u7684\u6709\u6548\u5e26\u5bbd\u5229\u7528\u7387<\/td><td>\u53ef\u52a8\u6001\u5206\u914d\u5e26\u5bbd\u8d44\u6e90\uff0c\u9002\u5e94\u4e0d\u540c\u5de5\u4f5c\u8d1f\u8f7d\u9700\u6c42\u3002<\/td><\/tr><tr><td>\u66f4\u7b80\u5355\u7684\u8f6f\u4ef6\u6808<\/td><td>\u65e0\u9700\u533a\u5206 intra-node \u4e0e inter-node \u901a\u4fe1\uff0c\u7b80\u5316\u8c03\u5ea6\u903b\u8f91\u3002<\/td><\/tr><tr><td>\u66f4\u5f3a\u7684\u53ef\u6269\u5c55\u6027<\/td><td>\u66f4\u5bb9\u6613\u652f\u6301\u8d85\u5927\u89c4\u6a21\u96c6\u7fa4\uff08\u5982 10k+ GPU\uff09\u3002<\/td><\/tr><tr><td>\u66f4\u597d\u7684\u8d1f\u8f7d\u5747\u8861\u80fd\u529b<\/td><td>\u53ef\u6839\u636e\u7f51\u7edc\u72b6\u6001\u52a8\u6001\u9009\u62e9\u6700\u4f18\u8def\u5f84\uff0c\u907f\u514d\u70ed\u70b9\u74f6\u9888\u3002<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\"><br>4.4 \u5e26\u5bbd\u4e89\u7528\u4e0e\u5ef6\u8fdf<\/h3>\n\n\n\n<p><br>\u5728 H800 GPU \u67b6\u6784\u4e2d\uff0cGPU \u548c CPU \u4e4b\u95f4\u7684 PCIe \u63a5\u53e3\u662f\u8fde\u63a5\u4e3b\u5b58\u548c\u663e\u5b58\u7684\u5173\u952e\u8def\u5f84\u3002\u7136\u800c\uff0c\u5728\u5927\u89c4\u6a21\u8bad\u7ec3\u6216\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u591a\u4e2a\u4efb\u52a1\u4f1a\u540c\u65f6\u4e89\u593a\u6709\u9650\u7684 PCIe \u5e26\u5bbd\uff0c\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d\u3002<br>\u6bd4\u5982\u5728\u4ee5\u4e0b\u4e24\u79cd\u573a\u666f\u4e0b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>KV Cache \u4f20\u8f93 \uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728\u63a8\u7406\u9636\u6bb5\uff0c\u9700\u8981\u9891\u7e41\u5c06 Key\/Value cache\u4ece CPU \u5185\u5b58\u62f7\u8d1d\u5230 GPU \u663e\u5b58\u3002<\/li>\n\n\n\n<li>\u8fd9\u7c7b\u64cd\u4f5c\u53ef\u80fd\u5360\u7528\u6570\u5341 GB\/s \u7684\u5e26\u5bbd\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u4e13\u5bb6\u5e76\u884c\uff08EP\uff09\u901a\u4fe1 \uff1a\n<ul class=\"wp-block-list\">\n<li>MoE \u6a21\u578b\u4e2d\u7684 All-to-All \u901a\u4fe1\u4e5f\u4f9d\u8d56 IB NIC \u548c GPU \u4e4b\u95f4\u7684\u6570\u636e\u4ea4\u6362\u3002<\/li>\n\n\n\n<li>\u5982\u679c EP \u901a\u4fe1\u548c KV Cache \u4f20\u8f93\u540c\u65f6\u8fdb\u884c\uff0cPCIe \u5e26\u5bbd\u5c31\u4f1a\u6210\u4e3a\u74f6\u9888\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u56e0\u6b64\uff0c\u4f1a\u5bfc\u81f4\u5b58\u5728\u4ee5\u4e0b\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6027\u80fd\u4e0b\u964d \uff1a\u5e26\u5bbd\u4e89\u7528\u4f1a\u5bfc\u81f4\u6574\u4f53\u541e\u5410\u4e0b\u964d\u3002<\/li>\n\n\n\n<li>\u5ef6\u8fdf\u589e\u52a0 \uff1a\u67d0\u4e9b\u4efb\u52a1\u7684\u5b8c\u6210\u65f6\u95f4\u88ab\u663e\u8457\u5ef6\u957f\u3002<\/li>\n\n\n\n<li>\u8d44\u6e90\u5229\u7528\u7387\u4f4e \uff1aGPU \u53ef\u80fd\u56e0\u7b49\u5f85\u6570\u636e\u800c\u5904\u4e8e\u7a7a\u95f2\u72b6\u6001<\/li>\n<\/ul>\n\n\n\n<p>\u4f18\u5316\u5efa\u8bae\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u52a8\u6001\u4f18\u5148\u7ea7\u8c03\u5ea6\u673a\u5236\uff0c\u6839\u636e\u4e0d\u540c\u7c7b\u578b\u7684\u6d41\u91cf\u9700\u6c42\uff0c\u52a8\u6001\u8c03\u6574\u5e26\u5bbd\u5206\u914d\u7b56\u7565\u3002\u4fdd\u8bc1\u5173\u952e\u4efb\u52a1\uff08\u5982\u63a8\u7406\u4efb\u52a1\uff09\u7684\u4f4e\u5ef6\u8fdf\u6267\u884c\u3002\u9632\u6b62\u975e\u5173\u952e\u4efb\u52a1\u5360\u7528\u8fc7\u591a\u5e26\u5bbd\uff0c\u63d0\u5347\u6574\u4f53\u6548\u7387\u3002\n<ul class=\"wp-block-list\">\n<li>\u5bf9\u4e0d\u540c\u7c7b\u578b\u6d41\u91cf\u8bbe\u5b9a\u4e0d\u540c\u4f18\u5148\u7ea7\uff0c\u4f8b\u5982\uff1a\n<ul class=\"wp-block-list\">\n<li>\u9ad8\u4f18\u5148\u7ea7\uff1aEP\u3001TP\u3001KV Cache \u4f20\u8f93<\/li>\n\n\n\n<li>\u4e2d\u4f4e\u4f18\u5148\u7ea7\uff1a\u5176\u4ed6\u540e\u53f0\u4efb\u52a1<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5728 PCIe \u4e0a\u66b4\u9732 Traffic Class (TC) \u7ed9\u7528\u6237\u7a7a\u95f4\u7a0b\u5e8f\uff0c\u5b9e\u73b0\u7075\u6d3b\u63a7\u5236\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5c06\u7f51\u5361\u76f4\u63a5\u96c6\u6210\u5230 I\/O Die \u4e2d\uff08Chiplet\u67b6\u6784\uff09\uff0c\u51cf\u5c11 PCIe \u6210\u4e3a\u74f6\u9888\u7684\u98ce\u9669\uff0c\u63d0\u5347\u6574\u4f53\u7cfb\u7edf\u7684\u53ef\u6269\u5c55\u6027\u548c\u7a33\u5b9a\u6027\uff0c\u5e76\u964d\u4f4e\u901a\u4fe1\u5ef6\u8fdf\n<ul class=\"wp-block-list\">\n<li>\u5c06 InfiniBand \u6216 RoCE \u7f51\u5361\u76f4\u63a5\u96c6\u6210\u5230\u5c01\u88c5\u7ea7\u522b\u7684 I\/O Die \u4e2d\u3002<\/li>\n\n\n\n<li>\u901a\u8fc7\u5c01\u88c5\u5185\u4e92\u8fde\u6280\u672f\uff08\u5982 NVLink\u3001Infinity Fabric\uff09\u8fde\u63a5 GPU Compute Die\u3002\u3001<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u4f7f\u7528 NVLink \u66ff\u4ee3 PCIe\uff0c\u7edf\u4e00 Scale-Up \u57df\u5185\u7684 CPU-GPU \u901a\u4fe1 \uff0c\u901a\u8fc7\u4f18\u5316\u8282\u70b9\u5185\u90e8 CPU \u4e0e GPU \u7684\u901a\u4fe1\u6548\u7387\uff0c\u907f\u514d PCIe \u6210\u4e3a\u74f6\u9888\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u7528 NVLink\u3001Infinity Fabric \u7b49\u9ad8\u901f\u4e92\u8fde\u6280\u672f\u66ff\u4ee3\u4f20\u7edf PCIe\u3002<\/li>\n\n\n\n<li>\u5c06 CPU \u548c GPU \u653e\u5165\u540c\u4e00\u4e2a Scale-Up \u57df\u4e2d\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">5 \u5927\u89c4\u6a21\u7f51\u7edc\u9a71\u52a8\u8bbe\u8ba1<\/h2>\n\n\n\n<p>\u968f\u7740 LLM\uff08\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff09\u53c2\u6570\u89c4\u6a21\u7684\u6307\u6570\u7ea7\u589e\u957f\uff0c\u4f20\u7edf\u7684\u4e09\u5c42 Fat-Tree\uff08FT3\uff09\u62d3\u6251\u5728\u6210\u672c\u3001\u53ef\u6269\u5c55\u6027\u548c\u9c81\u68d2\u6027\u65b9\u9762\u9010\u6e10\u66b4\u9732\u51fa\u5c40\u9650\u6027\u3002\u7279\u522b\u662f\u5728 MoE \u67b6\u6784\u4e0b\uff0cAll-to-All \u901a\u4fe1\u5bc6\u96c6\u578b\u64cd\u4f5c\u5bf9\u5e26\u5bbd\u63d0\u51fa\u4e86\u66f4\u9ad8\u8981\u6c42\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><br><strong>5.1&nbsp;<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=257770090&amp;content_type=Article&amp;match_order=1&amp;q=Multi-Plane+Fat-Tree+%E7%BD%91%E7%BB%9C%E6%8B%93%E6%89%91&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">Multi-Plane Fat-Tree \u7f51\u7edc\u62d3\u6251<\/a>\uff08MPFT\uff09<\/strong><\/h3>\n\n\n\n<p><br>DeepSeek-V3 \u90e8\u7f72\u4e86\u4e00\u4e2a\u57fa\u4e8e InfiniBand \u7684<strong>\u591a\u5e73\u9762\u4e24\u5c42 Fat-Tree \u7f51\u7edc\uff08MPFT\uff09<\/strong>\uff0c\u7528\u4e8e\u66ff\u4ee3\u4f20\u7edf\u7684\u4e09\u5c42 Fat-Tree \u67b6\u6784\uff0c\u5982\u56feFigure 3\u6240\u793a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"513\" height=\"390\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-a9c675f299cd8de76df7162e2b839dce_1440w.jpg\" alt=\"\" class=\"wp-image-541\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-a9c675f299cd8de76df7162e2b839dce_1440w.jpg 513w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-a9c675f299cd8de76df7162e2b839dce_1440w-300x228.jpg 300w\" sizes=\"auto, (max-width: 513px) 100vw, 513px\" \/><\/figure>\n\n\n\n<p><br>\u67b6\u6784\u7279\u70b9\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u4e2a\u8282\u70b9\u914d\u5907 8 \u5757 GPU \u548c 8 \u4e2a 400Gbps InfiniBand NIC \u3002<\/li>\n\n\n\n<li>\u6bcf\u5757 GPU \u5bf9\u5e94\u4e00\u4e2a\u72ec\u7acb\u7684 IB NIC\uff0c\u5c5e\u4e8e\u4e0d\u540c\u7684\u201c\u7f51\u7edc\u5e73\u9762\u201d\uff08plane\uff09\u3002<\/li>\n\n\n\n<li>\u6240\u6709\u8282\u70b9\u8fde\u63a5\u5230 64-port 400G IB \u4ea4\u6362\u673a \uff0c\u6784\u6210\u4e24\u5c42 Fat-Tree\u3002<\/li>\n\n\n\n<li>\u6b64\u5916\uff0c\u6bcf\u4e2a\u8282\u70b9\u8fd8\u914d\u7f6e\u4e86\u4e00\u4e2a RoCE NIC\uff0c\u7528\u4e8e\u8fde\u63a5\u5206\u5e03\u5f0f\u6587\u4ef6\u7cfb\u7edf\uff08\u5982 3FS\uff09\uff0c\u5b9e\u73b0\u5b58\u50a8\u7f51\u7edc\u9694\u79bb\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4f18\u70b9\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6210\u672c\u66f4\u4f4e\uff1a\u76f8\u6bd4\u4e09\u5c42 Fat-Tree\uff0cMPFT \u53ef\u8282\u7701\u9ad8\u8fbe 40% \u7684\u7f51\u7edc\u6210\u672c\u3002<\/li>\n\n\n\n<li>\u66f4\u9ad8\u7684\u53ef\u6269\u5c55\u6027\uff1a\u7406\u8bba\u4e0a\u652f\u6301\u6700\u591a 16,384 \u4e2a GPU\u3002<\/li>\n\n\n\n<li>\u6d41\u91cf\u9694\u79bb\uff1a\u6bcf\u4e2a\u5e73\u9762\u72ec\u7acb\u8fd0\u884c\uff0c\u907f\u514d\u8de8\u5e73\u9762\u62e5\u585e\u3002<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"517\" height=\"218\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-5db0ba7383b76aa87cfda3bbcbe1d110_1440w.jpg\" alt=\"\" class=\"wp-image-539\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-5db0ba7383b76aa87cfda3bbcbe1d110_1440w.jpg 517w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-5db0ba7383b76aa87cfda3bbcbe1d110_1440w-300x126.jpg 300w\" sizes=\"auto, (max-width: 517px) 100vw, 517px\" \/><\/figure>\n\n\n\n<p>\u5c3d\u7ba1 MPFT \u5728\u7406\u8bba\u4e0a\u5177\u6709\u5f3a\u5927\u7684\u4f18\u52bf\uff0c\u7406\u60f3\u60c5\u51b5\u5982Figure 4\u6240\u793a\uff0c<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1036\" height=\"351\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-e4effaaa70a2a04871419ae89151bcc0_1440w.jpg\" alt=\"\" class=\"wp-image-544\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-e4effaaa70a2a04871419ae89151bcc0_1440w.jpg 1036w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-e4effaaa70a2a04871419ae89151bcc0_1440w-300x102.jpg 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-e4effaaa70a2a04871419ae89151bcc0_1440w-1024x347.jpg 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-e4effaaa70a2a04871419ae89151bcc0_1440w-768x260.jpg 768w\" sizes=\"auto, (max-width: 1036px) 100vw, 1036px\" \/><\/figure>\n\n\n\n<p><br>\u4f46\u5b9e\u9645\u4e0a\u53d7\u9650\u4e8eIB ConnectX-7\u786c\u4ef6\uff0c\u4e3b\u8981\u8868\u73b0\u5728\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e0d\u652f\u6301\u771f\u6b63\u7684\u591a\u5e73\u9762\u901a\u4fe1\u3002<\/li>\n\n\n\n<li>QP \u6570\u636e\u5305\u65e0\u6cd5\u81ea\u52a8\u5728\u591a\u4e2a\u5e73\u9762\u4e0a\u8d1f\u8f7d\u5747\u8861\u3002<\/li>\n\n\n\n<li>\u82e5\u9700\u8981\u8de8\u5e73\u9762\u901a\u4fe1\uff0c\u5fc5\u987b\u901a\u8fc7\u8282\u70b9\u5185\u8f6c\u53d1\uff08intra-node forwarding\uff09\uff0c\u5f15\u5165\u989d\u5916\u5ef6\u8fdf\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4f46\u662fNVIDIA \u65b0\u4e00\u4ee3\u7f51\u5361 ConnectX-8 \u5df2\u539f\u751f\u652f\u6301 4 \u4e2a\u7f51\u7edc\u5e73\u9762\uff084-plane\uff09\uff0c\u53ef\u4ee5\u5728\u4e00\u4e2a QP \u4e0a\u5b9e\u73b0\u771f\u6b63\u7684 multi-path packet spraying\u3002\u5e76\u652f\u6301\u786c\u4ef6\u7ea7\u4e71\u5e8f\u5305\u5904\u7406\uff0c\u786e\u4fdd\u6570\u636e\u4e00\u81f4\u6027\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><br>5.2 \u4f4e\u5ef6\u8fdf\u7f51\u7edc<\/h3>\n\n\n\n<p><br>DeepSeek-V3 \u7684\u8bad\u7ec3\u4f9d\u8d56\u4e8e\u8d85\u5927\u89c4\u6a21\u96c6\u7fa4\uff08\u5982 2048 \u5757 H800 GPU\uff09\uff0c\u800c MoE\uff08Mixture of Experts\uff09\u67b6\u6784\u4e0b\u7684\u4e13\u5bb6\u5e76\u884c\uff08EP\uff09\u64cd\u4f5c\u5bf9\u5e26\u5bbd\u548c\u5ef6\u8fdf\u90fd\u975e\u5e38\u654f\u611f\u3002\u56e0\u6b64\uff0c\u7f51\u7edc\u5ef6\u8fdf\u7684\u4f18\u5316\u5bf9\u4e8e\u63d0\u5347\u6574\u4f53\u7cfb\u7edf\u6027\u80fd\u81f3\u5173\u91cd\u8981\u3002<br>1.&nbsp;<strong>InfiniBand vs. RoCE\uff1a\u6027\u80fd\u5bf9\u6bd4<\/strong><br>\u5728 DeepSeek-V3 \u7684\u90e8\u7f72\u4e2d\uff0c\u4f7f\u7528\u4e86 InfiniBand\uff08IB\uff09\u4f5c\u4e3a\u4e3b\u8981\u901a\u4fe1\u534f\u8bae\u3002\u4e3a\u4e86\u8bc4\u4f30\u5176\u6027\u80fd\u4f18\u52bf\uff0c\u8bba\u6587\u5bf9\u6bd4\u4e86 IB \u4e0e RoCE \u5728\u4e0d\u540c\u573a\u666f\u4e0b\u7684\u7aef\u5230\u7aef\u5ef6\u8fdf\u8868\u73b0\u3002<\/p>\n\n\n\n<p>64B \u6570\u636e\u4f20\u8f93\u5ef6\u8fdf\u5bf9\u6bd4\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"354\" height=\"113\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-c06ab1e5f43d68bb843215d2109d0983_1440w.jpg\" alt=\"\" class=\"wp-image-538\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-c06ab1e5f43d68bb843215d2109d0983_1440w.jpg 354w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/05\/v2-c06ab1e5f43d68bb843215d2109d0983_1440w-300x96.jpg 300w\" sizes=\"auto, (max-width: 354px) 100vw, 354px\" \/><\/figure>\n\n\n\n<p>\u53ef\u4ee5\u53d1\u73b0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>InfiniBand \u5ef6\u8fdf\u6700\u4f4e<\/strong>\uff0c\u662f\u5f53\u524d\u6700\u9002\u5408 MoE \u67b6\u6784\u4e0b All-to-All \u901a\u4fe1\u7684\u7f51\u7edc\u534f\u8bae\u3002<\/li>\n\n\n\n<li><strong>RoCE \u867d\u5177\u6210\u672c\u4f18\u52bf\uff0c\u4f46\u5ef6\u8fdf\u8f83\u9ad8<\/strong>\uff0c\u5c24\u5176\u5728\u8de8 Leaf \u4ea4\u6362\u673a\u65f6\u8868\u73b0\u4e0d\u4f73\u3002<\/li>\n\n\n\n<li><strong>NVLink<\/strong>\u00a0\u63d0\u4f9b\u4e86\u6700\u4f4e\u7684\u8282\u70b9\u5185\u901a\u4fe1\u5ef6\u8fdf\uff0c\u9002\u7528\u4e8e Tensor Parallelism \u548c KV Cache \u4f20\u8f93\u7b49 intra-node \u64cd\u4f5c\u3002<\/li>\n<\/ul>\n\n\n\n<p><br>\u5c3d\u7ba1 RoCE \u5728\u6210\u672c\u4e0a\u5177\u6709\u4e00\u5b9a\u4f18\u52bf\uff0c\u4f46\u5176\u5ef6\u8fdf\u548c\u6269\u5c55\u6027\u4ecd\u4e0d\u80fd\u5b8c\u5168\u6ee1\u8db3\u5927\u89c4\u6a21 AI \u7cfb\u7edf\u7684\u9700\u6c42\u3002\u4e3a\u6b64\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u4ee5\u4e0b\u51e0\u70b9\u6539\u8fdb\u5efa\u8bae\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e13\u7528\u4f4e\u5ef6\u8fdf RoCE \u4ea4\u6362\u673a\n<ul class=\"wp-block-list\">\n<li>\u5f53\u524d\u4ee5\u592a\u7f51\u4ea4\u6362\u673a\u5305\u542b\u5927\u91cf\u4e0d\u5fc5\u8981\u7684\u529f\u80fd\uff08\u5982 QoS\u3001VLAN \u7b49\uff09\uff0c\u589e\u52a0\u4e86\u5904\u7406\u5ef6\u8fdf\u3002<\/li>\n\n\n\n<li>\u5efa\u8bae\u5f00\u53d1\u4e13\u95e8\u9762\u5411 RDMA \u5de5\u4f5c\u8d1f\u8f7d\u7684 RoCE \u4ea4\u6362\u673a\uff0c\u53bb\u9664\u5197\u4f59\u529f\u80fd\uff0c\u63d0\u5347\u6027\u80fd\u3002<\/li>\n\n\n\n<li>\u793a\u4f8b\uff1aSlingshot \u67b6\u6784\u5df2\u8bc1\u660e\u57fa\u4e8e\u4ee5\u592a\u7f51\u7684\u8bbe\u8ba1\u53ef\u4ee5\u5b9e\u73b0\u63a5\u8fd1 IB \u7684\u5ef6\u8fdf\u8868\u73b0\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>AI \u8f6c\u53d1\u5934\uff08AIFH\uff09\u4e0e\u4f4e\u5ef6\u8fdf\u4ee5\u592a\u7f51\u4ea4\u6362\u673a\n<ul class=\"wp-block-list\">\n<li>Broadcom \u6700\u65b0\u63a8\u51fa\u7684 AI Forwarding Header\uff08AIFH\uff09\u534f\u8bae\u53ef\u652f\u6301\u66f4\u9ad8\u6548\u7684\u8def\u7531\u673a\u5236\u3002<\/li>\n\n\n\n<li>\u7ed3\u5408\u5373\u5c06\u53d1\u5e03\u7684\u4f4e\u5ef6\u8fdf\u4ee5\u592a\u7f51\u4ea4\u6362\u673a\uff0c\u6709\u671b\u5b9e\u73b0\u5ab2\u7f8e IB \u7684\u6027\u80fd\u3002<\/li>\n\n\n\n<li>\u8fd9\u4e9b\u6280\u672f\u4e3a\u672a\u6765\u6784\u5efa\u9ad8\u6027\u4ef7\u6bd4\u7684 AI \u96c6\u7fa4\u63d0\u4f9b\u4e86\u65b0\u8def\u5f84\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u62e5\u585e\u63a7\u5236\u7b97\u6cd5\u4f18\u5316\n<ul class=\"wp-block-list\">\n<li>\u5f53\u524d RoCE \u4f7f\u7528\u7684\u62e5\u585e\u63a7\u5236\u673a\u5236\u4e0d\u591f\u667a\u80fd\uff0c\u5bb9\u6613\u5bfc\u81f4 Head-of-Line Blocking\u3002<\/li>\n\n\n\n<li>\u5efa\u8bae\u91c7\u7528 endpoint-driven \u62e5\u585e\u63a7\u5236\u7b97\u6cd5\uff0c\u52a8\u6001\u8c03\u8282\u6ce8\u5165\u901f\u7387\uff0c\u907f\u514d\u4e25\u91cd\u62e5\u585e\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u5728 MoE \u67b6\u6784\u6216 All-to-All \u901a\u4fe1\u4e2d\uff0c\u5e38\u5e38\u6d89\u53ca\u5927\u91cf\u5c0f\u6570\u636e\u5305\u4f20\u8f93\uff0c \u5982\u679c\u7531 CPU \u5904\u7406\u8fd9\u4e9b\u901a\u4fe1\u4efb\u52a1\uff0cCPU \u5f88\u5bb9\u6613\u6210\u4e3a\u74f6\u9888\u3002\u7279\u522b\u662f\u5728\u9ad8\u5e76\u53d1\u573a\u666f\u4e0b\uff0cCPU \u7684\u4e32\u884c\u5904\u7406\u80fd\u529b\u65e0\u6cd5\u6ee1\u8db3\u9700\u6c42\u3002<br>InfiniBand GPUDirect Async\uff08IBGDA\uff09 \u662f NVIDIA \u63d0\u51fa\u7684\u4e00\u9879\u7528\u4e8e\u52a0\u901f GPU \u4e0e\u7f51\u7edc\u8bbe\u5907\u4e4b\u95f4\u901a\u4fe1\u7684\u6280\u672f\u3002\u5b83\u901a\u8fc7\u7ed5\u8fc7 CPU \u7684\u76f4\u63a5\u53c2\u4e0e\uff0c\u4f7f GPU \u80fd\u591f\u5f02\u6b65\u5730\u542f\u52a8\u548c\u7ba1\u7406 RDMA\u64cd\u4f5c\uff0c\u4ece\u800c\u663e\u8457\u964d\u4f4e\u901a\u4fe1\u5ef6\u8fdf\u5e76\u63d0\u5347\u6574\u4f53\u7cfb\u7edf\u541e\u5410\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>GPU \u76f4\u63a5\u586b\u5145 Work Request\n<ul class=\"wp-block-list\">\n<li>\u5141\u8bb8 GPU \u76f4\u63a5\u5411 NIC \u7684\u961f\u5217\u5bf9\uff08Queue Pair, QP\uff09\u5199\u5165 Work Request\uff08WR\uff09\u3002<\/li>\n\n\n\n<li>\u7ed5\u8fc7\u4e86\u4f20\u7edf\u6d41\u7a0b\u4e2d\u7684 CPU \u4ee3\u7406\u7ebf\u7a0b\uff0c\u51cf\u5c11\u4e0a\u4e0b\u6587\u5207\u6362\u548c\u540c\u6b65\u5f00\u9500\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5f02\u6b65 Doorbell \u5199\u5165\n<ul class=\"wp-block-list\">\n<li>GPU \u53ef\u4ee5\u76f4\u63a5\u5199\u5165 NIC \u7684 Doorbell \u5bc4\u5b58\u5668\u5730\u5740\u7a7a\u95f4\uff0c\u65e0\u9700\u7b49\u5f85 CPU\u3002<\/li>\n\n\n\n<li>\u4f7f\u7528\u5185\u5b58\u6620\u5c04 I\/O\uff08MMIO\uff09\u65b9\u5f0f\u5b9e\u73b0\u4f4e\u5ef6\u8fdf\u89e6\u53d1\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u96f6\u62f7\u8d1d RDMA \u901a\u4fe1\n<ul class=\"wp-block-list\">\n<li>\u6570\u636e\u53ef\u4ee5\u76f4\u63a5\u4ece GPU \u663e\u5b58\u901a\u8fc7 IB \u7f51\u7edc\u4f20\u8f93\u5230\u8fdc\u7a0b\u8282\u70b9\u7684 GPU \u663e\u5b58\uff0c\u65e0\u9700\u4e2d\u95f4 CPU \u6216\u7cfb\u7edf\u5185\u5b58\u62f7\u8d1d\u3002<\/li>\n\n\n\n<li>\u7279\u522b\u9002\u7528\u4e8e EP\uff08Expert Parallelism\uff09\u4e2d\u7684 All-to-All \u901a\u4fe1\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u8bb8\u591a\u7814\u7a76\uff08\u5305\u62ec DeepSeek \u56e2\u961f\u5f00\u53d1\u7684 DeepEP \uff09\u90fd\u91c7\u7528\u4e86 IBGDA \u6280\u672f\uff0c\u5e76\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">6. \u8ba8\u8bba\u4e0e\u5bf9\u672a\u6765\u786c\u4ef6\u67b6\u6784\u8bbe\u8ba1\u7684\u6d1e\u89c1<\/h2>\n\n\n\n<p>\u672c\u7ae0\u8282\u57fa\u4e8e DeepSeek-V3 \u7684\u5b9e\u8df5\u7ecf\u9a8c\uff0c\u7cfb\u7edf\u6027\u5730\u603b\u7ed3\u4e86\u5f53\u524d\u786c\u4ef6\u67b6\u6784\u5728\u5927\u89c4\u6a21 AI \u5de5\u4f5c\u8d1f\u8f7d\u4e0b\u7684\u5c40\u9650\u6027\uff0c\u5e76\u63d0\u51fa\u4e86\u9762\u5411\u672a\u6765\u7684\u521b\u65b0\u65b9\u5411\u3002\u4e3b\u8981\u56f4\u7ed5\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\u5c55\u5f00\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u786c\u4ef6\u9650\u5236\u7684\u5b9e\u9645\u5f71\u54cd<\/li>\n\n\n\n<li>\u5173\u952e\u6311\u6218\u4e0e\u74f6\u9888\u5206\u6790<\/li>\n\n\n\n<li>\u9762\u5411\u672a\u6765\u7684\u786c\u4ef6\u4f18\u5316\u5efa\u8bae<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">6.1 Robustness\uff08\u9c81\u68d2\u6027\uff09\u6311\u6218<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4e92\u8fde\u5931\u8d25\uff08Interconnect Failures\uff09\n<ul class=\"wp-block-list\">\n<li>\u5f53\u524d\u9ad8\u6027\u80fd\u4e92\u8fde\u534f\u8bae\uff08\u5982 InfiniBand \u548c NVLink\uff09\u5bb9\u6613\u51fa\u73b0\u95f4\u6b47\u6027\u65ad\u5f00\u3002<\/li>\n\n\n\n<li>\u5728\u901a\u4fe1\u5bc6\u96c6\u578b\u4efb\u52a1\uff08\u5982 MoE \u4e2d\u7684 EP \u4e13\u5bb6\u5e76\u884c\uff09\u4e2d\uff0c\u5373\u4f7f\u662f\u77ed\u6682\u4e2d\u65ad\u4e5f\u53ef\u80fd\u5bfc\u81f4\u6027\u80fd\u663e\u8457\u4e0b\u964d\u6216\u8bad\u7ec3\u5931\u8d25\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5355\u70b9\u786c\u4ef6\u6545\u969c\uff08Single Hardware Failures\uff09\n<ul class=\"wp-block-list\">\n<li>\u5305\u62ec\u8282\u70b9\u5d29\u6e83\u3001GPU \u6545\u969c\u3001ECC \u5185\u5b58\u9519\u8bef\u7b49\u3002<\/li>\n\n\n\n<li>\u5728\u8d85\u5927\u89c4\u6a21\u90e8\u7f72\u4e2d\uff0c\u8fd9\u7c7b\u6545\u969c\u7684\u6982\u7387\u968f\u7cfb\u7edf\u89c4\u6a21\u589e\u52a0\u800c\u4e0a\u5347\uff0c\u6062\u590d\u6210\u672c\u9ad8\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u9759\u9ed8\u6570\u636e\u635f\u574f\uff08Silent Data Corruption\uff09\n<ul class=\"wp-block-list\">\n<li>\u591a\u6bd4\u7279\u5185\u5b58\u7ffb\u8f6c\u6216\u8ba1\u7b97\u9519\u8bef\u53ef\u80fd\u672a\u88ab ECC \u68c0\u6d4b\u5230\u3002<\/li>\n\n\n\n<li>\u9519\u8bef\u5728\u957f\u671f\u4efb\u52a1\u4e2d\u4f20\u64ad\uff0c\u6700\u7ec8\u5f71\u54cd\u6a21\u578b\u8d28\u91cf\u3002<\/li>\n\n\n\n<li>\u76ee\u524d\u4f9d\u8d56\u5e94\u7528\u5c42\u542f\u53d1\u5f0f\u7b56\u7565\u8fdb\u884c\u68c0\u6d4b\uff0c\u65e0\u6cd5\u4fdd\u8bc1\u6574\u4f53\u7cfb\u7edf\u7684\u53ef\u9760\u6027\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u5efa\u8bae\u6539\u8fdb\u63aa\u65bd\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u589e\u5f3a\u7684\u9519\u8bef\u68c0\u6d4b\u673a\u5236\n<ul class=\"wp-block-list\">\n<li>\u786c\u4ef6\u5e94\u5f15\u5165\u6bd4\u4f20\u7edf ECC \u66f4\u5148\u8fdb\u7684\u9519\u8bef\u68c0\u6d4b\u673a\u5236\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5982 \u57fa\u4e8e\u6821\u9a8c\u548c\u7684\u9a8c\u8bc1 \uff08checksum-based validation\uff09<\/li>\n\n\n\n<li>\u6216 \u786c\u4ef6\u52a0\u901f\u5197\u4f59\u68c0\u67e5 \uff08hardware-accelerated redundancy checks\uff09<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u63d0\u4f9b\u5b8c\u6574\u7684\u8bca\u65ad\u5de5\u5177\u5305\n<ul class=\"wp-block-list\">\n<li>\u786c\u4ef6\u5382\u5546\u5e94\u4e3a\u7ec8\u7aef\u7528\u6237\u63d0\u4f9b\u5168\u9762\u7684\u8bca\u65ad\u5de5\u5177\uff0c\u5e2e\u52a9\u7528\u6237\u9a8c\u8bc1\u7cfb\u7edf\u5b8c\u6574\u6027\u3002<\/li>\n\n\n\n<li>\u8fd9\u4e9b\u5de5\u5177\u5e94\u4f5c\u4e3a\u6807\u51c6\u786c\u4ef6\u5957\u4ef6\u7684\u4e00\u90e8\u5206\uff0c\u652f\u6301\u6301\u7eed\u76d1\u63a7\u548c\u9884\u9632\u6027\u4fee\u590d\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u63d0\u5347\u7cfb\u7edf\u53ef\u4fe1\u5ea6\n<ul class=\"wp-block-list\">\n<li>\u901a\u8fc7\u5d4c\u5165\u5f0f\u7684\u8bca\u65ad\u548c\u9a8c\u8bc1\u673a\u5236\uff0c\u589e\u5f3a\u6574\u4e2a\u751f\u547d\u5468\u671f\u5185\u7684\u7cfb\u7edf\u900f\u660e\u6027\u548c\u53ef\u4fe1\u6027\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">6.2 CPU \u6027\u80fd\u74f6\u9888\u4e0e\u4e92\u8fde\u4f18\u5316<\/h3>\n\n\n\n<p><br>\u5c3d\u7ba1 GPU \u662f\u6df1\u5ea6\u5b66\u4e60\u7684\u4e3b\u8981\u8ba1\u7b97\u5355\u5143\uff0c\u4f46 CPU \u5728\u534f\u8c03\u8ba1\u7b97\u3001\u7ba1\u7406 I\/O \u548c\u7ef4\u6301\u541e\u5410\u65b9\u9762\u4ecd\u4e0d\u53ef\u66ff\u4ee3\u3002\u7136\u800c\uff0c\u5f53\u524d\u67b6\u6784\u5b58\u5728\u591a\u4e2a\u74f6\u9888\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>PCIe \u63a5\u53e3\u5e26\u5bbd\u4e0d\u8db3\n<ul class=\"wp-block-list\">\n<li>PCIe \u6210\u4e3a CPU \u4e0e GPU \u4e4b\u95f4\u7684\u4e3b\u8981\u901a\u4fe1\u63a5\u53e3\uff0c\u4f46\u5728\u5927\u89c4\u6a21\u53c2\u6570\u3001\u68af\u5ea6\u6216 KV Cache \u4f20\u8f93\u65f6\u6210\u4e3a\u74f6\u9888\u3002<\/li>\n\n\n\n<li>\u4f8b\u5982\uff0c\u9971\u548c 160 \u6761 PCIe 5.0 \u901a\u9053\u9700\u8981\u8d85\u8fc7 640 GB\/s \u7684\u5e26\u5bbd\uff0c\u5bf9\u5185\u5b58\u5e26\u5bbd\u63d0\u51fa\u9ad8\u8fbe 1 TB\/s per node \u7684\u9700\u6c42\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5185\u5b58\u5e26\u5bbd\u9650\u5236\n<ul class=\"wp-block-list\">\n<li>\u5f53\u524d DRAM \u67b6\u6784\u96be\u4ee5\u6ee1\u8db3\u5982\u6b64\u9ad8\u7684\u5e26\u5bbd\u8981\u6c42\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5ef6\u8fdf\u654f\u611f\u4efb\u52a1\u7684\u6027\u80fd\u74f6\u9888\n<ul class=\"wp-block-list\">\n<li>\u5185\u6838\u542f\u52a8\u3001\u7f51\u7edc\u5904\u7406\u7b49\u4efb\u52a1\u5bf9\u5355\u6838 CPU \u6027\u80fd\u8981\u6c42\u9ad8\uff0c\u901a\u5e38\u9700\u8981\u57fa\u7840\u9891\u7387\u9ad8\u4e8e 4 GHz \u3002<\/li>\n\n\n\n<li>\u6bcf\u4e2a GPU \u9700\u8981\u8db3\u591f\u591a\u7684 CPU \u6838\u5fc3\u6765\u907f\u514d\u63a7\u5236\u8def\u5f84\u4e0a\u7684\u74f6\u9888\u3002<\/li>\n\n\n\n<li>\u5bf9\u4e8e chiplet \u67b6\u6784\uff0c\u8fd8\u9700\u8981\u989d\u5916\u6838\u5fc3\u652f\u6301\u7f13\u5b58\u611f\u77e5\u7684\u4efb\u52a1\u5212\u5206\u4e0e\u9694\u79bb\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>6.3 Toward Intelligent Networks for AI<\/strong><\/h3>\n\n\n\n<p>\u4e3a\u5e94\u5bf9 AI \u8bad\u7ec3\u4e0e\u63a8\u7406\u4e2d\u5bf9\u4f4e\u5ef6\u8fdf\u548c\u9ad8\u541e\u5410\u7684\u4e25\u82db\u8981\u6c42\uff0c\u672a\u6765\u7684\u7f51\u7edc\u67b6\u6784\u5e94\u5177\u5907\u7684\u4e94\u5927\u5173\u952e\u80fd\u529b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f4e\u5ef6\u8fdf\u901a\u4fe1<\/li>\n\n\n\n<li>\u81ea\u9002\u5e94\u6d41\u91cf\u8c03\u5ea6<\/li>\n\n\n\n<li>\u9ad8\u6548\u7684\u62e5\u585e\u63a7\u5236<\/li>\n\n\n\n<li>\u5f3a\u9c81\u68d2\u6027\u548c\u5bb9\u9519\u673a\u5236<\/li>\n\n\n\n<li>\u7075\u6d3b\u7684\u5e26\u5bbd\u5206\u914d\u4e0e\u4f18\u5148\u7ea7\u7ba1\u7406<\/li>\n<\/ul>\n\n\n\n<p><strong>6.3.1 Co-Packaged Optics\uff08\u5171\u5c01\u88c5\u5149\u5b66\u6280\u672f\uff09<\/strong><\/p>\n\n\n\n<p>\u5728\u8d85\u5927\u89c4\u6a21\u5206\u5e03\u5f0f\u7cfb\u7edf\u4e2d\uff0c\u4f20\u7edf\u7535\u6c14\u63a5\u53e3\u7684\u5e26\u5bbd\u5bc6\u5ea6\u548c\u80fd\u6548\u5df2\u63a5\u8fd1\u6781\u9650\uff0c\u5f15\u5165&nbsp;<strong>\u7845\u5149\u5b50\uff08Silicon Photonics\uff09\u6280\u672f&nbsp;<\/strong>\uff0c\u5c06\u5149\u5b66\u8fde\u63a5\u76f4\u63a5\u96c6\u6210\u5230\u82af\u7247\u5c01\u88c5\u4e2d\uff0c\u53ef\u4ee5\u83b7\u5f97\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u66f4\u9ad8\u7684\u5e26\u5bbd\u5bc6\u5ea6\uff08Bandwidth Density\uff09<\/li>\n\n\n\n<li>\u66f4\u4f4e\u7684\u529f\u8017\uff08Energy Efficiency\uff09<\/li>\n<\/ul>\n\n\n\n<p><strong>6.3.2 Lossless Network\uff08\u65e0\u635f\u7f51\u7edc\uff09<\/strong><\/p>\n\n\n\n<p>AI\u5de5\u4f5c\u8d1f\u8f7d\uff08\u5c24\u5176\u662f MoE \u4e2d\u7684 All-to-All \u64cd\u4f5c\uff09\u5bf9\u4e22\u5305\u6781\u4e3a\u654f\u611f\uff0c\u5373\u4f7f\u662f\u6781\u5c0f\u7684\u4e22\u5305\u7387\u4e5f\u53ef\u80fd\u5bfc\u81f4\u6027\u80fd\u663e\u8457\u4e0b\u964d\uff0c\u53ef\u4ee5\u4f7f\u7528&nbsp;<strong>Credit-Based Flow Control (CBFC)&nbsp;<\/strong>\u5b9e\u73b0\u65e0\u4e22\u5305\u4f20\u8f93\uff0cCBFC \u539f\u7406\u662f\u901a\u8fc7\u63a5\u6536\u7aef\u53d1\u653e\u4fe1\u7528\u989d\u5ea6\uff08credit\uff09\uff0c\u53d1\u9001\u7aef\u53ea\u6709\u5728\u83b7\u5f97\u4fe1\u7528\u540e\u624d\u53ef\u53d1\u9001\u6570\u636e\u3002\u4f46\u4f1a\u5b58\u5728\u4ee5\u4e0b\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5982\u679c\u4ec5\u4f9d\u8d56 CBFC\uff0c\u5bb9\u6613\u5f15\u53d1\u4e25\u91cd\u7684\u00a0<strong>Head-of-Line Blocking\uff08HOLB\uff09\u00a0<\/strong>\uff0c\u5f71\u54cd\u6574\u4f53\u541e\u5410\u3002<\/li>\n\n\n\n<li>\u5728\u591a\u8def\u5f84\u901a\u4fe1\u4e2d\uff0c\u5355\u4e00\u8def\u5f84\u4e0a\u7684\u963b\u585e\u53ef\u80fd\u5bfc\u81f4\u6574\u4e2a\u6d41\u88ab\u5361\u4f4f<\/li>\n<\/ul>\n\n\n\n<p>\u901a\u8fc7\u5f15\u5165&nbsp;<strong>Endpoint-driven Congestion Control\uff0c<\/strong>\u5982 RTT-based CC\uff08RTTCC\uff09\u6216 Programmed Congestion Control\uff08PCC\uff09\uff0c\u80fd\u591f\u4e3b\u52a8\u8c03\u8282\u6ce8\u5165\u901f\u7387\uff0c\u907f\u514d\u6781\u7aef\u62e5\u585e\u573a\u666f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">6.3.3&nbsp;<strong>Adaptive Routing\uff08\u81ea\u9002\u5e94\u8def\u7531\uff09<\/strong><\/h3>\n\n\n\n<p>\u4f20\u7edf\u7684 ECMP\uff08Equal-Cost Multi-Path\uff09\u8def\u7531\u7b56\u7565\u65e0\u6cd5\u6709\u6548\u5904\u7406 AI \u5de5\u4f5c\u8d1f\u8f7d\u4e2d\u9ad8\u5ea6\u96c6\u4e2d\u4e14\u7f3a\u4e4f\u968f\u673a\u6027\u7684\u901a\u4fe1\u6a21\u5f0f\uff08\u5982 Data Parallelism \u7684 AllReduce \u6216 MoE \u7684 All-to-All\uff09\u3002\u81ea\u9002\u5e94\u8def\u7531\u65b9\u6cd5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u52a8\u6001\u9009\u62e9\u6700\u4f18\u8def\u5f84\uff0c\u907f\u514d\u70ed\u70b9\u74f6\u9888\u3002<\/li>\n\n\n\n<li>\u63d0\u5347\u7f51\u7edc\u5229\u7528\u7387\uff0c\u51cf\u5c11\u901a\u4fe1\u5ef6\u8fdf\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u53ef\u4ee5\u663e\u8457\u7f13\u89e3 All-to-All \u548c Reduce-Scatter \u7c7b\u64cd\u4f5c\u4e2d\u7684\u70ed\u70b9\u95ee\u9898\uff0c\u63d0\u9ad8\u96c6\u7fa4\u6574\u4f53\u901a\u4fe1\u6548\u7387\u548c\u7a33\u5b9a\u6027<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">6.3.4&nbsp;<strong>Efficient Fault-Tolerant Protocols\uff08\u9ad8\u6548\u5bb9\u9519\u534f\u8bae\uff09<\/strong><\/h3>\n\n\n\n<p>\u5927\u89c4\u6a21\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u7f51\u7edc\u94fe\u8def\u4e2d\u65ad\u3001\u786c\u4ef6\u6545\u969c\u7b49\u5f02\u5e38\u60c5\u51b5\u96be\u4ee5\u907f\u514d\u3002\u82e5\u4e0d\u80fd\u5feb\u901f\u6062\u590d\uff0c\u53ef\u80fd\u9020\u6210\u8bad\u7ec3\u4e2d\u65ad\u3001\u6a21\u578b\u635f\u574f\u7b49\u95ee\u9898\u3002\u53ef\u4ee5\u5f15\u5165\u4ee5\u4e0b\u65b9\u6cd5\u89e3\u51b3\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Self-healing Protocols\uff08\u81ea\u6108\u534f\u8bae\uff09 \uff1a\u81ea\u52a8\u68c0\u6d4b\u5e76\u4fee\u590d\u65ad\u5f00\u8fde\u63a5\u3002<\/li>\n\n\n\n<li>Redundant Ports\uff08\u5197\u4f59\u7aef\u53e3\uff09 \uff1a\u63d0\u4f9b\u591a\u6761\u8def\u5f84\uff0c\u9632\u6b62\u5355\u70b9\u6545\u969c\u3002<\/li>\n\n\n\n<li>Rapid Failover Techniques\uff08\u5feb\u901f\u5207\u6362\u673a\u5236\uff09 \uff1a\u5982\u94fe\u8def\u5c42\u91cd\u8bd5\u673a\u5236\u3001Selective Retransmission\uff08\u9009\u62e9\u6027\u91cd\u4f20\uff09<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>6.3.5 Dynamic Resource Management\uff08\u52a8\u6001\u8d44\u6e90\u7ba1\u7406\uff09<\/strong><\/h3>\n\n\n\n<p>AI \u7cfb\u7edf\u4e2d\u5e38\u5e38\u5b58\u5728\u6df7\u5408\u5de5\u4f5c\u8d1f\u8f7d\uff08\u5982\u540c\u65f6\u8fdb\u884c\u8bad\u7ec3\u4e0e\u63a8\u7406\uff09\uff0c\u4e0d\u540c\u4efb\u52a1\u5bf9\u5e26\u5bbd\u3001\u5ef6\u8fdf\u3001\u4f18\u5148\u7ea7\u7684\u9700\u6c42\u5dee\u5f02\u5927\u3002\u4f18\u5316\u65b9\u5411\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u52a8\u6001\u5e26\u5bbd\u5206\u914d\u00a0<\/strong>\uff1a\u6839\u636e\u4efb\u52a1\u7c7b\u578b\uff08\u5982 EP\u3001TP\u3001KV Cache \u4f20\u8f93\uff09\u52a8\u6001\u8c03\u6574\u53ef\u7528\u5e26\u5bbd\u3002<\/li>\n\n\n\n<li><strong>\u6d41\u91cf\u4f18\u5148\u7ea7\u7ba1\u7406\u00a0<\/strong>\uff1a\u533a\u5206\u63a8\u7406\u4efb\u52a1\u4e0e\u8bad\u7ec3\u4efb\u52a1\uff0c\u5728\u7edf\u4e00\u96c6\u7fa4\u4e2d\u4fdd\u8bc1\u63a8\u7406\u4efb\u52a1\u7684\u54cd\u5e94\u901f\u5ea6\u3002<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td>\u65b9\u5411<\/td><td>\u5173\u952e\u6280\u672f<\/td><td>\u4f18\u70b9<\/td><\/tr><tr><td>Co-Packaged Optics<\/td><td>\u7845\u5149\u5b50\u96c6\u6210<\/td><td>\u63d0\u5347\u5e26\u5bbd\u5bc6\u5ea6\u4e0e\u80fd\u6548<\/td><\/tr><tr><td>Lossless Network<\/td><td>CBFC + Endpoint-driven CC<\/td><td>\u907f\u514d\u4e22\u5305\uff0c\u9632\u6b62 HOLB\uff08Head-of-Line Blocking\uff09<\/td><\/tr><tr><td>Adaptive Routing<\/td><td>Packet Spraying \/ Congestion-aware Routing<\/td><td>\u7f13\u89e3\u70ed\u70b9\uff0c\u63d0\u5347\u8d1f\u8f7d\u5747\u8861<\/td><\/tr><tr><td>Fault-Tolerant Protocols<\/td><td>\u81ea\u6108\u673a\u5236 \/ \u5197\u4f59\u7aef\u53e3 \/ \u5feb\u901f\u5207\u6362<\/td><td>\u63d0\u5347\u7cfb\u7edf\u9c81\u68d2\u6027<\/td><\/tr><tr><td>Dynamic Resource Management<\/td><td>\u52a8\u6001\u5e26\u5bbd\u5206\u914d \/ \u6d41\u91cf\u4f18\u5148\u7ea7\u63a7\u5236<\/td><td>\u4f18\u5316\u6df7\u5408\u5de5\u4f5c\u8d1f\u8f7d\u8868\u73b0<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">6.4 \u5185\u5b58\u8bed\u4e49\u901a\u4fe1\u4e0e\u987a\u5e8f\u95ee\u9898\uff08Memory-Semantic Communication and Ordering Issue\uff09<\/h3>\n\n\n\n<p>\u4f7f\u7528 load\/store \u6a21\u578b\u8fdb\u884c\u8282\u70b9\u95f4\u901a\u4fe1\u867d\u7136\u9ad8\u6548\u4e14\u6613\u4e8e\u7f16\u7a0b\uff0c\u4f46\u5f53\u524d\u5b9e\u73b0\u53d7\u9650\u4e8e\u5185\u5b58\u987a\u5e8f\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u53d1\u9001\u65b9\u5199\u5165\u6570\u636e\u540e\u5fc5\u987b\u663e\u5f0f\u53d1\u51fa\u5185\u5b58\u5c4f\u969c\uff08fence\uff09\uff0c\u7136\u540e\u624d\u80fd\u66f4\u65b0\u6807\u5fd7\u901a\u77e5\u63a5\u6536\u65b9\u3002<\/li>\n\n\n\n<li>\u8fd9\u79cd\u4e25\u683c\u987a\u5e8f\u6027\u5f15\u5165\u4e86\u989d\u5916\u7684\u5f80\u8fd4\u65f6\u95f4\uff08RTT\uff09\u5ef6\u8fdf\uff0c\u5e76\u53ef\u80fd\u5bfc\u81f4\u7ebf\u7a0b\u963b\u585e\uff0c\u5f71\u54cd\u541e\u5410\u91cf\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6b64\u5916\uff0c\u5728\u6d88\u606f\u8bed\u4e49 RDMA \u4e2d\u4e5f\u5b58\u5728\u7c7b\u4f3c\u95ee\u9898\uff0c\u4f8b\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728 InfiniBand \u6216 NVIDIA BlueField-3 \u4e0a\u6267\u884c regular RDMA writes \u540e\u8fdb\u884c packet spraying \u7684 RDMA \u539f\u5b50\u64cd\u4f5c\uff0c\u4e5f\u53ef\u80fd\u5bfc\u81f4\u4e71\u5e8f\u540c\u6b65\u95ee\u9898\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>\u5efa\u8bae<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u63d0\u4f9b\u786c\u4ef6\u7ea7\u987a\u5e8f\u4fdd\u969c\u673a\u5236<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u786c\u4ef6\u5e94\u539f\u751f\u652f\u6301 memory-semantic \u901a\u4fe1\u7684\u987a\u5e8f\u4e00\u81f4\u6027\u3002<\/li>\n\n\n\n<li>\u7f16\u7a0b\u5c42\u5e94\u901a\u8fc7 acquire\/release \u8bed\u4e49\u786e\u4fdd\u4e00\u81f4\u6027\u3002<\/li>\n\n\n\n<li>\u63a5\u6536\u7aef\u5e94\u7531\u786c\u4ef6\u4fdd\u8bc1\u6570\u636e\u6709\u5e8f\u4ea4\u4ed8\uff0c\u907f\u514d\u989d\u5916\u5f00\u9500\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>Region Acquire\/Release (RAR) \u673a\u5236<\/strong>\n<ul class=\"wp-block-list\">\n<li>\u63a5\u6536\u7aef\u786c\u4ef6\u7ef4\u62a4\u4e00\u4e2a bitmap\uff0c\u8ddf\u8e2a RNR\uff08Receiver Not Ready\uff09\u5185\u5b58\u533a\u57df\u7684\u72b6\u6001\u3002<\/li>\n\n\n\n<li>acquire\/release \u64cd\u4f5c\u4f5c\u7528\u4e8e RAR \u5730\u5740\u8303\u56f4\u5185\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u4f18\u70b9<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6781\u4f4e bitmap \u5f00\u9500\uff1b<\/li>\n\n\n\n<li>\u5b9e\u73b0\u9ad8\u6548\u7684\u786c\u4ef6\u5f3a\u5236\u6392\u5e8f\uff1b<\/li>\n\n\n\n<li>\u6d88\u9664\u53d1\u9001\u7aef\u663e\u5f0f\u7684 fence \u64cd\u4f5c\uff1b<\/li>\n\n\n\n<li>\u652f\u6301 memory-semantic \u548c message-semantic RDMA \u64cd\u4f5c\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">6.5 \u7f51\u7edc\u5185\u8ba1\u7b97\u4e0e\u538b\u7f29\uff08In-Network Computation and Compression\uff09<\/h3>\n\n\n\n<p>MoE \u5305\u542b\u4e24\u4e2a All-to-All \u9636\u6bb5\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\uff081\uff09Dispatch \u9636\u6bb5\uff08\u8def\u7531\uff09<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7c7b\u4f3c\u5c0f\u89c4\u6a21\u591a\u64ad\uff08multicast\uff09\u64cd\u4f5c\u3002<\/li>\n\n\n\n<li>\u5355\u4e2a token \u9700\u8981\u88ab\u8f6c\u53d1\u5230\u591a\u4e2a\u4e13\u5bb6\u8282\u70b9\u3002<\/li>\n\n\n\n<li>\u5efa\u8bae\uff1a\u786c\u4ef6\u5e94\u652f\u6301\u81ea\u52a8\u5305\u590d\u5236\u4e0e\u591a\u76ee\u6807\u8f6c\u53d1\uff0c\u5927\u5e45\u51cf\u5c11\u901a\u4fe1\u5f00\u9500\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\uff082\uff09Combine \u9636\u6bb5\uff08\u805a\u5408\uff09<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7c7b\u4f3c\u5c0f\u89c4\u6a21\u5f52\u7ea6\uff08reduce\uff09\u64cd\u4f5c\u3002<\/li>\n\n\n\n<li>\u591a\u4e2a\u4e13\u5bb6\u8f93\u51fa\u7ed3\u679c\u9700\u805a\u5408\u56de\u539f\u59cb\u8282\u70b9\u3002<\/li>\n\n\n\n<li>\u5f53\u524d\u6311\u6218\uff1aEP \u7684\u5f52\u7ea6\u8303\u56f4\u8f83\u5c0f\u4e14\u8d1f\u8f7d\u4e0d\u5747\u8861\uff0c\u96be\u4ee5\u7075\u6d3b\u5b9e\u73b0\u7f51\u7edc\u5185\u805a\u5408\u3002<\/li>\n<\/ul>\n\n\n\n<p>LogFMT \u662f DeepSeek-V3 \u4e2d\u7528\u4e8e\u964d\u4f4e\u901a\u4fe1\u91cf\u7684\u65b0\u578b\u5bf9\u6570\u91cf\u5316\u683c\u5f0f\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u53ef\u4ee5\u5c06\u5176\u539f\u751f\u96c6\u6210\u8fdb\u7f51\u7edc\u786c\u4ef6\uff0c\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u71b5\u5bc6\u5ea6\uff0c\u964d\u4f4e\u5e26\u5bbd\u9700\u6c42\u3002<\/li>\n\n\n\n<li>\u53e6\u5916\uff0c\u5f00\u53d1\u786c\u4ef6\u52a0\u901f\u7684 LogFMT \u538b\u7f29\u4e0e\u89e3\u538b\u5355\u5143\uff0c\u5b9e\u73b0\u65e0\u7f1d\u96c6\u6210\u4e8e\u5206\u5e03\u5f0f\u7cfb\u7edf\uff0c\u63d0\u5347\u6574\u4f53\u541e\u5410\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">6.6 \u4ee5\u5185\u5b58\u4e3a\u4e2d\u5fc3\u7684\u521b\u65b0\uff08Memory-Centric Innovations\uff09<\/h3>\n\n\n\n<p>\u968f\u7740\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u5448\u6307\u6570\u589e\u957f\uff0c\u4f20\u7edf\u9ad8\u5e26\u5bbd\u5185\u5b58\uff08\u5982 HBM\uff09\u7684\u53d1\u5c55\u901f\u5ea6\u8fdc\u8fdc\u8ddf\u4e0d\u4e0a\u9700\u6c42\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>DRAM \u5806\u53e0\u52a0\u901f\u5668\uff08DRAM-Stacked Accelerators\uff09\n<ul class=\"wp-block-list\">\n<li>\u5229\u7528\u5148\u8fdb\u7684 3D \u5806\u53e0\u6280\u672f\uff0c\u5c06 DRAM \u5782\u76f4\u96c6\u6210\u5728\u903b\u8f91\u82af\u7247\u4e4b\u4e0a\u3002<\/li>\n\n\n\n<li>\u4f18\u70b9\uff1a\n<ul class=\"wp-block-list\">\n<li>\u6781\u9ad8\u7684\u5185\u5b58\u5e26\u5bbd\uff1b<\/li>\n\n\n\n<li>\u8d85\u4f4e\u5ef6\u8fdf\uff1b<\/li>\n\n\n\n<li>\u5b9e\u7528\u7684\u5185\u5b58\u5bb9\u91cf\uff08\u53d7\u5806\u53e0\u5c42\u6570\u9650\u5236\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5e94\u7528\u573a\u666f\uff1a\n<ul class=\"wp-block-list\">\n<li>\u7279\u522b\u9002\u7528\u4e8e MoE \u63a8\u7406\u4e2d\u5185\u5b58\u541e\u5410\u662f\u74f6\u9888\u7684\u573a\u666f\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u793a\u4f8b\u67b6\u6784\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5982 SeDRAM\uff0c\u4e3a\u5185\u5b58\u5bc6\u96c6\u578b\u4efb\u52a1\u63d0\u4f9b\u524d\u6240\u672a\u6709\u7684\u6027\u80fd\u63d0\u5347\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u6676\u5706\u7ea7\u7cfb\u7edf\uff08System-on-Wafer, SoW\uff09\n<ul class=\"wp-block-list\">\n<li>\u4f7f\u7528 wafer-scale \u96c6\u6210\u6280\u672f\uff0c\u6700\u5927\u5316\u8ba1\u7b97\u5bc6\u5ea6\u4e0e\u5185\u5b58\u5e26\u5bbd\u3002<\/li>\n\n\n\n<li>\u4f18\u70b9\uff1a\n<ul class=\"wp-block-list\">\n<li>\u652f\u6301\u8d85\u5927\u89c4\u6a21\u6a21\u578b\u90e8\u7f72\uff1b<\/li>\n\n\n\n<li>\u663e\u8457\u63d0\u5347\u7247\u4e0a\u901a\u4fe1\u6548\u7387\uff1b<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5e94\u7528\u573a\u666f\uff1a\n<ul class=\"wp-block-list\">\n<li>\u7279\u522b\u9002\u5408\u5904\u7406\u4e07\u4ebf\u53c2\u6570\u7ea7\u522b\u7684\u6a21\u578b\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">7. \u603b\u7ed3<\/h3>\n\n\n\n<p>DeepSeek-V3 \u662f\u786c\u4ef6\u4e0e\u8f6f\u4ef6\u534f\u540c\u8bbe\u8ba1\u5728\u5927\u89c4\u6a21 AI \u7cfb\u7edf\u4e2d\u53d6\u5f97\u7a81\u7834\u7684\u5178\u8303\u3002\u901a\u8fc7\u6df1\u5165\u5206\u6790\u5f53\u524d\u786c\u4ef6\u67b6\u6784\u7684\u5c40\u9650\u6027\uff0c\u5e76\u63d0\u51fa\u4e00\u7cfb\u5217\u5207\u5b9e\u53ef\u884c\u7684\u4f18\u5316\u5efa\u8bae\uff0c\u8be5\u7814\u7a76\u4e3a\u4e0b\u4e00\u4ee3\u9762\u5411\u5927\u6a21\u578b\u7684\u786c\u4ef6\u8bbe\u8ba1\u63d0\u4f9b\u4e86\u6e05\u6670\u7684\u53d1\u5c55\u84dd\u56fe\u3002<\/p>\n\n\n\n<p>\u8be5\u8bba\u6587\u4e0d\u4ec5\u662f DeepSeek-V3 \u7684\u6280\u672f\u603b\u7ed3\uff0c\u4e5f\u4e3a\u672a\u6765 AI \u786c\u4ef6\u67b6\u6784\u7684\u53d1\u5c55\u63d0\u4f9b\u4e86\u5b9d\u8d35\u7684\u5b9e\u8df5\u6307\u5bfc\u3002\u968f\u7740 AI \u5de5\u4f5c\u8d1f\u8f7d\u7684\u6301\u7eed\u589e\u957f\uff0c\u8fd9\u79cd\u201c\u6a21\u578b + \u786c\u4ef6\u201d\u7684\u8f6f\u786c\u4ef6\u534f\u540c\u4f18\u5316\u8bbe\u8ba1\u8bf6\u5c06\u6210\u4e3a\u6784\u5efa\u9ad8\u6548\u3001\u53ef\u6301\u7eed AI \u7cfb\u7edf\u7684\u5173\u952e\u8def\u5f84\u3002<\/p>\n\n\n\n<p>\u5176\u6838\u5fc3\u7ecf\u9a8c\u4e3b\u8981\u5305\u62ec\uff1a<\/p>\n\n\n\n<p><strong>\u672a\u6765\u786c\u4ef6\u8bbe\u8ba1\u65b9\u5411<\/strong>\uff08\u5982\u7edf\u4e00\u7f51\u7edc\u63a5\u53e3\u3001\u4e13\u7528\u534f\u5904\u7406\u5668\uff09\u63a8\u52a8\u4e0b\u4e00\u4ee3 AI \u7cfb\u7edf\u53d1\u5c55\u3002<\/p>\n\n\n\n<p><strong>\u6a21\u578b\u67b6\u6784\u521b\u65b0<\/strong>\uff08\u5982 MLA\u3001MoE\u3001MTP\uff09\u63d0\u5347\u6548\u7387\u3002<\/p>\n\n\n\n<p><strong>\u7f51\u7edc\u62d3\u6251\u4f18\u5316<\/strong>\uff08\u5982 MPFT\uff09\u964d\u4f4e\u6210\u672c\u5e76\u63d0\u5347\u6269\u5c55\u6027\u3002<\/p>\n\n\n\n<p><strong>\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u4e0e\u901a\u4fe1\u4f18\u5316<\/strong>\uff08\u5982 FP8\u3001\u538b\u7f29\u3001offload\uff09\u964d\u4f4e\u5e26\u5bbd\u9700\u6c42\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>https:\/\/www.arxiv.org\/pdf\/2505.09343 \u6df1\u5ea6\u89e3\u6790 DeepSeek-V3\uff1a\u6269 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6,11],"tags":[],"class_list":["post-535","post","type-post","status-publish","format-standard","hentry","category-ai","category-11"],"_links":{"self":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts\/535","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/comments?post=535"}],"version-history":[{"count":1,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts\/535\/revisions"}],"predecessor-version":[{"id":545,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts\/535\/revisions\/545"}],"wp:attachment":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/media?parent=535"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/categories?post=535"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/tags?post=535"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}