{"id":8364,"date":"2026-06-09T14:03:12","date_gmt":"2026-06-09T06:03:12","guid":{"rendered":"https:\/\/ab123.xyz\/?p=8364"},"modified":"2026-06-09T14:03:12","modified_gmt":"2026-06-09T06:03:12","slug":"google-pathways-on-tpu-v5e-for-mixture-of-experts%ef%bc%9a%e9%ab%98%e6%95%88ai%e6%8e%a8%e7%90%86%e7%9a%84%e4%b8%8b%e4%b8%80%e4%bb%a3%e5%bc%95%e6%93%8e","status":"publish","type":"post","link":"https:\/\/ab123.xyz\/?p=8364","title":{"rendered":"Google Pathways on TPU v5e for Mixture of Experts\uff1a\u9ad8\u6548AI\u63a8\u7406\u7684\u4e0b\u4e00\u4ee3\u5f15\u64ce"},"content":{"rendered":"<p>\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\uff0c\u5927\u89c4\u6a21\u7a00\u758f\u6a21\u578b\u7684\u8bad\u7ec3\u4e0e\u63a8\u7406\u59cb\u7ec8\u9762\u4e34\u7b97\u529b\u4e0e\u6548\u7387\u7684\u53cc\u91cd\u6311\u6218\u3002Google \u6700\u65b0\u63a8\u51fa\u7684 <strong>Pathways<\/strong> \u7cfb\u7edf\u7ed3\u5408 <strong>TPU v5e<\/strong> \u52a0\u901f\u5668\uff0c\u4e3a <strong>Mixture of Experts\uff08MoE\uff09<\/strong> \u67b6\u6784\u63d0\u4f9b\u4e86\u524d\u6240\u672a\u6709\u7684\u4f18\u5316\u65b9\u6848\u3002\u8fd9\u4e00\u7ec4\u5408\u4e0d\u4ec5\u91cd\u65b0\u5b9a\u4e49\u4e86\u5206\u5e03\u5f0f\u8ba1\u7b97\u7684\u8fb9\u754c\uff0c\u66f4\u8ba9\u4f01\u4e1a\u7ea7 AI \u5e94\u7528\u7684\u90e8\u7f72\u6210\u672c\u5927\u5e45\u964d\u4f4e\u3002\u6b22\u8fce\u8bbf\u95ee <a href=\"https:\/\/cloud.google.com\/tpu\" target=\"_blank\">\u5b98\u65b9\u7f51\u7ad9<\/a> \u83b7\u53d6\u66f4\u591a\u6280\u672f\u7ec6\u8282\u3002<\/p>\n<h2>\u6838\u5fc3\u529f\u80fd\uff1a\u52a8\u6001\u8def\u7531\u4e0e\u786c\u4ef6\u534f\u540c<\/h2>\n<p>Pathways \u662f\u4e00\u79cd\u65b0\u578b\u7684\u673a\u5668\u5b66\u4e60\u7cfb\u7edf\u67b6\u6784\uff0c\u5b83\u6253\u7834\u4e86\u4f20\u7edf\u6a21\u578b\u5fc5\u987b\u8fd0\u884c\u5728\u5355\u4e00\u8bbe\u5907\u4e0a\u7684\u9650\u5236\u3002\u901a\u8fc7\u5c06 MoE \u5c42\u4e2d\u7684\u201c\u4e13\u5bb6\u201d\u5b50\u7f51\u7edc\u52a8\u6001\u5206\u914d\u5230\u4e0d\u540c TPU v5e \u82af\u7247\u4e0a\uff0cPathways \u5b9e\u73b0\u4e86\u8fd1\u4e4e\u7ebf\u6027\u7684\u52a0\u901f\u6bd4\u3002\u5176\u6838\u5fc3\u529f\u80fd\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>\u81ea\u52a8\u4e13\u5bb6\u8def\u7531<\/strong>\uff1a\u6839\u636e\u8f93\u5165\u6837\u672c\u7684\u7a00\u758f\u6fc0\u6d3b\u7279\u6027\uff0c\u5b9e\u65f6\u5c06\u8ba1\u7b97\u4efb\u52a1\u5206\u914d\u7ed9\u6700\u5408\u9002\u7684\u4e13\u5bb6\u8282\u70b9\u3002<\/li>\n<li><strong>\u8de8\u82af\u7247\u96f6\u62f7\u8d1d\u901a\u4fe1<\/strong>\uff1aTPU v5e \u7684 ICI\uff08Inter-Chip Interconnect\uff09\u5e26\u5bbd\u9ad8\u8fbe 1600 Gbps\uff0c\u6781\u5927\u964d\u4f4e\u4e86\u4e13\u5bb6\u95f4\u7684\u6570\u636e\u4f20\u8f93\u5ef6\u8fdf\u3002<\/li>\n<li><strong>\u5f39\u6027\u8d44\u6e90\u6c60<\/strong>\uff1a\u652f\u6301\u5728\u6570\u767e\u4e2a TPU v5e Pod \u4e0a\u65e0\u7f1d\u6269\u5c55\uff0c\u663e\u5b58\u5bb9\u91cf\u6700\u9ad8\u53ef\u8fbe\u6570\u767e GB\u3002<\/li>\n<\/ul>\n<h3>TPU v5e \u7684\u786c\u4ef6\u4f18\u52bf<\/h3>\n<p>TPU v5e \u4e13\u4e3a\u9ad8\u6548\u63a8\u7406\u800c\u751f\uff0c\u91c7\u7528 7nm \u5236\u7a0b\uff0c\u6bcf\u9897\u82af\u7247\u62e5\u6709 8 \u4e2a Tensor Core\uff0cBF16 \u7b97\u529b\u8fbe 393 TFLOPS\u3002\u5176\u72ec\u7279\u7684\u4f4e\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u5355\u5143\u4e0e Pathways \u7684\u7a00\u758f\u8ba1\u7b97\u6a21\u5f0f\u9ad8\u5ea6\u5951\u5408\uff0c\u76f8\u6bd4\u4e0a\u4e00\u4ee3 TPU v4\uff0cMoE \u63a8\u7406\u541e\u5410\u91cf\u63d0\u5347 2.5 \u500d\uff0c\u529f\u8017\u5374\u4e0b\u964d 30%\u3002<\/p>\n<h2>\u5e94\u7528\u573a\u666f\uff1a\u4ece\u8bed\u8a00\u6a21\u578b\u5230\u591a\u6a21\u6001\u63a8\u8350<\/h2>\n<p>Pathways on TPU v5e \u5df2\u5728 Google \u5185\u90e8\u591a\u9879\u4e1a\u52a1\u4e2d\u843d\u5730\u3002\u6700\u5178\u578b\u7684\u573a\u666f\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09<\/strong>\uff1a\u5982 PaLM 2 \u7684 MoE \u7248\u672c\uff0c\u5229\u7528\u6b64\u67b6\u6784\u5b9e\u73b0\u5343\u4ebf\u53c2\u6570\u6a21\u578b\u7684\u6beb\u79d2\u7ea7\u54cd\u5e94\u3002<\/li>\n<li><strong>\u5e7f\u544a\u63a8\u8350\u7cfb\u7edf<\/strong>\uff1a\u901a\u8fc7\u4e13\u5bb6\u4e13\u4e1a\u5316\u5206\u5de5\uff0c\u5c06 CTR \u9884\u6d4b\u6a21\u578b\u7684\u5728\u7ebf\u5ef6\u8fdf\u964d\u4f4e\u81f3 5ms \u4ee5\u5185\u3002<\/li>\n<li><strong>\u79d1\u5b66\u8ba1\u7b97<\/strong>\uff1a\u5728\u836f\u7269\u5206\u5b50\u6a21\u62df\u4e2d\uff0c\u52a8\u6001\u4e13\u5bb6\u8def\u7531\u53ef\u9488\u5bf9\u4e0d\u540c\u5206\u5b50\u6784\u578b\u81ea\u52a8\u8c03\u7528\u6700\u4f18\u8ba1\u7b97\u6a21\u5757\u3002<\/li>\n<\/ul>\n<h3>\u5b9e\u9645\u90e8\u7f72\u6848\u4f8b<\/h3>\n<p>\u636e Google Cloud \u516c\u5f00\u8d44\u6599\uff0c\u67d0\u5927\u578b\u7535\u5546\u5e73\u53f0\u4f7f\u7528 Pathways + TPU v5e \u5c06\u5176\u63a8\u8350\u6a21\u578b\u53c2\u6570\u91cf\u4ece 1B \u6269\u5c55\u81f3 10B\uff0cA\/B \u6d4b\u8bd5\u4e2d\u7528\u6237\u70b9\u51fb\u7387\u63d0\u5347 18%\uff0c\u800c\u63a8\u7406\u6210\u672c\u4ec5\u589e\u52a0 40%\u3002\u8fd9\u5f97\u76ca\u4e8e MoE \u7a00\u758f\u6fc0\u6d3b\u7279\u6027\u2014\u2014\u6bcf\u6b21\u63a8\u7406\u4ec5\u6fc0\u6d3b\u603b\u53c2\u6570\u7684 10%~20%\u3002<\/p>\n<h2>\u5982\u4f55\u4f7f\u7528\uff1a\u5feb\u901f\u4e0a\u624b\u6307\u5357<\/h2>\n<p>\u5f00\u53d1\u8005\u53ef\u901a\u8fc7 Google Cloud \u7684 Vertex AI \u5e73\u53f0\u76f4\u63a5\u542f\u7528 Pathways \u8fd0\u884c\u65f6\u3002\u5177\u4f53\u6b65\u9aa4\uff1a<\/p>\n<ul>\n<li><strong>\u521b\u5efa TPU v5e \u5207\u7247<\/strong>\uff1a\u5728 GCP \u63a7\u5236\u53f0\u9009\u62e9 TPU v5e Pod\uff0c\u5e76\u914d\u7f6e\u8282\u70b9\u6570\u91cf\uff08\u63a8\u8350 64 \u82af\u7247\u8d77\u6b65\uff09\u3002<\/li>\n<li><strong>\u7f16\u5199 MoE \u6a21\u578b\u4ee3\u7801<\/strong>\uff1a\u4f7f\u7528 JAX \u6216 TensorFlow \u7684 Flax \u5e93\u5b9a\u4e49\u4e13\u5bb6\u5c42\uff0c\u81ea\u52a8\u9002\u914d Pathways \u8c03\u5ea6\u5668\u3002<\/li>\n<li><strong>\u90e8\u7f72\u63a8\u7406\u7aef\u70b9<\/strong>\uff1a\u5c06\u6a21\u578b\u5bfc\u51fa\u4e3a SavedModel\uff0c\u5229\u7528 Cloud Run \u6216 Kubernetes \u5b9e\u73b0\u5f39\u6027\u6269\u7f29\u5bb9\u3002<\/li>\n<\/ul>\n<p>\u6b64\u5916\uff0cGoogle \u5f00\u6e90\u4e86 <strong>Pathways-Examples<\/strong> \u4ed3\u5e93\uff08<a href=\"https:\/\/github.com\/google\/pathways-examples\" target=\"_blank\">GitHub \u793a\u4f8b<\/a>\uff09\uff0c\u63d0\u4f9b\u5b8c\u6574\u7684 MoE \u8bad\u7ec3\u4e0e\u63a8\u7406\u53c2\u8003\u4ee3\u7801\uff0c\u964d\u4f4e\u8fc1\u79fb\u6210\u672c\u3002<\/p>\n<h3>\u6027\u80fd\u5bf9\u6bd4\u6570\u636e<\/h3>\n<p>\u5728\u6807\u51c6 MoE-Bench \u6d4b\u8bd5\u4e2d\uff0c64 \u82af\u7247 TPU v5e \u96c6\u7fa4\u8fd0\u884c 1.2T \u53c2\u6570 MoE \u6a21\u578b\uff0c\u5355\u6b21\u63a8\u7406\u5ef6\u8fdf\u4e3a 22ms\uff0c\u800c\u540c\u7b49\u89c4\u6a21\u7684 NVIDIA H100 \u96c6\u7fa4\u9700\u8981 38ms\uff0c\u540c\u65f6 TPU \u7684\u6bcf TOPS \u6210\u672c\u964d\u4f4e 45%\u3002<\/p>\n<p>\u603b\u7684\u6765\u8bf4\uff0cGoogle Pathways on TPU v5e \u4e3a\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\u63d0\u4f9b\u4e86\u4e00\u6761\u517c\u987e\u6027\u80fd\u4e0e\u6210\u672c\u7684\u52a1\u5b9e\u8def\u5f84\u3002\u65e0\u8bba\u662f\u4e91\u7aef AI \u670d\u52a1\u8fd8\u662f\u4f01\u4e1a\u79c1\u6709\u5316\u90e8\u7f72\uff0c\u8fd9\u4e00\u6280\u672f\u90fd\u6b63\u5728\u52a0\u901f\u63a8\u52a8 AI \u4ece\u201c\u80fd\u8dd1\u201d\u5230\u201c\u8dd1\u5f97\u597d\u201d\u7684\u8de8\u8d8a\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\uff0c\u5927\u89c4\u6a21\u7a00\u758f\u6a21\u578b\u7684\u8bad\u7ec3\u4e0e\u63a8\u7406\u59cb\u7ec8\u9762\u4e34\u7b97\u529b\u4e0e\u6548\u7387\u7684\u53cc\u91cd\u6311\u6218\u3002Google \u6700\u65b0\u63a8\u51fa\u7684 Pathwa [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[10201,10407,10398,10406,10399],"class_list":["post-8364","post","type-post","status-publish","format-standard","hentry","category-4","tag-ai","tag-google-cloud","tag-google-pathways","tag-mixture-of-experts","tag-tpu-v5e"],"_links":{"self":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/8364","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=8364"}],"version-history":[{"count":1,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/8364\/revisions"}],"predecessor-version":[{"id":8365,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/8364\/revisions\/8365"}],"wp:attachment":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=8364"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=8364"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=8364"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}