{"id":868,"date":"2026-01-09T14:52:10","date_gmt":"2026-01-09T06:52:10","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=868"},"modified":"2026-01-12T10:20:32","modified_gmt":"2026-01-12T02:20:32","slug":"communication-characterization-of-ai-workloads-for-large-scale-multi-chiplet-accelerators","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=868","title":{"rendered":"Communication Characterization of AI Workloads for Large-scale Multi-chiplet Accelerators"},"content":{"rendered":"\n

\u8bba\u6587\u6765\u6e90\uff1aarXiv:2410.22262v2<\/a>
\u4f5c\u8005\uff1aMariam Musavi, Emmanuel Irabor, Abhijit Das, Eduard Alarc\u00f3n, Sergi Abadal
\u5355\u4f4d\uff1aNaNoNetworking Center in Catalunya (N3Cat), Universitat Polit\u00e8cnica de Catalunya (UPC)<\/p>\n\n\n\n

\u4e00\u3001\u7814\u7a76\u80cc\u666f\u4e0e\u6311\u6218<\/strong>
\u968f\u7740\u4eba\u5de5\u667a\u80fd\uff08AI\uff09\u6a21\u578b\u89c4\u6a21\u7684\u4e0d\u65ad\u589e\u957f\uff0c\u5355\u82af\u7247\u52a0\u901f\u5668\u5df2\u96be\u4ee5\u6ee1\u8db3\u7b97\u529b\u9700\u6c42\u3002Chiplet\uff08\u82af\u7c92\uff09<\/strong>\u6280\u672f\u5e94\u8fd0\u800c\u751f\uff0c\u901a\u8fc7\u5c06\u591a\u4e2a\u4e13\u7528\u52a0\u901f\u5668\u82af\u7c92\u5c01\u88c5\u5728\u4e00\u8d77\uff0c\u6784\u5efa\u6a2a\u5411\u6269\u5c55\uff08Scale-out\uff09\u7684\u591a\u82af\u7c92\u67b6\u6784\uff0c\u6210\u4e3a\u4e0b\u4e00\u4ee3 AI \u52a0\u901f\u5668\u7684\u4e3b\u6d41\u65b9\u6848\u3002
\u7136\u800c\uff0c\u591a\u82af\u7c92\u67b6\u6784\u9762\u4e34\u4e00\u4e2a\u5173\u952e\u6311\u6218\u2014\u2014\u901a\u4fe1\u74f6\u9888<\/strong>\uff1a
– \u80fd\u8017\u96c6\u4e2d\u4e8e\u6570\u636e\u642c\u8fd0<\/strong>\uff1a\u7814\u7a76\u8868\u660e\uff0cAI \u52a0\u901f\u5668\u82af\u7c92\u53ef\u80fd\u5c06\u8d85\u8fc7 90%<\/strong>\u7684\u7cfb\u7edf\u80fd\u91cf\u7528\u4e8e\u6570\u636e\u642c\u8fd0\u4efb\u52a1\u3002
– \u591a\u64ad\u901a\u4fe1\u9700\u6c42\u6fc0\u589e<\/strong>\uff1a\u73b0\u4ee3 AI \u6570\u636e\u6d41\uff08\u5982\u6743\u91cd\u5e7f\u64ad\u3001\u6fc0\u6d3b\u503c\u5206\u53d1\uff09\u4e2d\u5e7f\u6cdb\u5b58\u5728 \u591a\u64ad\uff08Multicast\uff09<\/strong> \u901a\u4fe1\u6a21\u5f0f\u3002
– \u82af\u7c92\u95f4\u4e92\u8fde\u901f\u5ea6\u53d7\u9650<\/strong>\uff1a\u76f8\u6bd4\u7247\u4e0a\u7f51\u7edc\uff08NoC\uff09\uff0c\u5c01\u88c5\u7ea7\u7f51\u7edc\uff08NoP\uff09\u7684\u4f20\u8f93\u5ef6\u8fdf\u66f4\u9ad8\uff0c\u4e25\u91cd\u5236\u7ea6\u7cfb\u7edf\u6269\u5c55\u6027\u3002
\u5c3d\u7ba1\u5df2\u6709 SIMBA\u3001GEMINI \u7b49\u5de5\u4f5c\u5173\u6ce8\u591a\u82af\u7c92\u67b6\u6784\u7684\u6620\u5c04\u4e0e\u8c03\u5ea6\u4f18\u5316\uff0c\u4f46\u73b0\u6709\u7814\u7a76\u7f3a\u4e4f\u5bf9\u901a\u4fe1\u6d41\u91cf\u7279\u5f81\uff08\u5c24\u5176\u662f\u591a\u64ad\u6d41\u91cf\uff09\u7684\u6df1\u5165\u91cf\u5316\u5206\u6790<\/strong>\u3002
\u672c\u8bba\u6587\u65e8\u5728\u586b\u8865\u8fd9\u4e00\u7a7a\u767d\uff0c\u7cfb\u7edf\u6027\u5730\u5256\u6790 AI \u8d1f\u8f7d\u5728\u591a\u82af\u7c92\u52a0\u901f\u5668\u4e0a\u7684\u6570\u636e\u79fb\u52a8\u7279\u5f81\uff0c\u63ed\u793a\u901a\u4fe1\u74f6\u9888\u7684\u6839\u6e90\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u6838\u5fc3\u65b9\u6cd5\uff1a\u57fa\u4e8e GEMINI \u7684\u901a\u4fe1\u7279\u5f81\u5206\u6790\u6846\u67b6<\/strong>
\u4e3a\u4e86\u91cf\u5316\u8bc4\u4f30 AI \u8d1f\u8f7d\u7684\u901a\u4fe1\u884c\u4e3a\uff0c\u4f5c\u8005\u57fa\u4e8e GEMINI<\/strong>\u6846\u67b6\u6784\u5efa\u4e86\u4e00\u5957\u589e\u5f3a\u578b\u5206\u6790\u65b9\u6cd5\u3002
2.1 \u5206\u6790\u6d41\u7a0b<\/strong>
\u6574\u4f53\u6d41\u7a0b\u5206\u4e3a\u56db\u4e2a\u6b65\u9aa4\uff1a
1. \u8f93\u5165\u914d\u7f6e<\/strong>\uff1a\u9009\u62e9\u76ee\u6807 AI \u8d1f\u8f7d\uff08\u5982 ResNet\u3001Transformer\uff09\u4e0e\u786c\u4ef6\u67b6\u6784\u53c2\u6570\uff08\u82af\u7c92\u6570\u91cf\u3001\u4e92\u8fde\u5e26\u5bbd\u7b49\uff09\u3002
2. \u6700\u4f18\u6620\u5c04<\/strong>\uff1a\u5229\u7528 GEMINI \u7684\u8bbe\u8ba1\u7a7a\u95f4\u63a2\u7d22\u5f15\u64ce\uff0c\u786e\u5b9a\u80fd\u8017\u5ef6\u8fdf\u79ef\uff08EDP\uff09\u6700\u4f18\u7684\u6620\u5c04\u65b9\u6848\u3002
3. \u901a\u4fe1\u8ffd\u8e2a<\/strong>\uff1a\u901a\u8fc7 C++ \u529f\u80fd\u811a\u672c\u6269\u5c55 GEMINI\uff0c\u8bb0\u5f55\u6240\u6709\u901a\u4fe1\u6570\u636e\u5305\u7684\u8be6\u7ec6\u4fe1\u606f\u3002
4. \u7279\u5f81\u63d0\u53d6<\/strong>\uff1a\u4f7f\u7528 Python \u811a\u672c\u89e3\u6790\u8ffd\u8e2a\u6570\u636e\uff0c\u63d0\u53d6\u5173\u952e\u901a\u4fe1\u6307\u6807\u5e76\u53ef\u89c6\u5316\u3002<\/p>\n\n\n\n

\"Communication<\/figure>\n\n\n\n

2.2 \u5173\u952e\u5ea6\u91cf\u6307\u6807<\/strong>
\u8bba\u6587\u805a\u7126\u4ee5\u4e0b\u56db\u7c7b\u901a\u4fe1\u5ea6\u91cf\uff1a<\/p>\n\n\n\n

\"Communication<\/figure>\n\n\n\n

\u4e09\u3001\u5b9e\u9a8c\u914d\u7f6e<\/strong>
\u4f5c\u8005\u9009\u53d6\u4e86 12 \u79cd\u4e3b\u6d41 AI \u63a8\u7406\u8d1f\u8f7d<\/strong> \u8fdb\u884c\u6d4b\u8bd5\uff0c\u6db5\u76d6\u56fe\u50cf\u5206\u7c7b\u4e0e\u81ea\u7136\u8bed\u8a00\u5904\u7406\u9886\u57df\uff1a
– \u6b8b\u5dee\u7f51\u7edc<\/strong>\uff1aResNet50\u3001ResNet101\u3001ResNet152\u3001ResNext50
– \u5bc6\u96c6\u7f51\u7edc<\/strong>\uff1aDenseNet\u3001Darknet19
– Inception \u7cfb\u5217<\/strong>\uff1aGoogleNet\u3001iRES
– Transformer \u7cfb\u5217<\/strong>\uff1aTF\u3001TF Cell
– \u5e8f\u5217\u6a21\u578b<\/strong>\uff1aLSTM\u3001GNMT
\u6d4b\u8bd5\u5728\u4e09\u79cd\u82af\u7c92\u9635\u5217\u914d\u7f6e\u4e0b\u8fdb\u884c\uff1a1\u00d72\uff082 \u82af\u7c92\uff09<\/strong>\u30013\u00d73\uff089 \u82af\u7c92\uff09<\/strong>\u30016\u00d73\uff0818 \u82af\u7c92\uff09<\/strong>\uff0c\u5747\u914d\u5907 4 \u4e2a DRAM \u82af\u7c92\u3002<\/p>\n\n\n\n

\u56db\u3001\u5b9e\u9a8c\u7ed3\u679c\u4e0e\u6838\u5fc3\u53d1\u73b0<\/strong>
4.1 \u901a\u4fe1\u65f6\u95f4\u5360\u6bd4\u5206\u6790<\/strong>
\u5b9e\u9a8c\u63ed\u793a\u4e86\u4e00\u4e2a\u5173\u952e\u4e8b\u5b9e\uff1a\u968f\u7740\u82af\u7c92\u6570\u91cf\u589e\u52a0\uff0c\u82af\u7c92\u95f4\u901a\u4fe1\u65f6\u95f4\u5360\u6bd4\u6025\u5267\u4e0a\u5347<\/strong>\u3002<\/p>\n\n\n\n

\"Communication<\/figure>\n\n\n\n

\u8fd9\u610f\u5473\u7740\uff1a
– \u7b80\u5355\u589e\u52a0\u82af\u7c92\u6570\u91cf\u65e0\u6cd5\u5e26\u6765\u7ebf\u6027\u6027\u80fd\u63d0\u5347<\/strong>
– “\u901a\u4fe1\u5899”\uff08Communication Wall\uff09<\/strong> \u6548\u5e94\u5728\u5927\u89c4\u6a21\u914d\u7f6e\u4e0b\u5c24\u4e3a\u663e\u8457
– Transformer \u7c7b\u6a21\u578b\u53d7\u901a\u4fe1\u74f6\u9888\u5f71\u54cd\u6700\u4e3a\u4e25\u91cd
4.2 \u591a\u64ad\u6d41\u91cf\u5206\u6790<\/strong>
\u591a\u64ad\u901a\u4fe1\u662f\u5bfc\u81f4\u901a\u4fe1\u74f6\u9888\u7684\u5173\u952e\u56e0\u7d20\uff1a
– DenseNet<\/strong> \u5728 6\u00d73 \u914d\u7f6e\u4e0b\u4ea7\u751f\u9ad8\u8fbe 4120 \u4e07\u6761<\/strong> \u591a\u64ad\u6d88\u606f
– Transformer<\/strong> \u4ea7\u751f\u7ea6 1360 \u4e07\u6761<\/strong> \u591a\u64ad\u6d88\u606f
– \u8bb8\u591a\u591a\u64ad\u6d88\u606f\u9700\u8981\u540c\u65f6\u53d1\u9001\u5230 **6 \u4e2a\u4ee5\u4e0a**<\/strong> \u76ee\u6807\u82af\u7c92\uff0c\u9020\u6210\u4e25\u91cd\u7684\u7f51\u7edc\u62e5\u585e
4.3 \u7a7a\u95f4\u5c40\u90e8\u6027\u5206\u6790<\/strong>
\u901a\u8fc7\u5206\u6790 NoP \u8df3\u6570\u5206\u5e03\uff0c\u4f5c\u8005\u53d1\u73b0\uff1a
– \u4f20\u7edf\u5c40\u90e8\u6027\u5047\u8bbe\u5931\u6548<\/strong>\uff1a\u4e0e CPU\/GPU \u5de5\u4f5c\u8d1f\u8f7d\u4e0d\u540c\uff0c\u591a\u82af\u7c92 AI \u52a0\u901f\u5668\u4e2d\u957f\u8ddd\u79bb\u901a\u4fe1\u975e\u5e38\u666e\u904d<\/strong>
– \u591a\u64ad\u503e\u5411\u4e8e\u957f\u8ddd\u79bb<\/strong>\uff1a\u5c3d\u7ba1\u591a\u64ad\u6d88\u606f\u603b\u6570\u5c11\u4e8e\u5355\u64ad\uff0c\u4f46\u5176\u5e73\u5747\u8df3\u6570\u66f4\u9ad8
– \u968f\u7740\u7cfb\u7edf\u89c4\u6a21\u6269\u5927\uff0c\u9ad8\u8df3\u6570\u6d88\u606f\u6570\u91cf\u5448\u8d85\u7ebf\u6027\u589e\u957f<\/strong><\/p>\n\n\n\n

\"Communication<\/figure>\n\n\n\n
\"Communication<\/figure>\n\n\n\n

\u4e94\u3001\u6838\u5fc3\u8d21\u732e\u4e0e\u610f\u4e49<\/strong>
1. \u9996\u6b21\u7cfb\u7edf\u6027\u523b\u753b\u591a\u82af\u7c92 AI \u52a0\u901f\u5668\u7684\u901a\u4fe1\u7279\u5f81<\/strong>\uff1a\u63d0\u4f9b\u4e86\u5355\u64ad\/\u591a\u64ad\u6d41\u91cf\u3001\u8df3\u6570\u5206\u5e03\u548c\u901a\u4fe1\u65f6\u95f4\u5360\u6bd4\u7684\u8be6\u7ec6\u91cf\u5316\u5206\u6790\u3002
2. \u63ed\u793a\u591a\u64ad\u901a\u4fe1\u7684\u4e3b\u5bfc\u5730\u4f4d<\/strong>\uff1a\u91cf\u5316\u8bc1\u660e\u4e86\u591a\u64ad\u6d41\u91cf\u5728\u7279\u5b9a AI \u8d1f\u8f7d\u4e2d\u7684\u5173\u952e\u4f5c\u7528\u53ca\u5176\u5bf9\u6269\u5c55\u6027\u7684\u5236\u7ea6\u3002
3. \u6311\u6218\u4f20\u7edf\u5c40\u90e8\u6027\u5047\u8bbe<\/strong>\uff1a\u8bc1\u660e\u4e86 AI \u8d1f\u8f7d\u901a\u4fe1\u4e2d\u5e7f\u6cdb\u5b58\u5728\u7684\u957f\u8ddd\u79bb\u4f9d\u8d56\uff0c\u8fd9\u5bf9\u4e92\u8fde\u67b6\u6784\u8bbe\u8ba1\u5177\u6709\u91cd\u8981\u6307\u5bfc\u610f\u4e49\u3002<\/p>\n\n\n\n

\u516d\u3001\u672a\u6765\u5c55\u671b<\/strong>
\u57fa\u4e8e\u4e0a\u8ff0\u53d1\u73b0\uff0c\u8bba\u6587\u63d0\u51fa\u591a\u82af\u7c92 AI \u52a0\u901f\u5668\u7684\u4e92\u8fde\u67b6\u6784\u8bbe\u8ba1\u5e94\u91cd\u70b9\u8003\u8651\uff1a
– \u65b0\u578b\u4e92\u8fde\u6280\u672f<\/strong>\uff1a
– \u65e0\u7ebf\u4e92\u8fde\uff08Wireless Interconnects\uff09<\/strong>\uff1a\u5929\u7136\u652f\u6301\u5e7f\u64ad\uff0c\u9002\u5408\u591a\u64ad\u5bc6\u96c6\u578b\u8d1f\u8f7d
– \u5149\u4e92\u8fde\uff08Optical Interconnects\uff09<\/strong>\uff1a\u4f4e\u5ef6\u8fdf\u957f\u8ddd\u79bb\u4f20\u8f93\uff0c\u7f13\u89e3\u8df3\u6570\u74f6\u9888
– \u7075\u6d3b\u7684\u4e92\u8fde\u67b6\u6784<\/strong>\uff1a\u8bbe\u8ba1\u80fd\u591f\u52a8\u6001\u9002\u5e94\u4e0d\u540c AI \u8d1f\u8f7d\u901a\u4fe1\u6a21\u5f0f\u7684 NoP \u67b6\u6784
– \u8fd1\u5b58\u8ba1\u7b97\u4e0e 3D \u5806\u53e0<\/strong>\uff1a\u901a\u8fc7\u8fd1\u5b58\u8ba1\u7b97\uff08NMC\uff09\u6216 3D \u5806\u53e0\u5185\u5b58\u51cf\u5c11\u6570\u636e\u642c\u8fd0\u8ddd\u79bb<\/p>\n\n\n\n

\u603b\u7ed3<\/strong>
\u672c\u6587\u901a\u8fc7\u7cfb\u7edf\u6027\u7684\u901a\u4fe1\u7279\u5f81\u5206\u6790\uff0c\u63ed\u793a\u4e86\u591a\u82af\u7c92 AI \u52a0\u901f\u5668\u4e2d\u901a\u4fe1\u74f6\u9888\u7684\u6839\u6e90\u5728\u4e8e\u591a\u64ad\u6d41\u91cf\u7684\u4e3b\u5bfc\u5730\u4f4d\u548c\u957f\u8ddd\u79bb\u901a\u4fe1\u7684\u666e\u904d\u6027<\/strong>\u3002\u8fd9\u4e9b\u6d1e\u5bdf\u4e3a\u4e0b\u4e00\u4ee3 AI \u52a0\u901f\u5668\u7684\u4e92\u8fde\u67b6\u6784\u8bbe\u8ba1\u63d0\u4f9b\u4e86\u91cd\u8981\u7684\u7406\u8bba\u4f9d\u636e\u548c\u5b9e\u8df5\u6307\u5bfc\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u8bba\u6587\u6765\u6e90\uff1aarXiv:2410.22262v2\u4f5c\u8005\uff1aMariam Musavi, Emmanuel Irabor, Abhijit Das, Eduard Alarc\u00f3n, Sergi Abadal\u5355\u4f4d\uff1aNaNoNetworking Center in Catalunya (N3Cat), Universitat Polit\u00e8cnica de Catalunya (UPC) \u4e00\u3001\u7814\u7a76\u80cc\u666f\u4e0e\u6311\u6218 \u968f\u7740\u4eba\u5de5\u667a\u80fd\uff08AI\uff09\u6a21\u578b\u89c4\u6a21\u7684\u4e0d\u65ad\u589e\u957f\uff0c\u5355\u82af\u7247\u52a0\u901f\u5668\u5df2\u96be\u4ee5\u6ee1\u8db3\u7b97\u529b\u9700\u6c42\u3002Chipl … \u7ee7\u7eed\u9605\u8bfb →<\/span><\/a><\/p>\n","protected":false},"author":6,"featured_media":876,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-868","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-rengongzhineng"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/868","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/6"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=868"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/868\/revisions"}],"predecessor-version":[{"id":877,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/868\/revisions\/877"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/876"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=868"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=868"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=868"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}