{"id":1161,"date":"2026-03-27T13:17:28","date_gmt":"2026-03-27T05:17:28","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1161"},"modified":"2026-03-27T13:17:42","modified_gmt":"2026-03-27T05:17:42","slug":"pegasus-a-universal-framework-for-scalable-deep-learning-inference","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1161","title":{"rendered":"Pegasus: A Universal Framework for Scalable Deep Learning Inference"},"content":{"rendered":"\n<p>SIGCOMM &#8217;25: ACM SIGCOMM 2025<\/p>\n\n\n\n<p>Conference<\/p>\n\n\n\n<p><em>September 8 &#8211; 11, 2025<\/em><\/p>\n\n\n\n<p><em>Coimbra, Portuga<\/em><\/p>\n\n\n\n<p><strong>1. \u6458\u8981\uff08Abstract\uff09<\/strong><\/p>\n\n\n\n<p>Pegasus\u9488\u5bf9\u5f53\u524dIntelligent DataPlane\uff08IDP\uff09\u4e2d\u6df1\u5ea6\u5b66\u4e60\uff08DL\uff09\u63a8\u7406\u9762\u4e34\u7684\u4e09\u5927\u6838\u5fc3\u9650\u5236\u2014\u2014\u7cbe\u5ea6\u4e0b\u964d\u3001\u53ef\u6269\u5c55\u6027\u5dee\u3001\u6cdb\u5316\u80fd\u529b\u4e0d\u8db3\u2014\u2014\u63d0\u51fa\u4e86\u4e00\u4e2a\u901a\u7528\u6846\u67b6\u3002\u5b83\u5c06DL\u64cd\u4f5c\u62bd\u8c61\u4e3a\u4e09\u79cd\u9762\u5411\u6570\u636e\u5e73\u9762\u7684\u539f\u8bed\uff1aPartition\uff08\u5206\u533a\uff09\u3001Map\uff08\u6620\u5c04\uff09\u548cSumReduce\uff08\u6c42\u548c\u5f52\u7ea6\uff09\u3002\u901a\u8fc7\u8fd9\u4e9b\u539f\u8bed\uff0cPegasus\u5b9e\u73b0\u4e86\u5bf9MLP\u3001RNN\u3001CNN\u3001AutoEncoder\u7b49\u591a\u79cd\u6a21\u578b\u7684\u76f4\u63a5\u652f\u6301\uff0c\u65e0\u9700\u4fee\u6539\u786c\u4ef6\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u5f3a\u8c03\uff0cPegasus\u91c7\u7528\u5168\u7cbe\u5ea6\u6743\u91cd+\u5b9a\u70b9\u6fc0\u6d3b\u503c\uff0c\u5e76\u5f15\u5165Fuzzy Matching\uff08\u6a21\u7cca\u5339\u914d\uff09\u548cPrimitive Fusion\uff08\u539f\u8bed\u878d\u5408\uff09\u6280\u672f\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u7cbe\u5ea6\u4e0e\u8d44\u6e90\u5229\u7528\u7387\u3002\u5728P4\u53ef\u7f16\u7a0b\u4ea4\u6362\u673a\u4e0a\u7684\u5b9e\u9a8c\u663e\u793a\uff0cPegasus\u5728\u7cbe\u5ea6\u4e0a\u5e73\u5747\u63d0\u5347\u6700\u9ad822.8%\uff0c\u6a21\u578b\u89c4\u6a21\u6269\u5927248\u500d\uff0c\u8f93\u5165\u89c4\u6a21\u6269\u5927212\u500d\uff0c\u5b9e\u73b0\u4e86\u771f\u6b63\u7ebf\u901f\u7684\u901a\u7528DL\u63a8\u7406\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"153\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/1-3-1024x153.png\"  class=\"wp-image-1162\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/1-3-1024x153.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/1-3-300x45.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/1-3-768x114.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/1-3.png 1269w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe\" \/><\/figure>\n\n\n\n<p><strong>2. \u5f15\u8a00\uff08Introduction\uff09<\/strong><\/p>\n\n\n\n<p>\u4f20\u7edfIDP\u4e3b\u8981\u4f9d\u8d56\u5339\u914d-\u52a8\u4f5c\u8868\uff08MAT\uff09\u62bd\u8c61\uff0c\u4f46MAT\u4e0eDL\u63a8\u7406\u7684\u8ba1\u7b97\u9700\u6c42\u4e25\u91cd\u4e0d\u5339\u914d\u3002\u73b0\u6709\u65b9\u6848\u5206\u4e3a\u4e24\u7c7b\uff1a\u8ba1\u7b97\u7b80\u5316\uff08\u5982N3IC\u7684\u4e8c\u503c\u5316MLP\uff09\u548c\u8ba1\u7b97\u7ed5\u8fc7\uff08\u5982BoS\u7684\u8f93\u5165-\u8f93\u51fa\u6620\u5c04\u8868\uff09\u3002\u4e24\u8005\u5747\u5b58\u5728\u7cbe\u5ea6\u635f\u5931\uff08\u4e8c\u503c\u5316\u5bfc\u81f4\u6570\u503c\u8303\u56f4\u7f29\u5c0f\uff09\u3001\u53ef\u6269\u5c55\u6027\u53d7\u9650\uff08\u8868\u9879\u7206\u70b8\u6216\u9636\u6bb5\u4e0d\u8db3\uff09\u548c\u6cdb\u5316\u80fd\u529b\u5dee\uff08\u4ec5\u652f\u6301\u7279\u5b9a\u5c42\u6216\u7279\u5b9a\u6a21\u578b\uff09\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p>Pegasus\u63d0\u51fa\u4e94\u9879\u7d27\u5bc6\u8026\u5408\u7684\u8bbe\u8ba1\uff1a\u2460\u5c06DL\u64cd\u4f5c\u8f6c\u5316\u4e3a\u4e09\u79cd\u539f\u8bed\uff1b\u2461\u91c7\u7528Partition+Map+SumReduce\u6d41\u7a0b\uff1b\u2462\u4f7f\u7528\u5168\u7cbe\u5ea6\u6743\u91cd\u4e0e\u5b9a\u70b9\u6fc0\u6d3b\uff1b\u2463\u5f15\u5165Fuzzy Matching\u51cf\u5c11\u67e5\u627e\u6b21\u6570\uff1b\u2464\u91c7\u7528Primitive Fusion\u5408\u5e76\u64cd\u4f5c\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u8be5\u6846\u67b6\u5728Tofino2\u4ea4\u6362\u673a\u4e0a\u6210\u529f\u90e8\u7f72\u591a\u79cd\u6a21\u578b\uff0c\u663e\u8457\u8d85\u8d8aN3IC\u3001BoS\u548cLeo\u7b49\u73b0\u6709\u5de5\u4f5c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"708\" height=\"786\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/3-2.png\"  class=\"wp-image-1163\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/3-2.png 708w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/3-2-270x300.png 270w\" sizes=\"auto, (max-width: 708px) 100vw, 708px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe1\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe1\" \/><\/figure>\n\n\n\n<p><strong>3. \u80cc\u666f\u4e0e\u52a8\u673a\uff08Background and Motivation\uff09<\/strong><\/p>\n\n\n\n<p>\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7531FC\u3001Conv\u3001Act\u3001Norm\u3001Pool\u3001Rec\u3001Emb\u7b49\u591a\u79cd\u5c42\u7ec4\u6210\uff0c\u6bcf\u5c42\u5305\u542bMatMul\u3001\u5143\u7d20\u7ea7\u53d8\u6362\u3001\u591a\u8f93\u5165\u64cd\u4f5c\u7b49\u590d\u6742\u8ba1\u7b97\u3002\u800cPISA\u67b6\u6784\u4ec5\u652f\u6301\u6574\u6570\u52a0\u51cf\u3001\u79fb\u4f4d\u3001\u4f4d\u8fd0\u7b97\uff0c\u4e0d\u652f\u6301\u4e58\u6cd5\u3001\u6d6e\u70b9\u6570\u548c\u6307\u6570\u8fd0\u7b97\uff0c\u8d44\u6e90\u4e5f\u6781\u4e3a\u6709\u9650\uff0820\u4e2aMAT\u9636\u6bb5\u3001\u6709\u9650SRAM\/TCAM\uff09\u3002<\/p>\n\n\n\n<p>\u73b0\u6709\u5de5\u4f5cN3IC\u901a\u8fc7XNOR+popcnt\u5b9e\u73b0\u4e8c\u503cMLP\uff0c\u4f46\u7cbe\u5ea6\u663e\u8457\u4e0b\u964d\u4e14\u65e0\u6cd5\u652f\u6301\u5176\u4ed6\u5c42\uff1bBoS\u901a\u8fc7\u5b8c\u6574\u6620\u5c04\u8868\u7ed5\u8fc7\u8ba1\u7b97\uff0c\u5374\u56e02^n\u8868\u9879\u7206\u70b8\u800c\u8f93\u5165\u89c4\u6a21\u6781\u5c0f\uff0818bit\uff09\u3002\u4e24\u8005\u5747\u65e0\u6cd5\u540c\u65f6\u6ee1\u8db3\u7cbe\u5ea6\u3001\u53ef\u6269\u5c55\u6027\u548c\u6cdb\u5316\u9700\u6c42\u3002Pegasus\u7684\u52a8\u673a\u6b63\u662f\u586b\u8865\u8fd9\u4e00\u7a7a\u767d\uff0c\u5b9e\u73b0\u5546\u54c1\u5316\u53ef\u7f16\u7a0b\u4ea4\u6362\u673a\u4e0a\u7684\u901a\u7528\u3001\u9ad8\u7cbe\u5ea6\u3001\u5927\u89c4\u6a21DL\u63a8\u7406\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"657\" height=\"627\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/5-1.png\"  class=\"wp-image-1164\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/5-1.png 657w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/5-1-300x286.png 300w\" sizes=\"auto, (max-width: 657px) 100vw, 657px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe2\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe2\" \/><\/figure>\n\n\n\n<p><strong>4. \u8bbe\u8ba1\u6982\u8ff0\uff08Design Overview\uff09<\/strong><\/p>\n\n\n\n<p>Pegasus\u7684\u6838\u5fc3\u662f\u4e09\u79cd\u539f\u8bed\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Partition<\/strong>\uff1a\u5c06\u9ad8\u7ef4\u8f93\u5165\u5411\u91cf\u62c6\u5206\u4e3a\u591a\u4e2a\u4f4e\u7ef4\u5b50\u5411\u91cf\uff1b<\/li>\n\n\n\n<li><strong>Map<\/strong>\uff1a\u5bf9\u6bcf\u4e2a\u5b50\u5411\u91cf\u8fdb\u884c\u5e76\u884c\u51fd\u6570\u6620\u5c04\uff08\u501f\u52a9Fuzzy Matching\uff09\uff1b<\/li>\n\n\n\n<li><strong>SumReduce<\/strong>\uff1a\u5bf9\u6620\u5c04\u7ed3\u679c\u8fdb\u884c\u5143\u7d20\u7ea7\u6c42\u548c\u5f52\u7ea6\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u539f\u8bed\u53ef\u4efb\u610f\u7ec4\u5408\uff0c\u8986\u76d6\u7edd\u5927\u591a\u6570DL\u64cd\u4f5c\u3002\u540c\u65f6\u5f15\u5165Primitive Fusion\uff08\u57fa\u672c\u878d\u5408\u4e0e\u9ad8\u7ea7\u878d\u5408\uff09\u548cFuzzy Matching\uff08\u805a\u7c7b\u6811+\u8d28\u5fc3\u8fd1\u4f3c\uff09\uff0c\u5927\u5e45\u964d\u4f4e\u8868\u9879\u6570\u91cf\u548c\u67e5\u627e\u6b21\u6570\u3002\u8bbe\u8ba1\u76ee\u6807\u660e\u786e\uff1a\u66f4\u9ad8\u7cbe\u5ea6\u3001\u66f4\u5927\u89c4\u6a21\u3001\u66f4\u5f3a\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"618\" height=\"705\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/2-3.png\"  class=\"wp-image-1165\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/2-3.png 618w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/2-3-263x300.png 263w\" sizes=\"auto, (max-width: 618px) 100vw, 618px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe3\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe3\" \/><\/figure>\n\n\n\n<p><strong>5. Pegasus\u539f\u8bed\uff08Pegasus Primitives\uff09<\/strong><\/p>\n\n\n\n<p><strong>Fuzzy Matching<\/strong>\uff1a\u6784\u5efa\u805a\u7c7b\u6811\uff08\u8d2a\u5fc3\u5206\u88c2\uff0c\u6700\u5c0f\u5316SSE\uff09\uff0c\u8f93\u5165\u5411\u91cf\u901a\u8fc7\u6bd4\u8f83\u5feb\u901f\u6620\u5c04\u5230\u53f6\u8282\u70b9\u8d28\u5fc3\uff0c\u5b9e\u73b0\u8fd1\u4f3c\u4f46\u9ad8\u6548\u7684\u67e5\u627e\u3002\u76f8\u6bd4\u7a77\u4e3e\u6620\u5c04\uff0c\u5b58\u50a8\u6548\u7387\u3001\u67e5\u627e\u6b21\u6570\u3001\u878d\u5408\u80fd\u529b\u548c\u6d41\u53ef\u6269\u5c55\u6027\u5747\u663e\u8457\u63d0\u5347\uff08\u7cbe\u5ea6\u635f\u5931\u7ea61%\uff09\u3002<\/p>\n\n\n\n<p><strong>Primitive Fusion<\/strong>\uff1a\u57fa\u672c\u878d\u5408\u5229\u7528\u7ebf\u6027\u91cd\u6392\u5e8f\u548c\u8fde\u7eedMap\u5408\u5e76\uff1b\u9ad8\u7ea7\u878d\u5408\u901a\u8fc7\u79fb\u9664\u4e2d\u95f4\u975e\u7ebf\u6027\u6620\u5c04\u6216\u4fdd\u7559\u6700\u7ec8SumReduce\uff0c\u5b9e\u73b0\u5355\u8868\u67e5\u627e\u5b8c\u6210\u591a\u5c42\u8ba1\u7b97\uff08\u7c7b\u4f3cNeural Additive Models\uff09\u3002\u6620\u5c04\u4f18\u5316\u8fd8\u91c7\u7528\u53cd\u5411\u4f20\u64ad\u52a8\u6001\u8c03\u6574\u8d28\u5fc3\uff0c\u8fdb\u4e00\u6b65\u964d\u4f4e\u8fd1\u4f3c\u8bef\u5dee\u3002<\/p>\n\n\n\n<p><strong>6. \u6df1\u5ea6\u5b66\u4e60\u7b97\u5b50\u5b9e\u73b0\uff08Deep Learning Operators\uff09<\/strong><\/p>\n\n\n\n<p>\u8bba\u6587\u8be6\u7ec6\u8bf4\u660e\u5982\u4f55\u7528\u4e09\u79cd\u539f\u8bed\u5b9e\u73b0\u5e38\u89c1DL\u7b97\u5b50\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Embedding Lookup \u2192 Map\uff1b<\/li>\n\n\n\n<li>\u5143\u7d20\u7ea7\u53d8\u6362\uff08BN\u3001ReLU\u3001tanh\uff09\u2192 Map\uff1b<\/li>\n\n\n\n<li>\u52a0\u6743\u805a\u5408\uff08MatMul\u3001Conv\uff09\u2192 Partition+Map+SumReduce\uff1b<\/li>\n\n\n\n<li>\u591a\u8f93\u5165\u64cd\u4f5c\uff08Softmax\u3001Pooling\uff09\u2192 Map+SumReduce+Map\u6216\u8fde\u7eedMap\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6240\u6709\u5b9e\u73b0\u5747\u4e25\u683c\u9075\u5faa\u6570\u636e\u5e73\u9762\u7ea6\u675f\uff0c\u4ec5\u4f9d\u8d56\u6bd4\u8f83\u3001\u67e5\u627e\u548c\u52a0\u6cd5\u3002<\/p>\n\n\n\n<p><strong>7. \u5b9e\u73b0\uff08Implementation\uff09<\/strong><\/p>\n\n\n\n<p>\u5728PISA\uff08Tofino2\uff09\u4e0a\u5b9e\u73b0\u65f6\uff0c\u9488\u5bf9\u5b9a\u70b9\u805a\u5408\u3001Fuzzy Indexing\u548c\u9636\u6bb5\u9650\u5236\u8fdb\u884c\u4e86\u591a\u9879\u5de5\u7a0b\u4f18\u5316\uff1a13-6-13\u5b9a\u70b9\u683c\u5f0f\u9632\u6ea2\u51fa\u3001Consecutive Range Coding\uff08CRC\uff09\u5b9e\u73b0\u8303\u56f4\u5339\u914d\u3001Pegasus Syntax\u9ad8\u5c42\u62bd\u8c61+\u81ea\u52a8\u7ffb\u8bd1\u5de5\u5177\u3002<\/p>\n\n\n\n<p>\u5b9e\u73b0\u4e86\u516d\u79cd\u4ee3\u8868\u6027\u6a21\u578b\uff1aMLP-B\u3001RNN-B\u3001CNN-B\/M\/L\u3001AutoEncoder\u3002\u5176\u4e2dCNN-L\u91c7\u7528Advanced Primitive Fusion\uff0c\u652f\u63013840bit\u8f93\u5165\u548c6083Kb\u6a21\u578b\uff0c\u4ec5\u5360\u752815\u4e2a\u6d41\u6c34\u7ebf\u9636\u6bb5\u3002<\/p>\n\n\n\n<p><strong>8. \u8bc4\u4f30\uff08Evaluation\uff09<\/strong><\/p>\n\n\n\n<p>\u5728PeerRush\u3001CICIOT\u3001ISCXVPN\u4e09\u4e2a\u771f\u5b9e\u6d41\u91cf\u5206\u7c7b\u6570\u636e\u96c6\u4e0a\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7cbe\u5ea6\uff1aCNN-L\u8f83N3IC\u63d0\u534722.8%\u3001\u8f83BoS\u63d0\u534717.9%\u3001\u8f83Leo\u63d0\u534717.2%\uff1b<\/li>\n\n\n\n<li>\u89c4\u6a21\uff1a\u6a21\u578b\u5927\u5c0f248\u00d7\uff0c\u8f93\u5165\u89c4\u6a21212\u00d7\uff1b<\/li>\n\n\n\n<li>\u8d44\u6e90\uff1aCNN-L\u4ec5\u5360\u75287.12% SRAM\u300113.33% TCAM\uff0c\u652f\u6301\u767e\u4e07\u6d41\u7ea7\u522b\u5e76\u53d1\uff1b<\/li>\n\n\n\n<li>\u65e0\u76d1\u7763\u68c0\u6d4b\uff1aAutoEncoder\u5728\u672a\u77e5\u6076\u610f\u6d41\u91cf\u4e0aAUC\u6700\u9ad8\u8fbe99.0%\uff1b<\/li>\n\n\n\n<li>\u4e0eCPU\/GPU\u5bf9\u6bd4\uff1a\u7cbe\u5ea6\u4ec5\u4e0b\u964d\u7ea61.08%\uff0c\u541e\u5410\u91cf\u63d0\u53473800\u00d7\uff08CPU\uff09\u548c600\u00d7\uff08GPU\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"265\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/6-2-1024x265.png\"  class=\"wp-image-1167\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/6-2-1024x265.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/6-2-300x78.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/6-2-768x199.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/6-2.png 1299w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe4\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe4\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"294\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/7-2-1024x294.png\"  class=\"wp-image-1168\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/7-2-1024x294.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/7-2-300x86.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/7-2-768x220.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/7-2.png 1329w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe5\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe5\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"243\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/8-1-1024x243.png\"  class=\"wp-image-1169\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/8-1-1024x243.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/8-1-300x71.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/8-1-768x182.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/8-1.png 1302w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe6\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference\u63d2\u56fe6\" \/><\/figure>\n\n\n\n<p><strong>9. \u8ba8\u8bba\u3001\u76f8\u5173\u5de5\u4f5c\u4e0e\u7ed3\u8bba\uff08Discussion, Related Work &amp; Conclusion\uff09<\/strong><\/p>\n\n\n\n<p>\u8ba8\u8bba\u4e86Fuzzy Matching\u5728\u5206\u5e03\u6f02\u79fb\u573a\u666f\u4e0b\u7684\u9c81\u68d2\u6027\u3001\u4e0e\u73b0\u6709IDP\u5de5\u4f5c\u7684\u4e92\u8865\u6027\uff0c\u4ee5\u53ca\u672a\u6765\u786c\u4ef6\u8f7b\u91cf\u8c03\u6574\u7684\u53ef\u80fd\u6027\u3002\u76f8\u5173\u5de5\u4f5c\u5bf9\u6bd4\u663e\u793a\uff0cPegasus\u662f\u9996\u4e2a\u5728\u5546\u54c1\u5316\u4ea4\u6362\u673a\u4e0a\u5b9e\u73b0\u591a\u6a21\u578b\u3001\u5927\u89c4\u6a21\u3001\u9ad8\u7cbe\u5ea6DL\u63a8\u7406\u7684\u901a\u7528\u6846\u67b6\u3002<\/p>\n\n\n\n<p>\u7ed3\u8bba\u6307\u51fa\uff1a\u9650\u5236IDP\u667a\u80fd\u5316\u7684\u4e0d\u662f\u8ba1\u7b97\u8d44\u6e90\uff0c\u800c\u662fMAT\u62bd\u8c61\u7684\u4f4e\u6548\u4f7f\u7528\u3002Pegasus\u4ee5\u7b80\u5355\u4e14\u5b9e\u7528\u7684\u539f\u8bed\u8868\u8fbe\u65b9\u5f0f\uff0c\u4e3a\u7ebf\u901fDL\u63a8\u7406\u63d0\u4f9b\u4e86\u53ef\u884c\u8def\u5f84\uff0c\u4e5f\u4e3a\u672a\u6765\u6570\u636e\u5e73\u9762\u6301\u7eed\u667a\u80fd\u5316\u5960\u5b9a\u4e86\u57fa\u7840\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>SIGCOMM &#8217;25: ACM SIGCOMM 2025 Conference September 8 &#8211; 11, 2025 Coimbra, Portuga 1. \u6458\u8981\uff08Abstract\uff09 Pegasus\u9488\u5bf9\u5f53\u524dIntelligent DataPlane\uff08IDP\uff09\u4e2d\u6df1\u5ea6\u5b66\u4e60\uff08DL\uff09\u63a8\u7406\u9762\u4e34\u7684\u4e09\u5927\u6838\u5fc3\u9650\u5236\u2014\u2014\u7cbe\u5ea6\u4e0b\u964d\u3001\u53ef\u6269\u5c55\u6027\u5dee\u3001\u6cdb\u5316\u80fd\u529b\u4e0d\u8db3\u2014\u2014\u63d0\u51fa\u4e86\u4e00\u4e2a\u901a\u7528\u6846\u67b6\u3002\u5b83\u5c06DL\u64cd\u4f5c\u62bd\u8c61\u4e3a\u4e09\u79cd\u9762\u5411\u6570\u636e\u5e73\u9762\u7684\u539f\u8bed\uff1aPartition\uff08\u5206\u533a\uff09\u3001Map\uff08\u6620\u5c04\uff09\u548c &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=1161\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":5,"featured_media":1166,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[21,17,5,23,6],"tags":[13],"class_list":["post-1161","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-nft","category-17","category-rengongzhineng","category-23","category-weilaiwangluo","tag-13"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1161","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\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1161"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1161\/revisions"}],"predecessor-version":[{"id":1170,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1161\/revisions\/1170"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/1166"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1161"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1161"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1161"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}