{"id":1080,"date":"2026-03-19T11:05:34","date_gmt":"2026-03-19T03:05:34","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1080"},"modified":"2026-03-19T11:05:34","modified_gmt":"2026-03-19T03:05:34","slug":"pegasus-a-universal-framework-for-scalable-deep-learning-inference-on-the-dataplane","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1080","title":{"rendered":"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane"},"content":{"rendered":"\n<p>SIGCOMM &#8217;25: ACM SIGCOMM 2025 Conference <em>September 8 &#8211; 11, 2025<\/em> <em>Coimbra, Portugal<\/em><\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>1. \u8bfe\u9898\u80cc\u666f\u4e0e\u73b0\u6709\u65b9\u6848\u7684\u5c40\u9650\u6027<\/strong><\/h4>\n\n\n\n<p>\u968f\u7740\u7f51\u7edc\u6d41\u91cf\u7684\u7206\u70b8\u5f0f\u589e\u957f\uff0c\u5728\u6570\u636e\u5e73\u9762\u76f4\u63a5\u8fdb\u884c\u5b9e\u65f6\u667a\u80fd\u5206\u6790\u5df2\u6210\u4e3a\u5fc5\u7136\u8d8b\u52bf\u3002\u7136\u800c\uff0c\u73b0\u6709\u7684\u7814\u7a76\u65b9\u6848\u5728\u5e94\u5bf9\u590d\u6742\u7684\u6df1\u5ea6\u5b66\u4e60\uff08DL\uff09\u6a21\u578b\u65f6\u9762\u4e34\u4e25\u91cd\u7684\u6027\u80fd\u74f6\u9888\u3002\u4f20\u7edf\u7684\u201c\u8ba1\u7b97\u7b80\u5316\u201d\u7c7b\u65b9\u6848\uff0c\u5982 N3IC \u7b49\uff0c\u4e3b\u8981\u4f9d\u8d56\u4e8e\u5bf9\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u6781\u5ea6\u7684\u91cf\u5316\u6216\u4e8c\u503c\u5316\u5904\u7406\uff0c\u867d\u7136\u8fd9\u964d\u4f4e\u4e86\u5bf9\u4ea4\u6362\u673a\u7b97\u529b\u7684\u8981\u6c42\uff0c\u4f46\u7531\u4e8e\u4e22\u5931\u4e86\u5927\u91cf\u7684\u7279\u5f81\u6570\u503c\u7cbe\u5ea6\uff0c\u5bfc\u81f4\u5176\u5728\u5904\u7406\u590d\u6742\u5206\u7c7b\u4efb\u52a1\u65f6\u7684\u51c6\u786e\u7387\u5927\u6253\u6298\u6263\u3002\u8fd9\u79cd\u65b9\u6848\u5f80\u5f80\u53ea\u80fd\u5904\u7406\u975e\u5e38\u7b80\u5355\u7684\u7ebf\u6027\u4efb\u52a1\uff0c\u4e00\u65e6\u9762\u5bf9\u5177\u6709\u590d\u6742\u975e\u7ebf\u6027\u7279\u5f81\u7684\u6d41\u91cf\u6570\u636e\uff0c\u5176\u9884\u6d4b\u7ed3\u679c\u7684\u53ef\u9760\u6027\u4fbf\u65e0\u6cd5\u6ee1\u8db3\u5de5\u4e1a\u7ea7\u7f51\u7edc\u5b89\u5168\u548c\u7ba1\u7406\u7684\u9700\u6c42\u3002<\/p>\n\n\n\n<p>\u4e0e\u6b64\u540c\u65f6\uff0c\u53e6\u4e00\u79cd\u201c\u8ba1\u7b97\u7ed5\u8fc7\u201d\u7c7b\u65b9\u6848\uff0c\u5982 BoS\uff0c\u8bd5\u56fe\u901a\u8fc7\u9884\u5148\u5b58\u50a8\u8f93\u5165\u5230\u8f93\u51fa\u7684\u6620\u5c04\u8868\u6765\u5b8c\u5168\u907f\u514d\u5b9e\u65f6\u8ba1\u7b97\u3002\u8fd9\u79cd\u65b9\u6cd5\u7684\u81f4\u547d\u7f3a\u9677\u5728\u4e8e\u5176\u6781\u5dee\u7684\u6269\u5c55\u6027\uff1a\u968f\u7740\u8f93\u5165\u7279\u5f81\u7ef4\u5ea6\u7684\u589e\u52a0\uff0c\u53ef\u80fd\u7684\u7ec4\u5408\u7a7a\u95f4\u5448\u6307\u6570\u7ea7\u81a8\u80c0\uff0c\u8fd9\u5728\u5b66\u672f\u4e0a\u88ab\u79f0\u4e3a\u201c\u7ec4\u5408\u7206\u70b8\u201d\u95ee\u9898\u3002\u7531\u4e8e\u4ea4\u6362\u673a\u7684\u5185\u5b58\u8d44\u6e90\uff08\u5982 SRAM \u548c TCAM\uff09\u6781\u5176\u5b9d\u8d35\u4e14\u5bb9\u91cf\u56fa\u5b9a\uff0c\u8fd9\u7c7b\u65b9\u6848\u5f80\u5f80\u53ea\u80fd\u652f\u6301\u6781\u5c0f\u89c4\u6a21\u7684\u8f93\u5165\uff0c\u65e0\u6cd5\u5e94\u5bf9\u73b0\u4ee3\u7f51\u7edc\u534f\u8bae\u4e2d\u52a8\u8f84\u6570\u767e\u6bd4\u7279\u7684\u62a5\u6587\u7279\u5f81\u3002\u56e0\u6b64\uff0c\u4e1a\u754c\u8feb\u5207\u9700\u8981\u4e00\u79cd\u65e2\u80fd\u4fdd\u8bc1\u8ba1\u7b97\u7cbe\u5ea6\uff0c\u53c8\u80fd\u5728\u7ebf\u901f\u73af\u5883\u4e0b\u5904\u7406\u5927\u89c4\u6a21\u3001\u9ad8\u7ef4\u5ea6\u8f93\u5165\u7684\u901a\u7528\u63a8\u7406\u6846\u67b6\uff0c\u8fd9\u6b63\u662f Pegasus \u8bfe\u9898\u7684\u7814\u7a76\u8d77\u70b9\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"161\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/1-2-1024x161.png\"  class=\"wp-image-1081\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/1-2-1024x161.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/1-2-300x47.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/1-2-768x121.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/1-2.png 1176w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe\" \/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>2. Pegasus \u6838\u5fc3\u539f\u8bed\u4f53\u7cfb\u7684\u8bbe\u8ba1\u903b\u8f91<\/strong><\/h4>\n\n\n\n<p>Pegasus \u7684\u6838\u5fc3\u8d21\u732e\u5728\u4e8e\u5c06\u590d\u6742\u7684\u6df1\u5ea6\u5b66\u4e60\u8ba1\u7b97\u62bd\u8c61\u4e3a\u4e09\u4e2a\u57fa\u7840\u539f\u8bed\uff1aPartition\uff08\u5212\u5206\uff09\u3001Map\uff08\u6620\u5c04\uff09\u548c SumReduce\uff08\u548c\u5f52\u7ea6\uff09\u3002Partition \u539f\u8bed\u7684\u4f5c\u7528\u662f\u5c06\u9ad8\u7ef4\u5ea6\u7684\u8f93\u5165\u7279\u5f81\u5411\u91cf\u5207\u5206\u4e3a\u591a\u4e2a\u4f4e\u7ef4\u7684\u5b50\u5411\u91cf\uff0c\u8fd9\u79cd\u201c\u5206\u800c\u6cbb\u4e4b\u201d\u7684\u7b56\u7565\u80fd\u591f\u5c06\u539f\u672c\u65e0\u6cd5\u5728\u5355\u6b21\u5339\u914d\u4e2d\u5b8c\u6210\u7684\u8d85\u5927\u89c4\u6a21\u7a7a\u95f4\u6620\u5c04\u964d\u7ef4\u3002Map \u539f\u8bed\u5219\u5229\u7528\u4ea4\u6362\u673a\u7684\u5339\u914d-\u52a8\u4f5c\u8868\uff08MAT\uff09\u5e76\u884c\u67e5\u627e\u6bcf\u4e2a\u5b50\u5411\u91cf\u5bf9\u5e94\u7684\u4e2d\u95f4\u8ba1\u7b97\u7ed3\u679c\u3002SumReduce \u539f\u8bed\u8d1f\u8d23\u5c06\u8fd9\u4e9b\u5206\u6563\u7684\u4e2d\u95f4\u7ed3\u679c\u8fdb\u884c\u5206\u5c42\u805a\u5408\uff0c\u6700\u7ec8\u5728\u4ea4\u6362\u673a\u6d41\u6c34\u7ebf\u7684\u51fa\u53e3\u5904\u8f93\u51fa\u5b8c\u6574\u7684\u63a8\u7406\u7ed3\u8bba\u3002\u8fd9\u79cd\u539f\u8bed\u5316\u7684\u8bbe\u8ba1\u903b\u8f91\uff0c\u5c06 DL \u8fd0\u7b97\u4ece\u6602\u8d35\u7684\u6d6e\u70b9\u6570\u77e9\u9635\u8fd0\u7b97\u8f6c\u5316\u4e3a\u4e86\u4ea4\u6362\u673a\u64c5\u957f\u7684\u591a\u7ea7\u67e5\u8868\u64cd\u4f5c\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"624\" height=\"741\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/2-1.png\"  class=\"wp-image-1082\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/2-1.png 624w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/2-1-253x300.png 253w\" sizes=\"auto, (max-width: 624px) 100vw, 624px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe1\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe1\" \/><\/figure>\n<\/div>\n\n\n<p>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u4f18\u5316\u8d44\u6e90\u5229\u7528\u7387\uff0cPegasus \u5f15\u5165\u4e86\u521b\u65b0\u7684\u201c\u6a21\u7cca\u5339\u914d\u201d\uff08Fuzzy Matching\uff09\u6280\u672f\u3002\u5b83\u4e0d\u518d\u8ffd\u6c42\u8f93\u5165\u503c\u7684\u5b8c\u5168\u7cbe\u786e\u5339\u914d\uff0c\u800c\u662f\u901a\u8fc7\u805a\u7c7b\u7b97\u6cd5\u5c06\u76f8\u8fd1\u7684\u8f93\u5165\u503c\u6620\u5c04\u5230\u540c\u4e00\u4e2a\u8d28\u5fc3\u7d22\u5f15\uff08Fuzzy Index\uff09\u4e0a\u3002\u8fd9\u79cd\u8bbe\u8ba1\u4e0d\u4ec5\u6781\u5927\u5730\u538b\u7f29\u4e86\u6620\u5c04\u8868\u6240\u9700\u7684\u5b58\u50a8\u7a7a\u95f4\uff0c\u8fd8\u8d4b\u4e88\u4e86\u6846\u67b6\u5904\u7406\u8fde\u7eed\u6570\u503c\u7279\u5f81\u7684\u80fd\u529b\u3002\u6b64\u5916\uff0c\u901a\u8fc7\u201c\u539f\u8bed\u878d\u5408\u201d\uff08Primitive Fusion\uff09\u6280\u672f\uff0cPegasus \u80fd\u591f\u5c06\u8bf8\u5982\u6279\u5f52\u4e00\u5316\uff08Batch Normalization\uff09\u3001\u6fc0\u6d3b\u51fd\u6570\uff08ReLU\/Sigmoid\uff09\u548c\u7ebf\u6027\u53d8\u6362\u7b49\u591a\u4e2a\u8fde\u7eed\u7684\u795e\u7ecf\u5143\u64cd\u4f5c\u5408\u5e76\u5230\u540c\u4e00\u7ec4\u6620\u5c04\u8868\u4e2d\u5b8c\u6210\u3002\u8fd9\u79cd\u878d\u5408\u7b56\u7565\u6781\u5927\u5730\u8282\u7701\u4e86\u4ea4\u6362\u673a\u7684\u6d41\u6c34\u7ebf\u9636\u6bb5\uff08Stages\uff09\uff0c\u4f7f\u5f97\u5728\u6709\u9650\u7684\u786c\u4ef6\u6b65\u6570\u5185\u6267\u884c\u66f4\u6df1\u5c42\u3001\u66f4\u590d\u6742\u7684\u7f51\u7edc\u6a21\u578b\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"629\" height=\"558\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/6-1.png\"  class=\"wp-image-1086\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/6-1.png 629w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/6-1-300x266.png 300w\" sizes=\"auto, (max-width: 629px) 100vw, 629px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe2\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe2\" \/><\/figure>\n<\/div>\n\n\n<h4 class=\"wp-block-heading\"><strong>3. \u652f\u6301\u591a\u6837\u5316\u6a21\u578b\u67b6\u6784\u7684\u901a\u7528\u6027\u5b9e\u73b0<\/strong><\/h4>\n\n\n\n<p>Pegasus \u4e0d\u4ec5\u4ec5\u662f\u4e00\u4e2a\u9488\u5bf9\u7279\u5b9a\u4efb\u52a1\u7684\u63d2\u4ef6\uff0c\u800c\u662f\u4e00\u4e2a\u5177\u5907\u9ad8\u5ea6\u901a\u7528\u6027\u7684\u6846\u67b6\uff0c\u80fd\u591f\u652f\u6301 MLP\uff08\u591a\u5c42\u611f\u77e5\u673a\uff09\u3001RNN\uff08\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff09\u3001CNN\uff08\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff09\u4ee5\u53ca AutoEncoder\uff08\u81ea\u52a8\u7f16\u7801\u5668\uff09\u7b49\u591a\u79cd\u4e3b\u6d41\u6a21\u578b\u3002\u5728\u5b9e\u73b0 CNN \u65f6\uff0cPegasus \u5de7\u5999\u5730\u5c06\u5377\u79ef\u5c42\u7684\u6ed1\u52a8\u7a97\u53e3\u64cd\u4f5c\u8f6c\u5316\u4e3a\u6570\u636e\u5e73\u9762\u4e0a\u7684\u5b57\u6bb5\u504f\u79fb\u5904\u7406\uff0c\u901a\u8fc7 Map \u539f\u8bed\u5e76\u884c\u63d0\u53d6\u7a7a\u95f4\u7279\u5f81\uff1b\u5728\u5904\u7406 RNN \u7b49\u5177\u6709\u65f6\u95f4\u5e8f\u5217\u7279\u6027\u7684\u6a21\u578b\u65f6\uff0c\u6846\u67b6\u5229\u7528\u4ea4\u6362\u673a\u5185\u7684\u5bc4\u5b58\u5668\uff08Registers\uff09\u6765\u5b58\u50a8\u9690\u85cf\u72b6\u6001\uff08Hidden State\uff09\uff0c\u5b9e\u73b0\u4e86\u8de8\u62a5\u6587\u7684\u4fe1\u606f\u4f20\u9012\u3002\u8fd9\u79cd\u5bf9\u4e0d\u540c\u6a21\u578b\u7b97\u5b50\u7684\u7075\u6d3b\u89e3\u6784\uff0c\u8bc1\u660e\u4e86 Pegasus \u4f5c\u4e3a\u5e95\u5c42\u63a8\u7406\u5f15\u64ce\u7684\u5f3a\u5927\u666e\u9002\u6027\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"621\" height=\"675\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/3-1.png\"  class=\"wp-image-1083\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/3-1.png 621w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/3-1-276x300.png 276w\" sizes=\"auto, (max-width: 621px) 100vw, 621px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe3\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe3\" \/><\/figure>\n<\/div>\n\n\n<p>\u5728\u8f6f\u4ef6\u5c42\u9762\uff0cPegasus \u63d0\u4f9b\u4e86\u4e00\u5957\u5b8c\u6574\u7684\u7f16\u8bd1\u5668\u5957\u4ef6\uff0c\u80fd\u591f\u5c06\u7528\u6237\u8bad\u7ec3\u597d\u7684 PyTorch \u6216 TensorFlow \u6a21\u578b\u81ea\u52a8\u8f6c\u6362\u4e3a P4 \u4ee3\u7801\u548c\u5bf9\u5e94\u7684\u6d41\u8868\u9879\u3002\u8fd9\u4e00\u8fc7\u7a0b\u5305\u62ec\u4e86\u81ea\u52a8\u5316\u7684\u91cf\u5316\u7b56\u7565\u3001\u539f\u8bed\u6620\u5c04\u7b97\u6cd5\u4ee5\u53ca\u6d41\u6c34\u7ebf\u8d44\u6e90\u8c03\u5ea6\u4f18\u5316\u3002\u5f00\u53d1\u8005\u65e0\u9700\u6df1\u5165\u4e86\u89e3\u4ea4\u6362\u673a\u5e95\u5c42\u590d\u6742\u7684\u786c\u4ef6\u903b\u8f91\uff0c\u53ea\u9700\u63d0\u4f9b\u6807\u51c6\u6a21\u578b\uff0cPegasus \u5c31\u80fd\u5728\u4fdd\u8bc1\u5ef6\u8fdf\u548c\u541e\u5410\u91cf\u7684\u524d\u63d0\u4e0b\uff0c\u81ea\u52a8\u5b8c\u6210\u4ece\u9ad8\u5c42\u7b97\u6cd5\u5230\u4f4e\u5c42\u786c\u4ef6\u914d\u7f6e\u7684\u6620\u5c04\u3002\u8fd9\u79cd\u201c\u7b97\u6cd5\u5230\u786c\u4ef6\u201d\u7684\u7aef\u5230\u7aef\u81ea\u52a8\u5316\u80fd\u529b\uff0c\u6781\u5927\u5730\u964d\u4f4e\u4e86\u5728\u7f51\u7edc\u8bbe\u5907\u4e0a\u90e8\u7f72\u6df1\u5ea6\u5b66\u4e60\u7b97\u6cd5\u7684\u95e8\u69db\uff0c\u4f7f\u5f97\u7f51\u7edc\u5de5\u7a0b\u5e08\u80fd\u591f\u50cf\u5728 GPU \u4e0a\u4e00\u6837\u7075\u6d3b\u5730\u90e8\u7f72 AI \u7b56\u7565\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"603\" height=\"585\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/5.png\"  class=\"wp-image-1085\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/5.png 603w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/5-300x291.png 300w\" sizes=\"auto, (max-width: 603px) 100vw, 603px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe4\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe4\" \/><\/figure>\n<\/div>\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"618\" height=\"384\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/4-1.png\"  class=\"wp-image-1084\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/4-1.png 618w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/4-1-300x186.png 300w\" sizes=\"auto, (max-width: 618px) 100vw, 618px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe5\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe5\" \/><\/figure>\n<\/div>\n\n\n<h4 class=\"wp-block-heading\"><strong>4. \u5b9e\u9a8c\u8bc4\u4f30\u4e0e\u6027\u80fd\u6307\u6807\u5206\u6790<\/strong><\/h4>\n\n\n\n<p>\u7814\u7a76\u56e2\u961f\u5728\u771f\u5b9e\u7684 Barefoot Tofino \u786c\u4ef6\u4ea4\u6362\u673a\u4e0a\u5bf9 Pegasus \u8fdb\u884c\u4e86\u4e25\u683c\u7684\u8bc4\u4f30\uff0c\u8986\u76d6\u4e86\u591a\u4e2a\u6743\u5a01\u7684\u7f51\u7edc\u6d41\u91cf\u6570\u636e\u96c6\u3002\u5b9e\u9a8c\u6570\u636e\u8868\u660e\uff0cPegasus \u5728\u6240\u6709\u6d4b\u8bd5\u7684\u6a21\u578b\u67b6\u6784\u4e2d\u5747\u8868\u73b0\u51fa\u663e\u8457\u7684\u4f18\u8d8a\u6027\uff0c\u5176\u9884\u6d4b\u7cbe\u5ea6\u76f8\u6bd4\u4e4b\u524d\u7684\u6700\u5148\u8fdb\u65b9\u6848\uff08SOTA\uff09\u5e73\u5747\u63d0\u5347\u4e86 22.8%\u3002\u7279\u522b\u662f\u5728\u5904\u7406\u6d89\u53ca\u957f\u5e8f\u5217\u3001\u9ad8\u7ef4\u7279\u5f81\u7684\u5165\u4fb5\u68c0\u6d4b\u548c\u534f\u8bae\u8bc6\u522b\u4efb\u52a1\u65f6\uff0c\u7531\u4e8e\u5176\u91c7\u7528\u4e86\u6a21\u7cca\u5339\u914d\u548c\u6df7\u5408\u7cbe\u5ea6\u6280\u672f\uff0cPegasus \u80fd\u591f\u5728\u6781\u4f4e\u7684\u786c\u4ef6\u5f00\u9500\u4e0b\uff0c\u4fdd\u6301\u4e0e\u79bb\u7ebf\u670d\u52a1\u5668\u7aef\u63a8\u7406\u51e0\u4e4e\u4e00\u81f4\u7684\u51c6\u786e\u7387\u3002\u8fd9\u79cd\u7cbe\u5ea6\u7684\u8dc3\u5347\uff0c\u6807\u5fd7\u7740\u6570\u636e\u5e73\u9762\u63a8\u7406\u4ece\u201c\u53ef\u7528\u201d\u771f\u6b63\u8d70\u5411\u4e86\u201c\u53ef\u9760\u201d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"260\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/7-1-1024x260.png\"  class=\"wp-image-1087\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/7-1-1024x260.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/7-1-300x76.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/7-1-768x195.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/7-1.png 1275w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe6\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe6\" \/><\/figure>\n\n\n\n<p>\u5728\u8d44\u6e90\u6269\u5c55\u6027\u65b9\u9762\uff0cPegasus \u5c55\u73b0\u51fa\u4e86\u78be\u538b\u6027\u7684\u4f18\u52bf\u3002\u76f8\u6bd4\u4e8e BoS \u7b49\u65b9\u6848\uff0c\u5b83\u652f\u6301\u7684\u6a21\u578b\u53c2\u6570\u91cf\u6269\u5927\u4e86 248 \u500d\uff0c\u8f93\u5165\u7ef4\u5ea6\u652f\u6301\u80fd\u529b\u63d0\u5347\u4e86 212 \u500d\u3002\u5728\u541e\u5410\u91cf\u65b9\u9762\uff0cPegasus \u6210\u529f\u5b9e\u73b0\u4e86\u7ebf\u901f\u5904\u7406\uff0c\u5373\u4f7f\u662f\u5728\u6ee1\u8d1f\u8377\u7684 Tbps \u7ea7\u7f51\u7edc\u73af\u5883\u4e2d\uff0c\u5176\u5f15\u5165\u7684\u989d\u5916\u5ef6\u8fdf\u4e5f\u4ec5\u5728\u5fae\u79d2\u91cf\u7ea7\uff0c\u5b8c\u5168\u53ef\u4ee5\u5ffd\u7565\u4e0d\u8ba1\u3002\u6b64\u5916\uff0c\u5b9e\u9a8c\u8fd8\u901a\u8fc7\u5bf9\u6bd4\u4e0d\u540c\u6570\u91cf\u7684\u8d28\u5fc3\uff08Centroids\uff09\u5bf9\u7cbe\u5ea6\u7684\u5f71\u54cd\uff0c\u9a8c\u8bc1\u4e86\u6846\u67b6\u5728\u8d44\u6e90\u9650\u5236\u4e0e\u6027\u80fd\u8868\u73b0\u4e4b\u95f4\u7684\u826f\u597d\u5e73\u8861\u3002\u8fd9\u4e9b\u91cf\u5316\u6307\u6807\u5145\u5206\u8bc1\u660e\u4e86 Pegasus \u662f\u76ee\u524d\u5df2\u77e5\u552f\u4e00\u80fd\u591f\u540c\u65f6\u517c\u987e\u9ad8\u541e\u5410\u3001\u4f4e\u5ef6\u8fdf\u3001\u9ad8\u7cbe\u5ea6\u548c\u5f3a\u6269\u5c55\u6027\u7684\u6570\u636e\u5e73\u9762 AI \u63a8\u7406\u65b9\u6848\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"276\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/8-1024x276.png\"  class=\"wp-image-1088\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/8-1024x276.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/8-300x81.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/8-768x207.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/8.png 1260w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe7\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe7\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"252\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/9-1024x252.png\"  class=\"wp-image-1089\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/9-1024x252.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/9-300x74.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/9-768x189.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/9.png 1263w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe8\" alt=\"Pegasus: A Universal Framework for Scalable Deep Learning Inference on the Dataplane\u63d2\u56fe8\" \/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>5. \u7ed3\u8bba\u4e0e\u672a\u6765\u7814\u7a76\u4ef7\u503c<\/strong><\/h4>\n\n\n\n<p>Pegasus \u7684\u6210\u529f\u7814\u53d1\u5f7b\u5e95\u6539\u53d8\u4e86\u4e1a\u754c\u5bf9\u53ef\u7f16\u7a0b\u4ea4\u6362\u673a\u201c\u8ba1\u7b97\u80fd\u529b\u8d2b\u7620\u201d\u7684\u56fa\u6709\u8ba4\u77e5\u3002\u5b83\u8bc1\u660e\u4e86\u867d\u7136\u4ea4\u6362\u673a\u7684\u6307\u4ee4\u96c6\u6781\u5176\u7b80\u5316\uff0c\u4f46\u901a\u8fc7\u5de7\u5999\u7684\u62bd\u8c61\u5efa\u6a21\u548c\u539f\u8bed\u91cd\u6784\uff0c\u540c\u6837\u53ef\u4ee5\u627f\u8f7d\u539f\u672c\u5c5e\u4e8e GPU \u7684\u6df1\u5ea6\u5b66\u4e60\u8d1f\u8f7d\u3002Pegasus \u63d0\u4f9b\u4e86\u4e00\u5957\u6807\u51c6\u5316\u7684\u8bbe\u8ba1\u8303\u5f0f\uff0c\u5c06\u7f51\u7edc\u9886\u57df\u7684\u9700\u6c42\u4e0e AI \u9886\u57df\u7684\u8ba1\u7b97\u529b\u8fdb\u884c\u4e86\u5b8c\u7f8e\u7684\u5de5\u7a0b\u8854\u63a5\u3002\u5b83\u4e0d\u4ec5\u4e3a\u5b9e\u65f6\u7684\u7f51\u7edc\u6d41\u91cf\u6e05\u6d17\u3001\u667a\u80fd\u8def\u7531\u63a7\u5236\u548c\u52a8\u6001\u62e5\u585e\u7ba1\u7406\u63d0\u4f9b\u4e86\u6280\u672f\u57fa\u7840\uff0c\u4e5f\u4e3a\u672a\u6765\u201c\u7f51\u7edc\u5373\u8ba1\u7b97\u201d\uff08Network-as-Computing\uff09\u7684\u613f\u666f\u63d0\u4f9b\u4e86\u6781\u5177\u8bf4\u670d\u529b\u7684\u5b9e\u8df5\u8303\u4f8b\u3002<\/p>\n\n\n\n<p>\u4ece\u66f4\u6df1\u8fdc\u7684\u610f\u4e49\u6765\u770b\uff0cPegasus \u7684\u5f00\u6e90\u548c\u63a8\u5e7f\u5c06\u63a8\u52a8\u667a\u80fd\u6570\u636e\u5e73\u9762\u751f\u6001\u7cfb\u7edf\u7684\u53d1\u5c55\u3002\u5b83\u6240\u63d0\u51fa\u7684\u6a21\u7cca\u5339\u914d\u548c\u539f\u8bed\u878d\u5408\u601d\u60f3\uff0c\u53ef\u4ee5\u88ab\u8fdb\u4e00\u6b65\u501f\u9274\u5230\u5176\u4ed6\u8d44\u6e90\u53d7\u9650\u7684\u786c\u4ef6\u73af\u5883\uff08\u5982\u8fb9\u7f18\u8ba1\u7b97\u8bbe\u5907\u6216\u5d4c\u5165\u5f0f FPGA\uff09\u4e2d\u3002\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\u53ef\u80fd\u5305\u62ec\u5982\u4f55\u5728 Pegasus \u6846\u67b6\u4e0b\u5b9e\u73b0\u6a21\u578b\u7684\u52a8\u6001\u5728\u7ebf\u66f4\u65b0\uff0c\u4ee5\u53ca\u5982\u4f55\u8fdb\u4e00\u6b65\u652f\u6301\u66f4\u52a0\u524d\u6cbf\u7684 Transformer \u7b49\u6ce8\u610f\u529b\u673a\u5236\u6a21\u578b\u3002\u603b\u800c\u8a00\u4e4b\uff0cPegasus \u662f\u6570\u636e\u5e73\u9762\u667a\u80fd\u5316\u8fdb\u7a0b\u4e2d\u7684\u4e00\u5ea7\u91cc\u7a0b\u7891\uff0c\u5b83\u4e3a\u6784\u5efa\u66f4\u5b89\u5168\u3001\u66f4\u9ad8\u6548\u3001\u66f4\u5177\u611f\u77e5\u529b\u7684\u4e0b\u4e00\u4ee3\u672a\u6765\u7f51\u7edc\u5960\u5b9a\u4e86\u575a\u5b9e\u7684\u7b97\u6cd5\u4e0e\u67b6\u6784\u57fa\u77f3\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>SIGCOMM &#8217;25: ACM SIGCOMM 2025 Conference September 8 &#8211; 11, 2025 Coimbra, Portugal 1. \u8bfe\u9898\u80cc\u666f\u4e0e\u73b0\u6709\u65b9\u6848\u7684\u5c40\u9650\u6027 \u968f\u7740\u7f51\u7edc\u6d41\u91cf\u7684\u7206\u70b8\u5f0f\u589e\u957f\uff0c\u5728\u6570\u636e\u5e73\u9762\u76f4\u63a5\u8fdb\u884c\u5b9e\u65f6\u667a\u80fd\u5206\u6790\u5df2\u6210\u4e3a\u5fc5\u7136\u8d8b\u52bf\u3002\u7136\u800c\uff0c\u73b0\u6709\u7684\u7814\u7a76\u65b9\u6848\u5728\u5e94\u5bf9\u590d\u6742\u7684\u6df1\u5ea6\u5b66\u4e60\uff08DL\uff09\u6a21\u578b\u65f6\u9762\u4e34\u4e25\u91cd\u7684\u6027\u80fd\u74f6\u9888\u3002\u4f20\u7edf\u7684\u201c\u8ba1\u7b97\u7b80\u5316\u201d\u7c7b\u65b9\u6848\uff0c\u5982 N3IC \u7b49\uff0c\u4e3b\u8981\u4f9d\u8d56\u4e8e\u5bf9\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u6781\u5ea6\u7684\u91cf\u5316\u6216\u4e8c\u503c\u5316\u5904\u7406\uff0c\u867d\u7136\u8fd9\u964d\u4f4e\u4e86\u5bf9\u4ea4\u6362\u673a &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=1080\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":5,"featured_media":1090,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[17,5,23],"tags":[13],"class_list":["post-1080","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-17","category-rengongzhineng","category-23","tag-13"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1080","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=1080"}],"version-history":[{"count":2,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1080\/revisions"}],"predecessor-version":[{"id":1092,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1080\/revisions\/1092"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/1090"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1080"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1080"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1080"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}