{"id":4086,"date":"2024-06-19T18:27:00","date_gmt":"2024-06-19T10:27:00","guid":{"rendered":"http:\/\/www.yuntou8.com\/?p=4086"},"modified":"2024-06-19T18:27:00","modified_gmt":"2024-06-19T10:27:00","slug":"ai%e8%8a%af%e7%89%87%e7%9a%84%e6%9c%aa%e6%9d%a5%ef%bc%8c%e6%9c%aa%e5%bf%85%e6%98%afgpu","status":"publish","type":"post","link":"http:\/\/www.yuntou8.com\/?p=4086","title":{"rendered":"AI\u82af\u7247\u7684\u672a\u6765\uff0c\u672a\u5fc5\u662fGPU"},"content":{"rendered":"

\u5728\u4eba\u5de5\u667a\u80fd\u8ba1\u7b97\u67b6\u6784\u7684\u5e03\u5c40\u4e2d\uff0cCPU\u4e0e\u52a0\u901f\u82af\u7247\u534f\u540c\u5de5\u4f5c\u7684\u6a21\u5f0f\u5df2\u6210\u4e3a\u4e00\u79cd\u5178\u578b\u7684AI\u90e8\u7f72\u65b9\u6848\u3002CPU\u626e\u6f14\u57fa\u7840\u7b97\u529b\u7684\u63d0\u4f9b\u8005\u89d2\u8272\uff0c\u800c\u52a0\u901f\u82af\u7247\u5219\u8d1f\u8d23\u63d0\u5347\u8ba1\u7b97\u6027\u80fd\uff0c\u52a9\u529b\u7b97\u6cd5\u9ad8\u6548\u6267\u884c\u3002\u5e38\u89c1\u7684AI\u52a0\u901f\u82af\u7247\u6309\u5176\u6280\u672f\u8def\u5f84\uff0c\u53ef\u5212\u5206\u4e3aGPU\u3001FPGA\u548cASIC\u4e09\u5927\u7c7b\u522b\u3002,\u5728\u8fd9\u573a\u7ade\u4e89\u4e2d\uff0cGPU\u51ed\u501f\u5176\u72ec\u7279\u7684\u4f18\u52bf\u6210\u4e3a\u4e3b\u6d41\u7684AI\u82af\u7247\u3002\u90a3\u4e48\uff0cGPU\u662f\u5982\u4f55\u5728\u4f17\u591a\u9009\u9879\u4e2d\u8131\u9896\u800c\u51fa\u7684\u5462\uff1f\u5c55\u671bAI\u7684\u672a\u6765\uff0cGPU\u662f\u5426\u4ecd\u662f*\u89e3\u5462\uff1f,
\n01 <\/strong>
\nGPU\u5982\u4f55\u5236\u80dc\u5f53\u4e0b\uff1f<\/strong><\/strong><\/strong>,AI\u4e0eGPU\u4e4b\u95f4\u5b58\u5728\u7740\u5bc6\u5207\u7684\u5173\u7cfb\u3002,
\n\u5f3a\u5927\u7684\u5e76\u884c\u8ba1\u7b97\u80fd\u529b<\/strong>,AI\u5927\u6a21\u578b\u6307\u7684\u662f\u89c4\u6a21\u5e9e\u5927\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\uff0c\u5b83\u4eec\u9700\u8981\u5904\u7406\u6d77\u91cf\u7684\u6570\u636e\u548c\u8fdb\u884c\u590d\u6742\u7684\u8ba1\u7b97\u3002GPU\u7684\u6838\u5fc3\u4f18\u52bf\u5c31\u5728\u4e8e\u5176\u5f3a\u5927\u7684\u5e76\u884c\u8ba1\u7b97\u80fd\u529b\u3002\u4e0e\u4f20\u7edf\u7684CPU\u76f8\u6bd4\uff0cGPU\u80fd\u591f\u540c\u65f6\u5904\u7406\u591a\u4e2a\u4efb\u52a1\uff0c\u7279\u522b\u9002\u5408\u5904\u7406\u5927\u89c4\u6a21\u6570\u636e\u96c6\u548c\u590d\u6742\u8ba1\u7b97\u4efb\u52a1\u3002\u5728\u6df1\u5ea6\u5b66\u4e60\u7b49\u9700\u8981\u5927\u91cf\u5e76\u884c\u8ba1\u7b97\u7684\u9886\u57df\uff0cGPU\u5c55\u73b0\u51fa\u4e86\u65e0\u53ef\u6bd4\u62df\u7684\u4f18\u52bf\u3002,
\n\u5b8c\u5584\u7684\u751f\u6001\u7cfb\u7edf<\/strong>,\u5176\u6b21\uff0c\u4e3a\u4e86\u4fbf\u4e8e\u5f00\u53d1\u8005\u5145\u5206\u5229\u7528GPU\u7684\u8ba1\u7b97\u80fd\u529b\uff0c\u5404\u5927\u5382\u5546\u63d0\u4f9b\u4e86\u4e30\u5bcc\u7684\u8f6f\u4ef6\u5e93\u3001\u6846\u67b6\u548c\u5de5\u5177\u3002\u4f8b\u5982\uff0c\u82f1\u4f1f\u8fbe\u7684CUDA\u5e73\u53f0\u5c31\u4e3a\u5f00\u53d1\u8005\u63d0\u4f9b\u4e86\u4e30\u5bcc\u7684\u5de5\u5177\u548c\u5e93\uff0c\u4f7f\u5f97AI\u5e94\u7528\u7684\u5f00\u53d1\u548c\u90e8\u7f72\u53d8\u5f97\u76f8\u5bf9\u5bb9\u6613\u3002\u8fd9\u4f7f\u5f97GPU\u5728\u9700\u8981\u5feb\u901f\u8fed\u4ee3\u548c\u9002\u5e94\u65b0\u7b97\u6cd5\u7684\u573a\u666f\u4e2d\u66f4\u5177\u7ade\u4e89\u529b\u3002,
\n\u901a\u7528\u6027\u597d<\/strong>,GPU\u6700\u521d\u662f\u7528\u4e8e\u56fe\u5f62\u6e32\u67d3\u7684\uff0c\u4f46\u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\uff0c\u5b83\u7684\u5e94\u7528\u9886\u57df\u9010\u6e10\u6269\u5927\u3002\u5982\u4eca\uff0cGPU\u4e0d\u4ec5\u5728\u56fe\u5f62\u5904\u7406\u4e2d\u53d1\u6325\u7740\u6838\u5fc3\u4f5c\u7528\uff0c\u8fd8\u5e7f\u6cdb\u5e94\u7528\u4e8e\u6df1\u5ea6\u5b66\u4e60\u3001\u5927\u6570\u636e\u5206\u6790\u7b49\u9886\u57df\u3002\u8fd9\u79cd\u901a\u7528\u6027\u4f7f\u5f97GPU\u80fd\u591f\u6ee1\u8db3\u591a\u79cd\u5e94\u7528\u9700\u6c42\uff0c\u800cASIC\u548cFPGA\u7b49\u4e13\u7528\u82af\u7247\u5219\u5c40\u9650\u4e8e\u7279\u5b9a\u573a\u666f\u3002,\u6709\u4eba\u5c06GPU\u6bd4\u4f5c\u4e00\u628a\u901a\u7528\u7684\u591a\u529f\u80fd\u53a8\u5177\uff0c\u9002\u7528\u4e8e\u5404\u79cd\u70f9\u996a\u9700\u6c42\u3002\u56e0\u6b64\u5728AI\u5e94\u7528\u7684\u5927\u591a\u6570\u60c5\u51b5\u4e0b\uff0cGPU\u90fd\u88ab\u89c6\u4e3a*\u9009\u62e9\u3002\u76f8\u5e94\u7684\uff0c\u529f\u80fd\u591a\u800c\u5e7f\u7684\u540c\u65f6\u5f80\u5f80\u4f34\u968f\u7740\u7279\u5b9a\u9886\u57df\u4e0d\u591f\u201c\u7cbe\u7ec6\u201d\uff0c,\u63a5\u4e0b\u6765\u770b\u4e00\u4e0b\uff0c\u76f8\u8f83\u5176\u4ed6\u7c7b\u578b\u7684\u52a0\u901f\u82af\u7247\uff0cGPU\u9700\u8981\u9762\u4e34\u54ea\u4e9b\u63a3\u8098\uff1f,
\n02 <\/strong>
\nGPU\u4e5f\u5b58\u5728\u5b83\u7684\u63a3\u8098<\/strong><\/strong><\/strong>,\u6587\u9996\u63d0\u5230\uff0c\u5e38\u89c1\u7684AI\u52a0\u901f\u82af\u7247\u6839\u636e\u5176\u6280\u672f\u8def\u5f84\uff0c\u53ef\u4ee5\u5212\u5206\u4e3aGPU\u3001FPGA\u548cASIC\u4e09\u5927\u7c7b\u522b\u3002,,FPGA\uff08Field Programmable Gate Array\uff0c\u73b0\u573a\u53ef\u7f16\u7a0b\u95e8\u9635\u5217\uff09\uff0c\u662f\u4e00\u79cd\u534a\u5b9a\u5236\u82af\u7247\u3002\u7528\u6237\u53ef\u4ee5\u6839\u636e\u81ea\u8eab\u7684\u9700\u6c42\u8fdb\u884c\u91cd\u590d\u7f16\u7a0b\u3002FPGA \u7684\u4f18\u70b9\u662f\u65e2\u89e3\u51b3\u4e86\u5b9a\u5236\u7535\u8def\u7684\u4e0d\u8db3\uff0c\u53c8\u514b\u670d\u4e86\u539f\u6709\u53ef\u7f16\u7a0b\u5668\u4ef6\u95e8\u7535\u8def\u6570\u6709\u9650\u7684\u7f3a\u70b9\uff0c\u5bf9\u82af\u7247\u786c\u4ef6\u5c42\u53ef\u4ee5\u7075\u6d3b\u7f16\u8bd1\uff0c\u529f\u8017\u5c0f\u4e8e CPU\u3001GPU\uff1b\u7f3a\u70b9\u662f\u786c\u4ef6\u7f16\u7a0b\u8bed\u8a00\u8f83\u96be\uff0c\u5f00\u53d1\u95e8\u69db\u8f83\u9ad8\uff0c\u82af\u7247\u6210\u672c\u3001\u4ef7\u683c\u8f83\u9ad8\u3002FPGA \u6bd4 GPU\u3001CPU \u66f4\u5feb\u662f\u56e0\u4e3a\u5176\u5177\u6709\u5b9a\u5236\u5316\u7684\u7ed3\u6784\u3002,ASIC\uff08Application Specific Integrated Circuit\u7279\u5b9a\u7528\u9014\u96c6\u6210\u7535\u8def\uff09\u6839\u636e\u4ea7\u54c1\u7684\u9700\u6c42\u8fdb\u884c\u7279\u5b9a\u8bbe\u8ba1\u548c\u5236\u9020\u7684\u96c6\u6210\u7535\u8def\uff0c\u5176\u5b9a\u5236\u7a0b\u5ea6\u76f8\u6bd4\u4e8e GPU \u548c FPGA \u66f4\u9ad8\u3002ASIC \u7b97\u529b\u6c34\u5e73\u4e00\u822c\u9ad8\u4e8eGPU\u3001FPGA\uff0c\u4f46\u521d\u59cb\u6295\u5165\u5927\uff0c\u4e13\u4e1a\u6027\u5f3a\u7f29\u51cf\u4e86\u5176\u901a\u7528\u6027\uff0c\u7b97\u6cd5\u4e00\u65e6\u6539\u53d8\uff0c\u8ba1\u7b97\u80fd\u529b\u4f1a\u5927\u5e45\u4e0b\u964d\uff0c\u9700\u8981\u91cd\u65b0\u5b9a\u5236\u3002,\u518d\u770bGPU\u76f8\u8f83\u4e8e\u8fd9\u4e24\u7c7b\u82af\u7247\u5b58\u5728\u54ea\u4e9b\u52a3\u52bf\u3002,
\n*\u70b9\uff0cGPU\u7684\u5355\u4f4d\u6210\u672c\u7406\u8bba\u6027\u80fd\u4f4e\u4e8eFPGA\u3001ASIC\u3002<\/strong><\/strong>,\u4ece\u6210\u672c\u89d2\u5ea6\u770b\uff0cGPU\u3001FPGA\u3001ASIC \u4e09\u79cd\u786c\u4ef6\u4ece\u5de6\u5230\u53f3\uff0c\u4ece\u8f6f\u4ef6\u5230\u786c\u4ef6\uff0c\u901a\u7528\u6027\u9010\u6e10\u964d\u4f4e\u3001\u8d8a\u4e13\u7528\uff0c\u53ef\u5b9a\u5236\u5316\u9010\u6e10\u63d0\u9ad8\uff0c\u76f8\u5e94\u7684\u8bbe\u8ba1\u3001\u5f00\u53d1\u6210\u672c\u9010\u6e10\u63d0\u9ad8\uff0c\u4f46\u662f\u5355\u4f4d\u6210\u672c\u7406\u8bba\u6027\u80fd\u8d8a\u9ad8\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5bf9\u4e8e\u8fd8\u5728\u5b9e\u9a8c\u5ba4\u9636\u6bb5\u7684\u7ecf\u5178\u7b97\u6cd5\u6216\u6df1\u5ea6\u5b66\u4e60\u7b97\u6cd5\uff0c\u4f7f\u7528GPU \u505a\u8f6f\u4ef6\u65b9\u9762\u7684\u63a2\u7d22\u5c31\u5f88\u5408\u9002\uff1b\u5bf9\u4e8e\u5df2\u7ecf\u9010\u6e10\u6210\u4e3a\u6807\u51c6\u7684\u6280\u672f\uff0c\u9002\u5408\u4f7f\u7528 FPGA \u505a\u786c\u4ef6\u52a0\u901f\u90e8\u7f72\uff1b\u5bf9\u4e8e\u5df2\u7ecf\u6210\u4e3a\u6807\u51c6\u7684\u8ba1\u7b97\u4efb\u52a1\uff0c\u5219\u76f4\u63a5\u63a8\u51fa\u4e13\u7528\u82af\u7247ASIC\u3002,\u4ece\u516c\u53f8\u7684\u89d2\u5ea6\u6765\u8bf4\uff0c\u540c\u6837\u5bf9\u4e8e\u5927\u6279\u91cf\u6570\u636e\u7684\u8ba1\u7b97\u4efb\u52a1\uff0c\u540c\u7b49\u5185\u5b58\u5927\u5c0f\u3001\u540c\u7b49\u7b97\u529b\u7684\u6210\u719f GPU \u548c FPGA \u7684\u90e8\u7f72\u6210\u672c\u76f8\u8fd1\u3002 \u5982\u679c\u516c\u53f8\u7684\u4e1a\u52a1\u903b\u8f91\u7ecf\u5e38\u53d8\u5316\uff0c\u6bd4\u59821-2\u5e74\u5c31\u8981\u53d8\u5316\u4e00\u6b21\uff0c\u90a3\u4e48GPU \u7684\u5f00\u53d1\u6210\u672c\u4f4e\u3001\u90e8\u7f72\u901f\u5ea6\u5feb\uff1b\u5982\u679c\u516c\u53f8\u4e1a\u52a15\u5e74\u5de6\u53f3\u624d\u53d8\u5316\u4e00\u6b21\uff0cFPGA \u5f00\u53d1\u6210\u672c\u867d\u9ad8\u3001\u4f46\u82af\u7247\u672c\u8eab\u7684\u6210\u672c\u76f8\u6bd4 GPU \u4f4e\u5f88\u591a\u3002,
\n\u7b2c\u4e8c\u70b9\uff0cGPU\u7684\u8fd0\u7b97\u901f\u5ea6\u8981\u900a\u8272\u4e8eFPGA\u548cASIC\u3002<\/strong><\/strong>,FPGA\u3001ASIC\u548cGPU\u5185\u90fd\u6709\u5927\u91cf\u7684\u8ba1\u7b97\u5355\u5143\uff0c\u56e0\u6b64\u5b83\u4eec\u7684\u8ba1\u7b97\u80fd\u529b\u90fd\u5f88\u5f3a\u3002\u5728\u8fdb\u884c\u795e\u7ecf\u7f51\u7edc\u8fd0\u7b97\u7684\u65f6\u5019\uff0c\u4e09\u8005\u7684\u901f\u5ea6\u4f1a\u6bd4CPU\u5feb\u5f88\u591a\u3002\u4f46\u662fGPU\u7531\u4e8e\u67b6\u6784\u56fa\u5b9a\uff0c\u786c\u4ef6\u539f\u751f\u652f\u6301\u7684\u6307\u4ee4\u4e5f\u5c31\u56fa\u5b9a\u4e86\uff0c\u800cFPGA\u548cASIC\u5219\u662f\u53ef\u7f16\u7a0b\u7684\uff0c\u5176\u53ef\u7f16\u7a0b\u6027\u662f\u5173\u952e\uff0c\u56e0\u4e3a\u5b83\u8ba9\u8f6f\u4ef6\u4e0e\u7ec8\u7aef\u5e94\u7528\u516c\u53f8\u80fd\u591f\u63d0\u4f9b\u4e0e\u5176\u7ade\u4e89\u5bf9\u624b\u4e0d\u540c\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u5e76\u4e14\u80fd\u591f\u7075\u6d3b\u5730\u9488\u5bf9\u81ea\u5df1\u6240\u7528\u7684\u7b97\u6cd5\u4fee\u6539\u7535\u8def\u3002,\u56e0\u6b64\u5728\u5f88\u591a\u573a\u666f\u7684\u5e94\u7528\u4e2d\uff0cFPGA\u548cASIC\u7684\u8fd0\u7b97\u901f\u5ea6\u8981\u5927\u5927\u4f18\u4e8eGPU\u3002,\u5177\u4f53\u5230\u573a\u666f\u5e94\u7528\uff0cGPU \u6d6e\u70b9\u8fd0\u7b97\u80fd\u529b\u5f88\u5f3a\uff0c\u9002\u5408\u9ad8\u7cbe\u5ea6\u7684\u795e\u7ecf\u7f51\u7edc\u8ba1\u7b97\uff1bFPGA \u5e76\u4e0d\u64c5\u957f\u6d6e\u70b9\u8fd0\u7b97\uff0c\u4f46\u662f\u5bf9\u4e8e\u7f51\u7edc\u6570\u636e\u5305\u3001\u89c6\u9891\u6d41\u53ef\u4ee5\u505a\u5230\u5f88\u5f3a\u7684\u6d41\u6c34\u7ebf\u5904\u7406\uff1bASIC \u5219\u6839\u636e\u6210\u672c\u6709\u51e0\u4e4e\u65e0\u9650\u7684\u7b97\u529b\uff0c\u53d6\u51b3\u4e8e\u786c\u4ef6\u8bbe\u8ba1\u8005\u3002,
\n\u7b2c\u4e09\u70b9\uff0cGPU\u7684\u529f\u8017\u8fdc\u8fdc\u5927\u4e8eFPGA\u548cASIC\u3002<\/strong><\/strong>,\u518d\u770b\u529f\u8017\u3002GPU\u7684\u529f\u8017\uff0c\u662f\u51fa\u4e86\u540d\u7684\u9ad8\uff0c\u5355\u7247\u53ef\u4ee5\u8fbe\u5230250W\uff0c\u751a\u81f3450W\uff08RTX4090\uff09\u3002\u800cFPGA\u4e00\u822c\u53ea\u670930~50W\u3002\u8fd9\u4e3b\u8981\u662f\u56e0\u4e3a\u5185\u5b58\u8bfb\u53d6\u3002GPU\u7684\u5185\u5b58\u63a5\u53e3\uff08GDDR5\u3001HBM\u3001HBM2\uff09\u5e26\u5bbd\u6781\u9ad8\uff0c\u5927\u7ea6\u662fFPGA\u4f20\u7edfDDR\u63a5\u53e3\u76844-5\u500d\u3002\u4f46\u5c31\u82af\u7247\u672c\u8eab\u6765\u8bf4\uff0c\u8bfb\u53d6DRAM\u6240\u6d88\u8017\u7684\u80fd\u91cf\uff0c\u662fSRAM\u7684100\u500d\u4ee5\u4e0a\u3002GPU\u9891\u7e41\u8bfb\u53d6DRAM\u7684\u5904\u7406\uff0c\u4ea7\u751f\u4e86\u6781\u9ad8\u7684\u529f\u8017\u3002\u53e6\u5916\uff0cFPGA\u7684\u5de5\u4f5c\u4e3b\u9891\uff08500MHz\u4ee5\u4e0b\uff09\u6bd4CPU\u3001GPU\uff081~3GHz\uff09\u4f4e\uff0c\u4e5f\u4f1a\u4f7f\u5f97\u81ea\u8eab\u529f\u8017\u66f4\u4f4e\u3002,\u518d\u770bASIC\uff0cASIC\u7684\u6027\u80fd\u548c\u529f\u8017\u4f18\u5316\u662f\u9488\u5bf9\u7279\u5b9a\u5e94\u7528\u8fdb\u884c\u7684\uff0c\u56e0\u6b64\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u6027\u80fd\u66f4\u9ad8\u3001\u529f\u8017\u66f4\u4f4e\u3002\u7531\u4e8e\u8bbe\u8ba1\u662f\u9488\u5bf9\u7279\u5b9a\u529f\u80fd\u7684\uff0cASIC\u5728\u6267\u884c\u6548\u7387\u548c\u80fd\u6548\u6bd4\u65b9\u9762\u901a\u5e38\u4f18\u4e8eFPGA\u3002,\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5728\u667a\u80fd\u9a7e\u9a76\u8fd9\u6837\u7684\u9886\u57df\uff0c\u73af\u5883\u611f\u77e5\u3001\u7269\u4f53\u8bc6\u522b\u7b49\u6df1\u5ea6\u5b66\u4e60\u5e94\u7528\u8981\u6c42\u8ba1\u7b97\u54cd\u5e94\u65b9\u9762\u5fc5\u987b\u66f4\u5feb\u7684\u540c\u65f6\uff0c\u529f\u8017\u4e5f\u4e0d\u80fd\u8fc7\u9ad8\uff0c\u5426\u5219\u5c31\u4f1a\u5bf9\u667a\u80fd\u6c7d\u8f66\u7684\u7eed\u822a\u91cc\u7a0b\u9020\u6210\u8f83\u5927\u5f71\u54cd\u3002,
\n\u7b2c\u56db\u70b9\uff0cGPU\u65f6\u5ef6\u9ad8\u4e8eFPGA\u3001ASIC\u3002<\/strong>FPGA\u76f8\u5bf9\u4e8eGPU\u5177\u6709\u66f4\u4f4e\u7684\u5ef6\u8fdf\u3002GPU\u901a\u5e38\u9700\u8981\u5c06\u4e0d\u540c\u7684\u8bad\u7ec3\u6837\u672c\uff0c\u5212\u5206\u6210\u56fa\u5b9a\u5927\u5c0f\u7684\u201cBatch\uff08\u6279\u6b21\uff09\u201d\uff0c\u4e3a\u4e86*\u5316\u8fbe\u5230\u5e76\u884c\u6027\uff0c\u9700\u8981\u5c06\u6570\u4e2aBatch\u90fd\u96c6\u9f50\uff0c\u518d\u7edf\u4e00\u8fdb\u884c\u5904\u7406\u3002,FPGA\u7684\u67b6\u6784\uff0c\u662f\u65e0\u6279\u6b21\u7684\u3002\u6bcf\u5904\u7406\u5b8c\u6210\u4e00\u4e2a\u6570\u636e\u5305\uff0c\u5c31\u80fd\u9a6c\u4e0a\u8f93\u51fa\uff0c\u65f6\u5ef6\u66f4\u6709\u4f18\u52bf\u3002ASIC\u4e5f\u662f\u5b9e\u73b0\u6781\u4f4e\u5ef6\u8fdf\u7684\u53e6\u4e00\u79cd\u6280\u672f\u3002\u5728\u9488\u5bf9\u7279\u5b9a\u4efb\u52a1\u8fdb\u884c\u4f18\u5316\u540e\uff0cASIC\u901a\u5e38\u80fd\u591f\u5b9e\u73b0\u6bd4FPGA\u66f4\u4f4e\u7684\u5ef6\u8fdf\uff0c\u56e0\u4e3a\u5b83\u53ef\u4ee5\u6d88\u9664FPGA\u4e2d\u53ef\u80fd\u5b58\u5728\u7684\u989d\u5916\u7f16\u7a0b\u548c\u914d\u7f6e\u5f00\u9500\u3002,\u65e2\u5982\u6b64\uff0c\u4e3a\u4ec0\u4e48GPU\u8fd8\u4f1a\u6210\u4e3a\u73b0\u4e0bAI\u8ba1\u7b97\u7684\u5927\u70ed\u95e8\u5462\uff1f,\u5728\u5f53\u524d\u7684\u5e02\u573a\u73af\u5883\u4e0b\uff0c\u7531\u4e8e\u5404\u5927\u5382\u5546\u5bf9\u4e8e\u6210\u672c\u548c\u529f\u8017\u7684\u8981\u6c42\u5c1a\u672a\u8fbe\u5230\u4e25\u82db\u7684\u7a0b\u5ea6\uff0c\u52a0\u4e4b\u82f1\u4f1f\u8fbe\u5728GPU\u9886\u57df\u7684\u957f\u671f\u6295\u5165\u548c\u79ef\u7d2f\uff0c\u4f7f\u5f97GPU\u6210\u4e3a\u4e86\u5f53\u524d\u6700\u9002\u5408\u5927\u6a21\u578b\u5e94\u7528\u7684\u786c\u4ef6\u4ea7\u54c1\u3002\u5c3d\u7ba1FPGA\u548cASIC\u5728\u7406\u8bba\u4e0a\u5177\u6709\u6f5c\u5728\u7684\u4f18\u52bf\uff0c\u4f46\u5b83\u4eec\u7684\u5f00\u53d1\u8fc7\u7a0b\u76f8\u5bf9\u590d\u6742\uff0c\u76ee\u524d\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u4ecd\u9762\u4e34\u8bf8\u591a\u6311\u6218\uff0c\u96be\u4ee5\u5e7f\u6cdb\u666e\u53ca\u3002\u56e0\u6b64\uff0c\u4f17\u591a\u5382\u5546\u7eb7\u7eb7\u9009\u62e9GPU\u4f5c\u4e3a\u89e3\u51b3\u65b9\u6848\uff0c\u8fd9\u4e5f\u5bfc\u81f4\u4e86\u7b2c\u4e94\u70b9\u6f5c\u5728\u95ee\u9898\u7684\u6d6e\u73b0\u3002,
\n\u7b2c\u4e94\u70b9\uff0c\u9ad8\u7aefGPU\u7684\u4ea7\u80fd\u95ee\u9898\u4e5f\u4ee4\u4eba\u7126\u8651\u3002<\/strong>,OpenAI \u9996\u5e2d\u79d1\u5b66\u5bb6 IlyaSutskever\u2002\u8868\u793a\uff0cGPU \u5c31\u662f\u65b0\u65f6\u4ee3\u7684\u6bd4\u7279\u5e01\u3002\u5728\u7b97\u529b\u6fc0\u589e\u7684\u80cc\u666f\u4e0b\uff0c\u82f1\u4f1f\u8fbe\u7684B\u7cfb\u5217\u548cH\u7cfb\u5217 GPU \u6210\u4e3a\u201c\u786c\u901a\u8d27\u201d\u3002,\u7136\u800c\uff0c\u867d\u7136\u8be5\u7cfb\u5217\u9700\u6c42\u5341\u5206\u65fa\u76db\uff0c\u4f46\u8003\u8651\u5230HBM\u548cCoWos\u4f9b\u9700\u7d27\u5f20\uff0c\u4ee5\u53ca\u53f0\u79ef\u7535\u5148\u8fdb\u4ea7\u80fd\u5403\u7d27\u7684\u60c5\u51b5\uff0cGPU\u4ea7\u80fd\u5b9e\u5728\u65e0\u6cd5\u8ddf\u5f97\u4e0a\u9700\u6c42\u3002,\u8981\u77e5\u9053\u201c\u5de7\u5987\u96be\u4e3a\u65e0\u7c73\u4e4b\u708a\u201d\uff0c\u5728\u8fd9\u79cd\u5f62\u52bf\u4e0b\uff0c\u79d1\u6280\u5de8\u5934\u4eec\u9700\u8981\u66f4\u52a0\u7075\u6d3b\u5730\u5e94\u5bf9\u5e02\u573a\u53d8\u5316\uff0c\u56e4\u79ef\u66f4\u591a\u7684GPU\u4ea7\u54c1\u6216\u8005\u5bfb\u627e\u66ff\u4ee3\u65b9\u6848\u3002,\u5982\u4eca\u5df2\u7ecf\u6709\u4e0d\u5c11\u5382\u5546\u5f00\u59cb\u53e6\u8f9f\u8e4a\u5f84\uff0c\u5728GPU\u4e4b\u5916\u7684\u9053\u8def\u4e0a\u63a2\u7d22\u5e76\u7814\u53d1\u66f4\u4e3a\u4e13\u4e1a\u5316\u3001\u7cbe\u7ec6\u5316\u7684\u8ba1\u7b97\u8bbe\u5907\u548c\u89e3\u51b3\u65b9\u6848\u3002\u90a3\u4e48\u672a\u6765\u7684AI\u52a0\u901f\u82af\u7247\u53c8\u5c06\u5982\u4f55\u53d1\u5c55\uff1f,
\n03 <\/strong>
\n\u79d1\u6280\u5de8\u5934\u53e6\u8f9f\u8e4a\u5f84<\/strong><\/strong><\/strong>,\u5728\u5f53\u4e0b\u8fd9\u4e2a\u79d1\u6280\u53d1\u5c55\u6781\u5feb\u3001\u7b97\u6cd5\u4ee5\u6708\u4e3a\u5355\u4f4d\u66f4\u8fed\u7684\u5927\u6570\u636e\u65f6\u4ee3\uff0cGPU \u786e\u5b9e\u9002\u5408\u66f4\u591a\u4eba\uff1b\u4f46\u662f\u4e00\u65e6\u672a\u6765\u7684\u5546\u4e1a\u9700\u6c42\u56fa\u5b9a\u4e0b\u6765\uff0cFPGA \u751a\u81f3 ASIC \u5219\u4f1a\u6210\u4e3a\u66f4\u597d\u7684\u5e95\u5c42\u8ba1\u7b97\u8bbe\u5907\u3002,\u5404\u82af\u7247\u9f99\u5934\u548c\u79d1\u6280\u9f99\u5934\u4e5f\u65e9\u5df2\u5f00\u59cb\u7814\u53d1\u751f\u4ea7\u4e13\u7528\u4e8e\u6df1\u5ea6\u5b66\u4e60\u3001DNN \u7684\u8fd0\u7b97\u82af\u7247\u6216\u57fa\u4e8e FPGA \u67b6\u6784\u7684\u534a\u5b9a\u5236\u82af\u7247\uff0c\u4ee3\u8868\u4ea7\u54c1\u6709 Google \u7814\u53d1\u7684\u5f20\u91cf\u8ba1\u7b97\u5904\u7406\u5668 TPU\u3001 Intel \u65d7\u4e0b\u7684 Altera Stratix V FPGA\u7b49\u3002,
\nGoogle\u62bc\u6ce8\u5b9a\u5236\u5316\u7684 ASIC \u82af\u7247\uff1aTPU<\/strong>,Google \u65e9\u5728 2013 \u5e74\u5c31\u79d8\u5bc6\u7814\u53d1\u4e13\u6ce8 AI\u673a\u5668\u5b66\u4e60\u7b97\u6cd5\u82af\u7247\uff0c\u5e76\u7528\u4e8e\u4e91\u8ba1\u7b97\u6570\u636e\u4e2d\u5fc3\uff0c\u53d6\u4ee3\u82f1\u4f1f\u8fbe GPU\u3002,\u8fd9\u6b3eTPU\u81ea\u7814\u82af\u72472016\u5e74\u516c\u5f00\uff0c\u4e3a\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u6267\u884c\u5927\u89c4\u6a21\u77e9\u9635\u8fd0\u7b97\uff0c\u5982\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u8ba1\u7b97\u673a\u89c6\u89c9\u548c\u63a8\u8350\u7cfb\u7edf\u6a21\u578b\u3002Google \u5176\u5b9e\u5728 2020 \u5e74\u7684\u8d44\u6599\u4e2d\u5fc3\u4fbf\u5efa\u6784 AI \u82af\u7247 TPU v4\uff0c\u76f4\u5230 2023 \u5e74 4 \u6708\u624d\u9996\u6b21\u516c\u5f00\u7ec6\u8282\u3002,\u503c\u5f97\u6ce8\u610f\u7684\u662fTPU\u662f\u4e00\u79cd\u5b9a\u5236\u5316\u7684 ASIC \u82af\u7247\uff0c\u5b83\u7531\u8c37\u6b4c\u4ece\u5934\u8bbe\u8ba1\uff0c\u5e76\u4e13\u95e8\u7528\u4e8e\u673a\u5668\u5b66\u4e60\u5de5\u4f5c\u8d1f\u8f7d\u3002,2023\u5e7412\u67086\u65e5\uff0c\u8c37\u6b4c\u5b98\u5ba3\u4e86\u5168\u65b0\u7684\u591a\u6a21\u6001\u5927\u6a21\u578bGemini\uff0c\u5305\u542b\u4e86\u4e09\u4e2a\u7248\u672c\uff0c\u6839\u636e\u8c37\u6b4c\u7684\u57fa\u51c6\u6d4b\u8bd5\u7ed3\u679c\uff0c\u5176\u4e2d\u7684Gemini Ultra\u7248\u672c\u5728\u8bb8\u591a\u6d4b\u8bd5\u4e2d\u90fd\u8868\u73b0\u51fa\u4e86\u201c*\u8fdb\u7684\u6027\u80fd\u201d\uff0c\u751a\u81f3\u5728\u5927\u90e8\u5206\u6d4b\u8bd5\u4e2d\u5b8c\u5168\u51fb\u8d25\u4e86OpenAI\u7684GPT-4\u3002,\u800c\u5728Gemini\u51fa\u5c3d\u4e86\u98ce\u5934\u7684\u540c\u65f6\uff0c\u8c37\u6b4c\u8fd8\u4e22\u51fa\u4e86\u53e6\u4e00\u4e2a\u91cd\u78c5\u70b8\u5f39\u2014\u2014\u5168\u65b0\u7684\u81ea\u7814\u82af\u7247TPU v5p\uff0c\u5b83\u4e5f\u662f\u8fc4\u4eca\u4e3a\u6b62\u529f\u80fd\u6700\u5f3a\u5927\u7684TPU\u3002\u6839\u636e\u5b98\u65b9\u63d0\u4f9b\u7684\u6570\u636e\uff0c\u6bcf\u4e2aTPU v5p pod\u5728\u4e09\u7ef4\u73af\u5f62\u62d3\u6251\u7ed3\u6784\u4e2d\uff0c\u901a\u8fc7\u6700\u9ad8\u5e26\u5bbd\u7684\u82af\u7247\u95f4\u4e92\u8054\uff08ICI\uff09\uff0c\u4ee54800 Gbps\/chip\u7684\u901f\u5ea6\u5c068960\u4e2a\u82af\u7247\u7ec4\u5408\u5728\u4e00\u8d77\uff0c\u4e0eTPU v4\u76f8\u6bd4\uff0cTPU v5p\u7684FLOPS\u548c\u9ad8\u5e26\u5bbd\u5185\u5b58\uff08HBM\uff09\u5206\u522b\u63d0\u9ad8\u4e862\u500d\u548c3\u500d\u3002,\u968f\u540e\u5728\u4eca\u5e745\u6708\uff0c\u8c37\u6b4c\u53c8\u5ba3\u5e03\u4e86\u7b2c\u516d\u4ee3\u6570\u636e\u4e2d\u5fc3 AI \u82af\u7247 Tensor \u5904\u7406\u5668\u5355\u5143–Trillium\uff0c\u5e76\u8868\u793a\u5c06\u4e8e\u4eca\u5e74\u665a\u4e9b\u65f6\u5019\u63a8\u51fa\u4ea4\u4ed8\u3002\u8c37\u6b4c\u8868\u793a\uff0c\u7b2c\u516d\u4ee3Trillium\u82af\u7247\u7684\u8ba1\u7b97\u6027\u80fd\u6bd4TPU v5e\u82af\u7247\u63d0\u9ad84.7\u500d\uff0c\u80fd\u6548\u6bd4v5e\u9ad8\u51fa67%\u3002\u8fd9\u6b3e\u82af\u7247\u65e8\u5728\u4e3a\u4ece\u5927\u6a21\u578b\u4e2d\u751f\u6210\u6587\u672c\u548c\u5176\u4ed6\u5185\u5bb9\u7684\u6280\u672f\u63d0\u4f9b\u52a8\u529b\u3002\u8c37\u6b4c\u8fd8\u8868\u793a\uff0c\u7b2c\u516d\u4ee3Trillium\u82af\u7247\u5c06\u5728\u4eca\u5e74\u5e74\u5e95\u53ef\u4f9b\u5176\u4e91\u5ba2\u6237\u4f7f\u7528\u3002,\u636e\u6089\uff0c\u82f1\u4f1f\u8fbe\u5728AI\u82af\u7247\u5e02\u573a\u7684\u5e02\u5360\u9ad8\u8fbe80%\u5de6\u53f3\uff0c\u5176\u4f5920%\u7684\u7edd\u5927\u90e8\u5206\u7531\u5404\u79cd\u7248\u672c\u7684\u8c37\u6b4cTPU\u6240\u63a7\u5236\u3002\u8c37\u6b4c\u81ea\u8eab\u4e0d\u51fa\u552e\u82af\u7247\uff0c\u800c\u662f\u901a\u8fc7\u5176\u4e91\u8ba1\u7b97\u5e73\u53f0\u79df\u7528\u8bbf\u95ee\u6743\u9650\u3002,
\n\u5fae\u8f6f\uff1a\u63a8\u51fa\u57fa\u4e8eArm\u67b6\u6784\u7684\u901a\u7528\u578b\u82af\u7247Cobalt\u3001ASIC\u82af\u7247Maia 100<\/strong>,2023\u5e7411\u6708\uff0c\u5fae\u8f6f\u5728Ignite\u6280\u672f\u5927\u4f1a\u4e0a\u53d1\u5e03\u4e86*\u81ea\u5bb6\u7814\u53d1\u7684AI\u82af\u7247Azure Maia 100\uff0c\u4ee5\u53ca\u5e94\u7528\u4e8e\u4e91\u7aef\u8f6f\u4ef6\u670d\u52a1\u7684\u82af\u7247Azure Cobalt\u3002\u4e24\u6b3e\u82af\u7247\u5c06\u7531\u53f0\u79ef\u7535\u4ee3\u5de5\uff0c\u91c7\u75285nm\u5236\u7a0b\u6280\u672f\u3002,\u636e\u6089\uff0c\u82f1\u4f1f\u8fbe\u7684\u9ad8\u7aef\u4ea7\u54c1\u4e00\u9897\u6709\u65f6\u53ef\u5356\u52303\u4e07\u52304\u4e07\u7f8e\u5143\uff0c\u7528\u4e8eChatGPT\u7684\u82af\u7247\u88ab\u8ba4\u4e3a\u5927\u6982\u5c31\u9700\u8981\u67091\u4e07\u9897\uff0c\u8fd9\u5bf9AI\u516c\u53f8\u662f\u4e2a\u5e9e\u5927\u6210\u672c\u3002\u6709\u5927\u91cfAI\u82af\u7247\u9700\u6c42\u7684\u79d1\u6280\u5927\u5382\u6781\u529b\u5bfb\u6c42\u53ef\u66ff\u4ee3\u7684\u4f9b\u5e94\u6765\u6e90\uff0c\u5fae\u8f6f\u9009\u62e9\u81ea\u884c\u7814\u53d1\uff0c\u4fbf\u662f\u5e0c\u671b\u589e\u5f3aChatGPT\u7b49\u751f\u6210\u5f0fAI\u4ea7\u54c1\u7684\u6027\u80fd\uff0c\u540c\u65f6\u964d\u4f4e\u6210\u672c\u3002,Cobalt\u662f\u57fa\u4e8eArm\u67b6\u6784\u7684\u901a\u7528\u578b\u82af\u7247\uff0c\u5177\u6709128\u4e2a\u6838\u5fc3\uff0cMaia 100\u662f\u4e00\u6b3e\u4e13\u4e3a Azure \u4e91\u670d\u52a1\u548c AI \u5de5\u4f5c\u8d1f\u8f7d\u8bbe\u8ba1\u7684 ASIC \u82af\u7247\uff0c\u7528\u4e8e\u4e91\u7aef\u8bad\u7ec3\u548c\u63a8\u7406\u7684\uff0c\u6676\u4f53\u7ba1\u6570\u91cf\u8fbe\u52301050\u4ebf\u4e2a\u3002\u8fd9\u4e24\u6b3e\u82af\u7247\u5c06\u5bfc\u5165\u5fae\u8f6fAzure\u6570\u636e\u4e2d\u5fc3\uff0c\u652f\u6301OpenAI\u3001Copilot\u7b49\u670d\u52a1\u3002,\u8d1f\u8d23Azure\u82af\u7247\u90e8\u95e8\u7684\u526f\u603b\u88c1Rani Borkar\u8868\u793a\uff0c\u5fae\u8f6f\u5df2\u5f00\u59cb\u7528Bing\u548cOffice AI\u4ea7\u54c1\u6d4b\u8bd5Maia 100\u82af\u7247\uff0c\u5fae\u8f6f\u4e3b\u8981AI\u5408\u4f5c\u4f19\u4f34\u3001ChatGPT\u5f00\u53d1\u5546OpenAI\uff0c\u4e5f\u5728\u8fdb\u884c\u6d4b\u8bd5\u4e2d\u3002\u6709\u5e02\u573a\u8bc4\u8bba\u8ba4\u4e3a\uff0c\u5fae\u8f6f AI \u82af\u7247\u7acb\u9879\u7684\u65f6\u673a\u5f88\u5de7\uff0c\u6b63\u597d\u5728\u5fae\u8f6f\u3001OpenAI \u7b49\u516c\u53f8\u57f9\u517b\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\u5df2\u7ecf\u5f00\u59cb\u817e\u98de\u4e4b\u9645\u3002,\u4e0d\u8fc7\uff0c\u5fae\u8f6f\u5e76\u4e0d\u8ba4\u4e3a\u81ea\u5df1\u7684 AI \u82af\u7247\u53ef\u4ee5\u5e7f\u6cdb\u66ff\u4ee3\u82f1\u4f1f\u8fbe\u7684\u4ea7\u54c1\u3002\u6709\u5206\u6790\u8ba4\u4e3a\uff0c\u5fae\u8f6f\u7684\u8fd9\u4e00\u52aa\u529b\u5982\u679c\u6210\u529f\u7684\u8bdd\uff0c\u4e5f\u6709\u53ef\u80fd\u5e2e\u52a9\u5b83\u5728\u672a\u6765\u4e0e\u82f1\u4f1f\u8fbe\u7684\u8c08\u5224\u4e2d\u66f4\u5177\u4f18\u52bf\u3002,\u636e\u6089\uff0c\u5fae\u8f6f\u6709\u671b\u5728\u5373\u5c06\u5230\u6765\u7684Build\u6280\u672f\u5927\u4f1a\u4e0a\u53d1\u5e03\u4e00\u7cfb\u5217\u4e91\u7aef\u8f6f\u786c\u4ef6\u6280\u672f\u65b0\u8fdb\u5c55\u3002\u800c\u5907\u53d7\u5173\u6ce8\u7684\u662f\uff0c\u5fae\u8f6f\u5c06\u5411Azure\u7528\u6237\u5f00\u653e\u5176\u81ea\u7814\u7684AI\u82af\u7247Cobalt 100\u7684\u4f7f\u7528\u6743\u9650\u3002,
\n\u82f1\u7279\u5c14\u62bc\u6ce8FPGA\u82af\u7247<\/strong>,\u82f1\u7279\u5c14\u8868\u793a\uff0c\u65e9\u671f\u7684\u4eba\u5de5\u667a\u80fd\u5de5\u4f5c\u8d1f\u8f7d\uff0c\u6bd4\u5982\u56fe\u50cf\u8bc6\u522b\uff0c\u5f88\u5927\u7a0b\u5ea6\u4e0a\u4f9d\u8d56\u4e8e\u5e76\u884c\u6027\u80fd\u3002\u56e0\u4e3a GPU \u662f\u4e13\u95e8\u9488\u5bf9\u89c6\u9891\u548c\u663e\u5361\u8bbe\u8ba1\u7684\uff0c\u56e0\u6b64\uff0c\u5c06\u5176\u5e94\u7528\u4e8e\u673a\u5668\u5b66\u4e60\u548c\u6df1\u5ea6\u5b66\u4e60\u53d8\u5f97\u5f88\u666e\u904d\u3002GPU \u5728\u5e76\u884c\u5904\u7406\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u5e76\u884c\u6267\u884c\u5927\u91cf\u8ba1\u7b97\u64cd\u4f5c\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5982\u679c\u5fc5\u987b\u591a\u6b21\u5feb\u901f\u6267\u884c\u540c\u4e00\u5de5\u4f5c\u8d1f\u8f7d\uff0c\u5b83\u4eec\u53ef\u4ee5\u5b9e\u73b0\u4ee4\u4eba\u96be\u4ee5\u7f6e\u4fe1\u7684\u901f\u5ea6\u63d0\u9ad8\u3002,\u4f46\u662f\uff0c\u5728 GPU \u4e0a\u8fd0\u884c\u4eba\u5de5\u667a\u80fd\u662f\u5b58\u5728\u5c40\u9650\u7684\u3002GPU \u4e0d\u80fd\u591f\u63d0\u4f9b\u4e0e ASIC \u76f8\u5ab2\u7f8e\u7684\u6027\u80fd\uff0c\u540e\u8005\u662f\u4e00\u79cd\u9488\u5bf9\u7ed9\u5b9a\u7684\u6df1\u5ea6\u5b66\u4e60\u5de5\u4f5c\u8d1f\u8f7d\u4e13\u95e8\u6784\u5efa\u7684\u82af\u7247\u3002,\u800c FPGA \u5219\u80fd\u591f\u501f\u52a9\u96c6\u6210\u7684\u4eba\u5de5\u667a\u80fd\u63d0\u4f9b\u786c\u4ef6\u5b9a\u5236\uff0c\u5e76\u4e14\u53ef\u4ee5\u901a\u8fc7\u7f16\u7a0b\u63d0\u4f9b\u4e0e GPU \u6216 ASIC \u76f8\u7c7b\u4f3c\u7684\u5de5\u4f5c\u65b9\u5f0f\u3002FPGA \u53ef\u91cd\u65b0\u7f16\u7a0b\u3001\u91cd\u65b0\u914d\u7f6e\u7684\u6027\u8d28\u4f7f\u5176\u683c\u5916\u9002\u5408\u5e94\u7528\u4e8e\u98de\u901f\u6f14\u53d8\u7684\u4eba\u5de5\u667a\u80fd\u9886\u57df\uff0c\u8fd9\u6837\uff0c\u8bbe\u8ba1\u4eba\u5458\u5c31\u80fd\u591f\u5feb\u901f\u6d4b\u8bd5\u7b97\u6cd5\uff0c\u5e76\u5c06\u4ea7\u54c1\u52a0\u901f\u63a8\u5411\u5e02\u573a\u3002,\u82f1\u7279\u5c14FPGA \u5bb6\u65cf\u5305\u62ec\u82f1\u7279\u5c14 Cyclone 10 GX FPGA\u3001\u82f1\u7279\u5c14 Arria 10 GX FPGA \u548c\u82f1\u7279\u5c14Stratix 10 GX FPGA\u7b49\u3002\u8fd9\u4e9b\u4ea7\u54c1\u5177\u5907 I\/O \u7075\u6d3b\u6027\u3001\u4f4e\u529f\u8017\uff08\u6216\u6bcf\u6b21\u63a8\u7406\u7684\u80fd\u8017\uff09\u548c\u4f4e\u65f6\u5ef6\uff0c\u672c\u5c31\u53ef\u5728 AI \u63a8\u7406\u4e0a\u5e26\u6765\u4f18\u52bf\u3002\u8fd9\u4e9b\u4f18\u52bf\u5728\u4e09\u4e2a\u5168\u65b0\u7684\u82f1\u7279\u5c14 FPGA \u548c\u7247\u4e0a\u7cfb\u7edf\u5bb6\u65cf\u7684\u4ea7\u54c1\u4e2d\u53c8\u5f97\u5230\u4e86\u8865\u5145\uff0c\u4f7f\u5f97 AI \u63a8\u7406\u6027\u80fd\u8fdb\u4e00\u6b65\u83b7\u5f97\u4e86\u663e\u8457\u63d0\u5347\u3002\u8fd9\u4e09\u4e2a\u5bb6\u65cf\u5206\u522b\u662f\u82f1\u7279\u5c14 Stratix 10 NX FPGA \u4ee5\u53ca\u82f1\u7279\u5c14 Agilex FPGA \u5bb6\u65cf\u7684\u65b0\u6210\u5458\uff1a\u82f1\u7279\u5c14 Agilex D \u7cfb\u5217 FPGA\uff0c\u548c\u4ee3\u53f7\u4e3a\u201cSundance Mesa\u201d\u7684\u5168\u65b0\u82f1\u7279\u5c14 Agilex \u8bbe\u5907\u5bb6\u65cf\u3002\u8fd9\u4e9b\u82f1\u7279\u5c14FPGA \u548c SoC \u5bb6\u65cf\u5305\u542b\u4e13\u95e8\u9762\u5411\u5f20\u91cf\u6570\u5b66\u8fd0\u7b97\u4f18\u5316\u7684\u4e13\u7528 DSP \u6a21\u5757\uff0c\u4e3a\u52a0\u901f AI \u8ba1\u7b97\u5960\u5b9a\u4e86\u57fa\u7840\u3002,\u4eca\u5e743\u6708\uff0c\u82af\u7247\u5de8\u5934\u82f1\u7279\u5c14\u5ba3\u5e03\u6210\u7acb\u5168\u65b0\u72ec\u7acb\u8fd0\u8425\u7684FPGA\u516c\u53f8\u2014\u2014Altera\u3002\u82f1\u7279\u5c14\u57282015\u5e746\u6708\u4ee5167\u4ebf\u7f8e\u5143\u6536\u8d2dAltera\uff0c\u88ab\u6536\u8d2d\u65f6Altera\u662f\u5168\u7403\u7b2c\u4e8c\u5927FPGA\u516c\u53f8\uff0c\u4e5d\u5e74\u540e\u82f1\u7279\u5c14\u51b3\u5b9a\u8ba9FPGA\u4e1a\u52a1\u72ec\u7acb\u8fd0\u8425\uff0c\u518d\u6b21\u9009\u62e9\u4ee5Altera\u547d\u540d\u3002,NPU(Neural Processing Unit)\u4e5f\u662f\u4e00\u79cd\u53c2\u8003\u4eba\u4f53\u795e\u7ecf\u7a81\u89e6\u7684 ASIC \u82af\u7247\u3002\u968f\u7740\u6df1\u5ea6\u5b66\u4e60\u795e\u7ecf\u7f51\u7edc\u7684\u5174\u8d77\uff0cCPU\u548c GPU \u9010\u6e10\u96be\u4ee5\u6ee1\u8db3\u6df1\u5ea6\u5b66\u4e60\u7684\u9700\u8981\uff0c\u4e13\u95e8\u7528\u4e8e\u795e\u7ecf\u7f51\u7edc\u6df1\u5ea6\u5b66\u4e60\u7684\u5904\u7406\u5668NPU\u5e94\u8fd0\u800c\u751f\u3002NPU \u91c7\u7528\u201c\u6570\u636e\u9a71\u52a8\u5e76\u884c\u8ba1\u7b97\u201d\u7684\u67b6\u6784\uff0c\u7279\u522b\u64c5\u957f\u5904\u7406\u89c6\u9891\u3001\u56fe\u50cf\u7c7b\u7684\u6d77\u91cf\u591a\u5a92\u4f53\u6570\u636e\u3002\u533a\u522b\u4e8e CPU \u4ee5\u53ca GPU \u6240\u9075\u5faa\u7684\u51af\u8bfa\u4f9d\u66fc\u67b6\u6784\uff0cNPU \u53c2\u8003\u4eba\u4f53\u7684\u795e\u7ecf\u7a81\u89e6\u7ed3\u6784\uff0c\u5c06\u5b58\u50a8\u4e0e\u8fd0\u7b97\u7ed3\u4e3a\u4e00\u4f53\u3002,
\nArm \u8fd1\u65e5\u5ba3\u5e03\u63a8\u51fa Ethos-U85 NPU\u3002<\/strong>\u4f5c\u4e3a Arm \u9762\u5411\u8fb9\u7f18 AI \u7684\u7b2c\u4e09\u4ee3 NPU \u4ea7\u54c1\uff0cEthos-U85 \u9002\u7528\u4e8e\u5de5\u4e1a\u81ea\u52a8\u5316\u548c\u89c6\u9891\u76d1\u63a7\u7b49\u573a\u666f\uff0c\u5728\u6027\u80fd\u65b9\u9762\u63d0\u5347\u4e86\u56db\u500d\u3002Ethos-U85 \u8f83\u4e0a\u4e00\u4ee3\u4ea7\u54c1\u5728\u80fd\u6548\u65b9\u9762\u62e5\u6709 20% \u7684\u63d0\u5347\uff0c\u8fd8\u53ef\u5728\u5e38\u7528\u795e\u7ecf\u7f51\u7edc\u4e0a\u5b9e\u73b0 85% \u7684\u5229\u7528\u7387\u3002\u5176\u5728\u8bbe\u8ba1\u4e0a\u9002\u5408\u57fa\u4e8e Arm Cortex-M \/ A \u5904\u7406\u5668\u5185\u6838\u7684\u7cfb\u7edf\uff0c\u80fd\u63a5\u53d7\u8f83\u9ad8\u7684\u5185\u5b58\u5ef6\u8fdf\u3002,\u534f\u540c\u3001\u8bad\u7ec3\u63a8\u7406\u878d\u5408\u3001\u5177\u5907\u7edf\u4e00\u751f\u6001\u7684\u7cfb\u5217\u5316\u667a\u80fd\u82af\u7247\u4ea7\u54c1\u548c\u5e73\u53f0\u5316\u57fa\u7840\u7cfb\u7edf\u8f6f\u4ef6\u3002\u5bd2\u6b66\u7eaa\u4ea7\u54c1\u5e7f\u6cdb\u5e94\u7528\u4e8e\u670d\u52a1\u5668\u5382\u5546\u548c\u4ea7\u4e1a\u516c\u53f8\uff0c\u9762\u5411\u4e92\u8054\u7f51\u3001\u91d1\u878d\u3001\u4ea4\u901a\u3001\u80fd\u6e90\u3001\u7535\u529b\u548c\u5236\u9020\u7b49\u3002,\u6b64\u5916\uff0c
\nOpenAI\u4e5f\u6b63\u5728\u63a2\u7d22\u81ea\u7814AI\u82af\u7247<\/strong>\uff0c\u540c\u65f6\u5f00\u59cb\u8bc4\u4f30\u6f5c\u5728\u6536\u8d2d\u76ee\u6807\u3002AWS\u81ea\u7814AI\u82af\u7247\u9635\u5bb9\u5305\u62ec\u63a8\u7406\u82af\u7247Inferentia\u548c\u8bad\u7ec3\u82af\u7247Trainium\u3002 \u7535\u52a8\u6c7d\u8f66\u5236\u9020\u5546\u7279\u65af\u62c9\u4e5f\u79ef\u6781\u53c2\u4e0eAI\u52a0\u901f\u5668\u82af\u7247\u7684\u5f00\u53d1\u3002\u7279\u65af\u62c9\u4e3b\u8981\u56f4\u7ed5\u81ea\u52a8\u9a7e\u9a76\u9700\u6c42\uff0c\u8fc4\u4eca\u4e3a\u6b62\u63a8\u51fa\u4e86\u4e24\u6b3eAI\u82af\u7247\uff1a\u5168\u81ea\u52a8\u9a7e\u9a76\uff08FSD\uff09\u82af\u7247\u548cDojo D1\u82af\u7247\u3002,\u53bb\u5e745\u6708
\nMeta\u62ab\u9732\u4e86\u65d7\u4e0b\u6570\u636e\u4e2d\u5fc3\u9879\u76ee\u652f\u6301AI\u5de5\u4f5c\u7684\u7ec6\u8282<\/strong>\uff0c\u63d0\u5230\u5df2\u7ecf\u6253\u9020\u4e00\u6b3e\u5b9a\u5236\u82af\u7247\uff0c\u7b80\u79f0MTIA\uff0c\u7528\u4e8e\u52a0\u5feb\u751f\u6210\u5f0fAI\u6a21\u578b\u7684\u8bad\u7ec3\u3002\u8fd9\u662fMeta\u9996\u6b21\u63a8\u51faAI\u5b9a\u5236\u82af\u7247\u3002Meta\u79f0\uff0cMTIA\u662f\u52a0\u5febAI\u8bad\u7ec3\u548c\u63a8\u7406\u5de5\u4f5c\u8d1f\u8f7d\u7684\u82af\u7247\u201c\u5bb6\u65cf\u201d\u7684\u4e00\u5206\u5b50\u3002\u6b64\u5916\uff0cMeta\u4ecb\u7ecd\uff0cMTIA\u91c7\u7528\u5f00\u6e90\u82af\u7247\u67b6\u6784RISC-V\uff0c\u5b83\u7684\u529f\u8017\u4ec5\u670925\u74e6\uff0c\u8fdc\u4f4e\u4e8e\u82f1\u4f1f\u8fbe\u7b49\u4e3b\u6d41\u82af\u7247\u5382\u5546\u7684\u4ea7\u54c1\u529f\u8017\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4eca\u5e744\u6708\uff0cMeta\u516c\u5e03\u4e86\u81ea\u4e3b\u7814\u53d1\u82af\u7247MTIA\u7684\u6700\u65b0\u7248\u672c\u3002\u5206\u6790\u6307\u51fa\uff0cMeta\u7684\u76ee\u6807\u662f\u964d\u4f4e\u5bf9\u82f1\u4f1f\u8fbe\u7b49\u82af\u7247\u5382\u5546\u7684\u4f9d\u8d56\u3002,\u3010\u672c\u6587\u7531\u5408\u4f5c\u4f19\u4f34\u5fae\u4fe1\u516c\u4f17\u53f7\uff1a\u534a\u5bfc\u4f53\u4ea7\u4e1a\u7eb5\u6a2a\u6388\u6743\u53d1\u5e03\uff0c\u672c\u5e73\u53f0\u4ec5\u63d0\u4f9b\u4fe1\u606f\u5b58\u50a8\u670d\u52a1\u3002\u3011\u5982\u6709\u4efb\u4f55\u7591\u95ee\uff0c\u8bf7\u8054\u7cfb\u5904\u7406\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u5728\u4eba\u5de5\u667a\u80fd\u8ba1\u7b97\u67b6\u6784\u7684\u5e03\u5c40\u4e2d\uff0cCPU\u4e0e\u52a0\u901f\u82af\u7247\u534f\u540c\u5de5\u4f5c\u7684\u6a21\u5f0f\u5df2\u6210\u4e3a\u4e00\u79cd\u5178\u578b\u7684AI\u90e8\u7f72\u65b9\u6848\u3002CPU\u626e\u6f14\u57fa\u7840\u7b97\u529b\u7684\u63d0…<\/p>\n","protected":false},"author":169,"featured_media":4087,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-4086","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news"],"_links":{"self":[{"href":"http:\/\/www.yuntou8.com\/index.php?rest_route=\/wp\/v2\/posts\/4086","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.yuntou8.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.yuntou8.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.yuntou8.com\/index.php?rest_route=\/wp\/v2\/users\/169"}],"replies":[{"embeddable":true,"href":"http:\/\/www.yuntou8.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4086"}],"version-history":[{"count":0,"href":"http:\/\/www.yuntou8.com\/index.php?rest_route=\/wp\/v2\/posts\/4086\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.yuntou8.com\/index.php?rest_route=\/wp\/v2\/media\/4087"}],"wp:attachment":[{"href":"http:\/\/www.yuntou8.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4086"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.yuntou8.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4086"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.yuntou8.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4086"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}