\u8981\u7406\u89e3GPU\u70ba\u4f55\u80fd\u6210\u70ba\u300c\u904a\u6232\u898f\u5247\u6539\u8b8a\u8005\u300d\uff0c\u9996\u5148\u9700\u5c0d\u6bd4\u5176\u8207\u50b3\u7d71CPU\u7684\u8a2d\u8a08\u5dee\u7570\u3002CPU\u5c08\u70ba\u5e8f\u5217\u8655\u7406\u6700\u4f73\u5316\uff0c\u50c5\u914d\u5099\u5c11\u91cf\u9ad8\u983b\u6838\u5fc3\uff0c\u64c5\u9577\u8655\u7406\u8907\u96dc\u63a7\u5236\u908f\u8f2f\uff1b\u800cGPU\u5247\u4ee5\u5e73\u884c\u70ba\u6838\u5fc3\u8a2d\u8a08\uff0c\u6574\u5408\u6578\u5343\u500b\u4f4e\u983b\u6838\u5fc3\uff0c\u80fd\u9ad8\u6548\u57f7\u884c\u300c\u5c0d\u591a\u7d44\u6578\u64da\u91cd\u8907\u540c\u4e00\u6307\u4ee4\u300d\u7684\u904b\u7b97\u3002\u9019\u7a2e\u7279\u6027\u4f7f\u5176\u6210\u70ba\u73fe\u4ee3\u641c\u5c0b\u5f15\u64ce\u4e2d\u300c\u904b\u7b97\u5bc6\u96c6\u578b\u3001\u6578\u64da\u5e73\u884c\u5316\u300d\u4efb\u52d9\u7684\u7406\u60f3\u9078\u64c7\u3002<\/p>\n

\u4ee5\u4e0b\u662f\u5169\u8005\u7684\u67b6\u69cb\u5c0d\u6bd4\uff1a<\/p>\n\n\n\n\n\n\n\n

\u6307\u6a19<\/th>\n	CPU<\/th>\n	GPU<\/th>\n<\/tr>\n
\u6838\u5fc3\u6578\u91cf<\/td>\n	8\u201364<\/td>\n	2,000\u20134,000+<\/td>\n<\/tr>\n
\u6642\u8108\u983b\u7387<\/td>\n	3\u20135 GHz<\/td>\n	1\u20132 GHz<\/td>\n<\/tr>\n
\u8a18\u61b6\u9ad4\u983b\u5bec<\/td>\n	50\u2013100 GB\/s<\/td>\n	400\u20131,000 GB\/s<\/td>\n<\/tr>\n
\u64c5\u9577\u7684\u4efb\u52d9<\/td>\n	\u8907\u96dc\u908f\u8f2f\u63a7\u5236<\/td>\n	\u5927\u898f\u6a21\u5e73\u884c\u904b\u7b97<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n [\/vc_column_text][\/vc_column][\/vc_row][vc_row el_class=”blog-detail-section”][vc_column][vc_column_text css=””]<\/p>\n GPU\u9a45\u52d5\u7684\u641c\u5c0b\u5f15\u64ce\u904b\u4f5c\u6838\u5fc3\u968e\u6bb5<\/strong><\/h2>\n \u641c\u5c0b\u5f15\u64ce\u7684\u904b\u4f5c\u5305\u542b\u64f7\u53d6\u3001\u7d22\u5f15\u3001\u6392\u5e8f\u3001\u6a5f\u5668\u5b78\u7fd2\u56db\u5927\u6838\u5fc3\u968e\u6bb5\uff0cGPU\u900f\u904e\u5c08\u9580\u7684\u5e73\u884c\u5316\u7b56\u7565\uff0c\u70ba\u6bcf\u500b\u968e\u6bb5\u63d0\u4f9b\u6548\u80fd\u63d0\u5347\u3002<\/p>\n 1. \u7db2\u9801\u64f7\u53d6\uff1a\u6578\u64da\u63a1\u96c6\u7684\u5e73\u884c\u5316\u5347\u7d1a<\/strong><\/h3>\n \u64f7\u53d6\u968e\u6bb5\u9700\u5f9e\u7db2\u969b\u7db2\u8def\u53d6\u5f97\u7db2\u9801\u6578\u64da\uff0c\u4f46\u53cd\u722c\u87f2\u6a5f\u5236\u7684\u9650\u5236\u8207\u6d77\u91cfURL\u7684\u8655\u7406\u58d3\u529b\uff0c\u4e00\u76f4\u662f\u8a72\u968e\u6bb5\u7684\u74f6\u9838\u3002GPU\u900f\u904e\u5206\u6563\u5f0f\u5e73\u884cURL\u8655\u7406\uff0c\u5fb9\u5e95\u6539\u8b8a\u4e86\u9019\u4e00\u5c40\u9762\uff1a<\/p>\n \n \u57fa\u65bc\u53e2\u96c6\u7684\u4efb\u52d9\u6392\u7a0b\uff1aGPU\u53e2\u96c6\u5e73\u884c\u7ba1\u7406\u6578\u767e\u842c\u500bURL\uff0c\u5c07\u64f7\u53d6\u3001\u6e32\u67d3\u3001\u89e3\u6790\u4efb\u52d9\u5206\u914d\u5230\u6578\u5343\u500b\u6838\u5fc3\u4e0a\u540c\u6b65\u57f7\u884c\u3002<\/li>\n JavaScript\u6e32\u67d3\u52a0\u901f\uff1a\u73fe\u4ee3\u7db2\u7ad9\u5927\u91cf\u4f9d\u8cf4\u7528\u6236\u7aef\u6307\u4ee4\u78bc\uff0c\u50b3\u7d71\u722c\u87f2\u96e3\u4ee5\u9ad8\u6548\u8655\u7406\uff0c\u800cGPU\u53ef\u5378\u8f09JS\u57f7\u884c\u4efb\u52d9\uff0c\u5728\u5be6\u969b\u5834\u666f\u4e2d\u80fd\u5c07\u6e32\u67d3\u5ef6\u9072\u964d\u4f4e40%\u3002<\/li>\n \u7db2\u8defI\/O\u6700\u4f73\u5316\uff1a\u5728\u9999\u6e2f\u4f3a\u670d\u5668\u79df\u7528\u9019\u985e\u9ad8\u983b\u5bec\u74b0\u5883\u4e2d\uff0cGPU\u8207\u4f4e\u5ef6\u9072\u7db2\u8def\u5354\u540c\uff0c\u6700\u5927\u7a0b\u5ea6\u6e1b\u5c11\u722c\u87f2\u7bc0\u9ede\u9593\u7684\u6578\u64da\u50b3\u8f38\u958b\u92b7\u3002<\/li>\n<\/ul>\n 2. \u7d22\u5f15\u5efa\u69cb\uff1a\u5927\u898f\u6a21\u5efa\u7f6e\u9ad8\u6548\u80fd\u6578\u64da\u7d50\u69cb<\/strong><\/h3>\n \u7d22\u5f15\u968e\u6bb5\u9700\u5c07\u64f7\u53d6\u5230\u7684\u6578\u64da\u8f49\u63db\u70ba\u53ef\u641c\u5c0b\u7d50\u69cb\uff0c\u904e\u7a0b\u4e2d\u6d89\u53ca\u5927\u91cf\u6587\u5b57\u5206\u6790\u8207\u6578\u64da\u8f49\u63db\u64cd\u4f5c\u3002GPU\u6191\u85c9\u5e73\u884c\u8655\u7406\u80fd\u529b\uff0c\u5728\u81ea\u7136\u8a9e\u8a00\u8655\u7406\uff08NLP\uff09\u8207\u5012\u6392\u7d22\u5f15\u5efa\u69cb\u4e2d\u8868\u73fe\u7a81\u51fa\uff1a<\/p>\n \n NLP\u7ba1\u7dda\u52a0\u901f\uff1a\u5206\u8a5e\u3001\u8a5e\u5e79\u63d0\u53d6\u3001\u8a5e\u6027\u6a19\u8a3b\u7b49\u6838\u5fc3NLP\u4efb\u52d9\u5728GPU\u6838\u5fc3\u4e0a\u5e73\u884c\u8655\u7406\uff0c\u6bcf\u79d2\u53ef\u8655\u7406\u6578\u767e\u842c\u500b\u8a5e\u5143\uff08token\uff09\u3002<\/li>\n \u5012\u6392\u7d22\u5f15\u5206\u7247\uff1a\u5c07\u7d22\u5f15\u7247\u6bb5\u5206\u6563\u5f0f\u5132\u5b58\u5230GPU\u8a18\u61b6\u9ad4\u4e2d\uff0c\u641c\u5c0b\u5f15\u64ce\u53ef\u6e1b\u5c11\u78c1\u789fI\/O\u74f6\u9838\u3002\u8a18\u61b6\u9ad4\u983b\u5bec\u512a\u52e2\u80fd\u5be6\u73fe\u300c\u8a5e-\u6587\u4ef6\u77e9\u9663\u300d\u7684\u6b21\u6beb\u79d2\u7d1a\u5b58\u53d6\uff0c\u9019\u5c0d\u5373\u6642\u67e5\u8a62\u8655\u7406\u81f3\u95dc\u91cd\u8981\u3002<\/li>\n \u58d3\u7e2e\u6f14\u7b97\u6cd5\u6700\u4f73\u5316\uff1aGPU\u5c0d\u7d22\u5f15\u6578\u64da\u9032\u884c\u7121\u640d\u58d3\u7e2e\u6700\u4f73\u5316\uff0c\u5728\u78ba\u4fdd\u5132\u5b58\u6548\u7387\u7684\u540c\u6642\uff0c\u517c\u9867\u67e5\u8a62\u57f7\u884c\u6642\u7684\u5feb\u901f\u89e3\u58d3\u9700\u6c42\u3002<\/li>\n<\/ol>\n 3. \u7d50\u679c\u6392\u5e8f\uff1a\u5373\u6642\u76f8\u95dc\u6027\u8a08\u7b97<\/strong><\/h3>\n \u6392\u5e8f\u968e\u6bb5\u9700\u6839\u64da\u67e5\u8a62\u76f8\u95dc\u6027\u6c7a\u5b9a\u7d50\u679c\u9806\u5e8f\uff0c\u5982\u4eca\u8a72\u904e\u7a0b\u5df2\u7531\u6a5f\u5668\u5b78\u7fd2\u6a21\u578b\u4e3b\u5c0e\u3002GPU\u70ba\u8907\u96dc\u6a21\u578b\u63d0\u4f9b\u8fd1\u5373\u6642\u7684\u63a8\u8ad6\u80fd\u529b\uff1a<\/p>\n \u8a5e\u983b-\u53cd\u5411\u6587\u4ef6\u983b\u7387\uff08TF-IDF\uff09\u7b49\u5411\u91cf\u7a7a\u9593\u6a21\u578b\uff0c\u4f9d\u8cf4\u96a8\u6578\u64da\u96c6\u898f\u6a21\u7dda\u6027\u589e\u9577\u7684\u77e9\u9663\u904b\u7b97\u3002GPU\u5229\u7528\u5e73\u884c\u6838\u5fc3\u540c\u6b65\u8a08\u7b97\u9918\u5f26\u76f8\u4f3c\u5ea6\u8207\u76f8\u95dc\u6027\u5206\u6578\uff0c\u5c07\u904b\u7b97\u6642\u9593\u5f9e\u6beb\u79d2\u7d1a\u58d3\u7e2e\u81f3\u5fae\u79d2\u7d1a\u3002\u5c0d\u65bcBERT\u8207Transformer\u7cfb\u5217\u7b49\u6df1\u5ea6\u5b78\u7fd2\u6a21\u578b\uff0cGPU\u652f\u63f4\u6df7\u5408\u7cbe\u5ea6\u8a13\u7df4\u8207\u63a8\u8ad6\uff1a<\/p>\n \n FP16\u6d6e\u9ede\u904b\u7b97\uff1a\u63a1\u7528\u534a\u7cbe\u5ea6\u6d6e\u9ede\u6578\uff0cGPU\u8655\u7406\u795e\u7d93\u7db2\u8def\u5c64\u7684\u901f\u5ea6\u53ef\u9054CPU\u76842\u500d\uff0c\u4e14\u4e0d\u6703\u5c0e\u81f4\u986f\u8457\u7cbe\u5ea6\u640d\u5931\u3002<\/li>\n \u5f35\u91cf\u6838\u5fc3\uff08Tensor Core\uff09\u6700\u4f73\u5316\uff1a\u73fe\u4ee3GPU\u4e2d\u7684\u5c08\u7528\u77e9\u9663\u8655\u7406\u55ae\u5143\uff0c\u53ef\u5c07\u6df1\u5ea6\u5b78\u7fd2\u904b\u7b97\u901f\u5ea6\u63d0\u534750%\uff0c\u9019\u5c0d\u9ad8\u6d41\u91cf\u5834\u666f\u4e0b\u7684\u5373\u6642\u6392\u5e8f\u81f3\u95dc\u91cd\u8981\u3002<\/li>\n<\/ul>\n 4. \u6a5f\u5668\u5b78\u7fd2\uff1a\u8a13\u7df4\u4e0b\u4e00\u4ee3\u6f14\u7b97\u6cd5<\/strong><\/h3>\n \u8a13\u7df4\u5927\u898f\u6a21\u6a21\u578b\u9700\u5c07\u904b\u7b97\u4efb\u52d9\u5206\u6563\u5f0f\u4f48\u7f72\u5230\u6578\u5343\u500b\u7bc0\u9ede\u4e0a\u3002GPU\u7d50\u5408Horovod\u3001TensorFlow\u7b49\u6846\u67b6\uff0c\u53ef\u5be6\u73fe\u9ad8\u6548\u80fd\u7684\u5206\u6563\u5f0f\u8a13\u7df4\uff1a<\/p>\n \n \u6578\u64da\u5e73\u884c\u5316\uff1a\u5c07\u8a13\u7df4\u6578\u64da\u62c6\u5206\u5230GPU\u53e2\u96c6\u4e2d\uff0c\u900f\u904e\u300c\u5168\u6b78\u7d04\uff08all-reduce\uff09\u300d\u64cd\u4f5c\u540c\u6b65\u68af\u5ea6\uff0c\u76f8\u6bd4\u7d14CPU\u65b9\u6848\uff0c\u53ef\u5c07\u8a13\u7df4\u9031\u671f\uff08epoch\uff09\u6642\u9593\u7e2e\u77ed70%\u3002<\/li>\n \u6a21\u578b\u5e73\u884c\u5316\uff1a\u5c07\u5927\u578b\u6a21\u578b\u62c6\u5206\u5230\u591a\u500bGPU\u4e0a\uff0c\u6bcf\u500bGPU\u8655\u7406\u90e8\u5206\u7db2\u8def\u5c64\uff0c\u9019\u5c0d\u8a13\u7df4\u6578\u5341\u5104\u53c3\u6578\u7684\u6a21\u578b\u81f3\u95dc\u91cd\u8981\u3002<\/li>\n \u8d85\u53c3\u6578\u6700\u4f73\u5316\uff1a\u5728GPU\u53e2\u96c6\u4e0a\u5e73\u884c\u57f7\u884c\u7db2\u683c\u641c\u5c0b\uff0c\u5de5\u7a0b\u5e2b\u53ef\u5728\u5e7e\u5c0f\u6642\u5167\uff08\u800c\u975e\u5e7e\u5929\uff09\u8a55\u4f30\u6578\u5343\u7a2e\u53c3\u6578\u914d\u7f6e\u3002<\/li>\n<\/ol>\n [\/vc_column_text][\/vc_column][\/vc_row][vc_row el_class=”blog-detail-section”][vc_column][vc_column_text css=””]<\/p>\n \u5be6\u6230\u6548\u80fd\u63d0\u5347\uff1a\u95dc\u9375\u6307\u6a19\u8207\u771f\u5be6\u5834\u666f<\/strong><\/h2>\n \u63a1\u7528GPU\u5f8c\uff0c\u95dc\u9375\u6548\u80fd\u6307\u6a19\u53ef\u5be6\u73fe\u53ef\u91cf\u5316\u63d0\u5347\uff0c\u5c24\u5176\u5728\u9999\u6e2f\u4f3a\u670d\u5668\u4ee3\u7ba1\u9019\u985e\u5c0d\u4f4e\u5ef6\u9072\u3001\u9ad8\u541e\u5410\u91cf\u6709\u56b4\u683c\u8981\u6c42\u7684\u74b0\u5883\u4e2d\uff0c\u512a\u52e2\u66f4\u70ba\u660e\u986f\u3002<\/p>\n 1. \u5ef6\u9072\u964d\u4f4e\uff1a\u5f9e\u6beb\u79d2\u5230\u5fae\u79d2\u7684\u8de8\u8d8a<\/strong><\/h3>\n \u63a1\u7528GPU\u52a0\u901f\u7ba1\u7dda\u7684\u641c\u5c0b\u5f15\u64ce\uff0c\u64da\u5831\u544a\u4e2d\u4f4d\u56de\u61c9\u6642\u9593\u5df2\u5f9e300\u6beb\u79d2\u964d\u81f3100\u6beb\u79d2\u4ee5\u4e0b\uff0c\u5c3e\u90e8\u5ef6\u9072\uff08p99\uff09\u6539\u5584\u5e45\u5ea6\u905460%\u3002\u9019\u4e00\u63d0\u5347\u4e3b\u8981\u4f9d\u8cf4\u65bc\uff1a<\/p>\n \n \u5e73\u884c\u67e5\u8a62\u57f7\u884c\uff1a\u5728GPU\u6838\u5fc3\u4e0a\u540c\u6b65\u8655\u7406\u591a\u500b\u4f7f\u7528\u8005\u67e5\u8a62\uff0c\u800c\u975e\u5728CPU\u57f7\u884c\u7dd2\u4e2d\u6392\u968a\u7b49\u5f85\u3002<\/li>\n \u8a18\u61b6\u9ad4\u99d0\u7559\u7d22\u5f15\uff1a\u5c07\u9ad8\u983b\u5b58\u53d6\u7684\u7d22\u5f15\u7247\u6bb5\u5132\u5b58\u5728GPU\u8a18\u61b6\u9ad4\u4e2d\uff0c\u907f\u514d\u71b1\u9ede\u6578\u64da\u7684\u78c1\u789f\u5b58\u53d6\u5ef6\u9072\u3002<\/li>\n<\/ul>\n 2. \u541e\u5410\u91cf\u63d0\u5347\uff1a\u61c9\u5c0d\u5c16\u5cf0\u6d41\u91cf<\/strong><\/h3>\n GPU\u53e2\u96c6\u64c5\u9577\u8655\u7406\u9ad8\u4f75\u767c\u5834\u666f\uff0c\u8ca0\u8f09\u6e2c\u8a66\u986f\u793a\uff0c\u55ae\u4e00\u7bc0\u9ede\u53ef\u652f\u63f4\u8d85\u904e10\u842c\u6b21\/\u79d2\uff08QPS\uff09\u7684\u67e5\u8a62\u8acb\u6c42\u3002\u9019\u5c0d\u8655\u7406\u7522\u54c1\u767c\u5e03\u3001\u7a81\u767c\u65b0\u805e\u7b49\u6d41\u91cf\u5c16\u5cf0\u81f3\u95dc\u91cd\u8981\u2014\u2014\u5728\u9019\u985e\u5834\u666f\u4e0b\uff0c\u50b3\u7d71CPU\u7cfb\u7d71\u5f80\u5f80\u6703\u51fa\u73fe\u9650\u6d41\u6216\u505c\u6a5f\u3002<\/p>\n 3. \u80fd\u6548\u512a\u52e2\uff1a\u6c38\u7e8c\u6027\u63d0\u5347<\/strong><\/h3>\n \u5118\u7ba1GPU\u7684\u521d\u59cb\u529f\u8017\u8f03\u9ad8\uff0c\u4f46\u5176\u5353\u8d8a\u7684\u904b\u7b97\u5bc6\u5ea6\u4f7f\u5176\u5728\u5927\u898f\u6a21\u61c9\u7528\u4e2d\u5177\u5099\u66f4\u512a\u80fd\u6548\u3002\u9999\u6e2f\u63a1\u7528GPU\u52a0\u901f\u4f3a\u670d\u5668\u7684\u73fe\u4ee3\u6578\u64da\u4e2d\u5fc3\u5831\u544a\u986f\u793a\uff0c\u76f8\u6bd4\u7d14CPU\u67b6\u69cb\uff0c\u6bcf\u500b\u67e5\u8a62\u7684\u80fd\u8017\u964d\u4f4e50%\uff0c\u9019\u8207\u5168\u7403\u7da0\u8272\u904b\u7b97\u8da8\u52e2\u9ad8\u5ea6\u5951\u5408\u3002<\/p>\n [\/vc_column_text][\/vc_column][\/vc_row][vc_row el_class=”blog-detail-section”][vc_column][vc_column_text css=””]<\/p>\n GPU\u6574\u5408\u9762\u81e8\u7684\u6311\u6230\uff1a\u7a81\u7834\u6280\u8853\u969c\u7919<\/strong><\/h2>\n \u5118\u7ba1\u512a\u52e2\u986f\u8457\uff0cGPU\u4e5f\u5e36\u4f86\u4e86\u7368\u7279\u6311\u6230\uff0c\u9700\u8981\u5de5\u7a0b\u5e2b\u91dd\u5c0d\u6027\u89e3\u6c7a\u3002<\/p>\n 1. \u786c\u9ad4\u9650\u5236\u8207\u6563\u71b1\u554f\u984c<\/strong><\/h3>\n \n \u8a18\u61b6\u9ad4\u5bb9\u91cf\u9650\u5236\uff1a\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u8207\u5bc6\u96c6\u578b\u7d22\u5f15\u7d50\u69cb\u53ef\u80fd\u8d85\u904eGPU\u8a18\u61b6\u9ad4\u5bb9\u91cf\uff0c\u89e3\u6c7a\u65b9\u6848\u5305\u62ec\u8a18\u61b6\u9ad4\u96c6\u6c60\u5316\u3001\u5378\u8f09\u81f3\u9ad8\u901fNVMe\u5132\u5b58\uff0c\u4ee5\u53ca\u63a1\u7528\u5206\u5c64\u8a18\u61b6\u9ad4\u7ba1\u7406\u3002<\/li>\n \u6563\u71b1\u7ba1\u7406\uff1a\u9ad8\u5bc6\u5ea6GPU\u6a5f\u6ac3\u6703\u7522\u751f\u5927\u91cf\u71b1\u91cf\uff0c\u9700\u63a1\u7528\u6db2\u51b7\u5faa\u74b0\u7b49\u5148\u9032\u6563\u71b1\u65b9\u6848\uff0c\u76f8\u6bd4\u98a8\u51b7\u53ef\u5c07\u904b\u4f5c\u6eab\u5ea6\u964d\u4f4e20\u201330\u00b0C\u3002<\/li>\n<\/ul>\n 2. \u8edf\u9ad4\u8907\u96dc\u5ea6\uff1a\u7570\u69cb\u7cfb\u7d71\u7a0b\u5f0f\u8a2d\u8a08<\/strong><\/h3>\n \u70baCPU-GPU\u6df7\u5408\u67b6\u69cb\u958b\u767c\u7a0b\u5f0f\uff0c\u9700\u5177\u5099CUDA\u3001OpenCL\u7b49\u7570\u69cb\u904b\u7b97\u6846\u67b6\u7684\u5c08\u696d\u77e5\u8b58\u3002\u5de5\u7a0b\u5e2b\u9700\u5e73\u8861CPU\uff08\u8ca0\u8cac\u63a7\u5236\u908f\u8f2f\uff09\u8207GPU\uff08\u8ca0\u8cac\u6578\u64da\u5e73\u884c\uff09\u7684\u4efb\u52d9\u5206\u914d\uff0c\u9019\u4e00\u904e\u7a0b\u5305\u62ec\uff1a<\/p>\n \n \u6838\u5fc3\uff08Kernel\uff09\u6700\u4f73\u5316\uff1a\u8abf\u6574GPU\u6838\u5fc3\u4ee5\u6e1b\u5c11\u555f\u52d5\u958b\u92b7\u3001\u63d0\u9ad8\u6838\u5fc3\u4f7f\u7528\u7387\uff0c\u9019\u901a\u5e38\u9700\u8981\u501f\u52a9NVIDIA Nsight\u7b49\u5de5\u5177\u9032\u884c\u8fed\u4ee3\u5206\u6790\u3002<\/li>\n \u8a18\u61b6\u9ad4\u540c\u6b65\uff1a\u900f\u904e\u91d8\u9078\u8a18\u61b6\u9ad4\uff08pinned memory\uff09\u3001\u975e\u540c\u6b65\u8907\u88fd\u7b49\u6280\u8853\uff0c\u9ad8\u6548\u7ba1\u7406CPU\u8207GPU\u9593\u7684\u6578\u64da\u50b3\u8f38\u3002<\/li>\n<\/ol>\n 3. \u6210\u672c\u8207\u8cc7\u6e90\u7ba1\u7406<\/strong><\/h3>\n GPU\u786c\u9ad4\u7684\u521d\u59cb\u6295\u5165\u8f03\u9ad8\uff0c\u4f01\u696d\u7d1aGPU\u7684\u6210\u672c\u901a\u5e38\u662f\u9ad8\u968eCPU\u76845\u201310\u500d\u3002\u70ba\u8b49\u660e\u6295\u8cc7\u5408\u7406\u6027\uff0c\u7d44\u7e54\u9700\u63a1\u53d6\u4ee5\u4e0b\u63aa\u65bd\uff1a<\/p>\n \n \u8cc7\u6e90\u7de8\u6392\uff1a\u4f7f\u7528\u652f\u63f4GPU\u6392\u7a0b\u5916\u639b\u7a0b\u5f0f\u7684Kubernetes\uff0c\u52d5\u614b\u5206\u914dGPU\u6838\u5fc3\u81f3\u591a\u500b\u5de5\u4f5c\u8ca0\u8f09\uff0c\u78ba\u4fdd\u8cc7\u6e90\u4f7f\u7528\u7387\u9054\u523080%\u4ee5\u4e0a\u3002<\/li>\n \u6548\u80fd\u57fa\u6e96\u6e2c\u8a66\uff1a\u5b9a\u671f\u5be9\u6838GPU\u4f7f\u7528\u6307\u6a19\uff08\u5982\u904b\u7b97\u4f7f\u7528\u7387\u3001\u8a18\u61b6\u9ad4\u983b\u5bec\u3001PCIe\u541e\u5410\u91cf\uff09\uff0c\u8b58\u5225\u672a\u5145\u5206\u5229\u7528\u7684\u8cc7\u6e90\u3002<\/li>\n<\/ul>\n [\/vc_column_text][\/vc_column][\/vc_row][vc_row el_class=”blog-detail-section”][vc_column][vc_column_text css=””]<\/p>\n GPU\u52a0\u901f\u5728\u641c\u5c0b\u6280\u8853\u4e2d\u7684\u672a\u4f86\u8da8\u52e2<\/strong><\/h2>\n \u96a8\u8457\u786c\u9ad4\u8207\u8edf\u9ad4\u751f\u614b\u7684\u6301\u7e8c\u767c\u5c55\uff0c\u4ee5\u4e0b\u8da8\u52e2\u6b63\u5851\u9020GPU\u52a0\u901f\u641c\u5c0b\u5f15\u64ce\u7684\u672a\u4f86\u3002<\/p>\n 1. \u786c\u9ad4\u5275\u65b0<\/strong><\/h3>\n \n \u5148\u9032\u88fd\u7a0b\u6280\u8853\uff1a\u63a1\u75283nm\u30012nm\u88fd\u7a0b\u7684\u4e0b\u4e00\u4ee3GPU\uff0c\u5c07\u5177\u5099\u66f4\u9ad8\u96fb\u6676\u9ad4\u5bc6\u5ea6\uff0c\u5728\u66f4\u4f4e\u529f\u8017\u4e0b\u5be6\u73fe\u66f4\u591a\u6838\u5fc3\u8207\u66f4\u5feb\u8a18\u61b6\u9ad4\u3002<\/li>\n \u5c08\u7528\u52a0\u901f\u5668\u5354\u540c\uff1a\u8207\u5f35\u91cf\u8655\u7406\u55ae\u5143\uff08TPU\uff09\u3001\u795e\u7d93\u7db2\u8def\u8655\u7406\u55ae\u5143\uff08NPU\uff09\u3001\u73fe\u5834\u53ef\u7a0b\u5f0f\u5316\u9598\u9663\u5217\uff08FPGA\uff09\u5354\u540c\u8a2d\u8a08\uff0c\u5f62\u6210\u91dd\u5c0d\u7279\u5b9a\u4efb\u52d9\uff08\u5982\u77e5\u8b58\u5716\u8b5c\u641c\u5c0b\u7684\u5716\u8655\u7406\uff09\u6700\u4f73\u5316\u7684\u6df7\u5408\u67b6\u69cb\u3002<\/li>\n<\/ol>\n 2. \u8edf\u9ad4\u5347\u7d1a<\/strong><\/h3>\n \u81ea\u52d5\u5316\u662f\u964d\u4f4eGPU\u4f7f\u7528\u9580\u6abb\u7684\u95dc\u9375\uff1a<\/p>\n \n \u81ea\u52d5\u6700\u4f73\u5316\u5de5\u5177\uff1aTensorRT\u3001ONNX Runtime\u7b49\u6846\u67b6\u53ef\u81ea\u52d5\u70baGPU\u6700\u4f73\u5316\u6a21\u578b\u63a8\u8ad6\uff0c\u5c07\u624b\u52d5\u8abf\u6821\u5de5\u4f5c\u91cf\u6e1b\u5c1180%\u3002<\/li>\n \u7121\u4f3a\u670d\u5668GPU\u6846\u67b6\uff1a\u652f\u63f4\u6309\u9700\u5206\u914dGPU\u8cc7\u6e90\u7684\u5e73\u53f0\uff0c\u53ef\u5be6\u73fe\u8cc7\u6e90\u52d5\u614b\u64f4\u5c55\uff0c\u6309\u5be6\u969b\u4f7f\u7528\u6642\u9593\u8a08\u8cbb\u2014\u2014\u9019\u5c0d\u4e2d\u5c0f\u578b\u5718\u968a\u5c24\u70ba\u53cb\u597d\u3002<\/li>\n<\/ul>\n 3. \u908a\u7de3\u904b\u7b97\u8207\u8fd1\u57df\u6700\u4f73\u5316<\/strong><\/h3>\n \u5728\u9999\u6e2f\u9019\u985e\u9700\u5411\u6771\u5357\u4e9e\u4f7f\u7528\u8005\u63d0\u4f9b\u4f4e\u5ef6\u9072\u670d\u52d9\u7684\u5730\u5340\uff0c\u908a\u7de3GPU\u7bc0\u9ede\u6b63\u6210\u70ba\u6838\u5fc3\u53e2\u96c6\u7684\u88dc\u5145\u3002\u9019\u4e9b\u908a\u7de3\u7bc0\u9ede\u8655\u7406\u5728\u5730\u5316\u67e5\u8a62\uff0c\u53ef\u5c07\u5340\u57df\u4f7f\u7528\u8005\u7684\u5f80\u8fd4\u5ef6\u9072\u964d\u4f4e30\u201340%\uff0c\u540c\u6642\u6e1b\u8f15\u6838\u5fc3\u6578\u64da\u4e2d\u5fc3\u7684\u983b\u5bec\u8ca0\u64d4\u3002<\/p>\n [\/vc_column_text][\/vc_column][\/vc_row][vc_row el_class=”blog-detail-section”][vc_column][vc_column_text css=””]<\/p>\n \u7d50\u8a9e\uff1a\u64c1\u62b1\u641c\u5c0b\u7684\u5e73\u884c\u904b\u7b97\u672a\u4f86<\/strong><\/h2>\n GPU\u5df2\u5f9e\u300c\u5c0f\u773e\u52a0\u901f\u5668\u300d\u8f49\u8b8a\u70ba\u73fe\u4ee3\u641c\u5c0b\u5f15\u64ce\u67b6\u69cb\u7684\u6838\u5fc3\u5143\u4ef6\uff0c\u70ba\u901f\u5ea6\u3001\u53ef\u64f4\u5145\u6027\u8207\u6a5f\u5668\u5b78\u7fd2\u80fd\u529b\u5e36\u4f86\u7a81\u7834\u6027\u63d0\u5347\u3002\u5118\u7ba1\u786c\u9ad4\u7ba1\u7406\u8207\u8edf\u9ad4\u8907\u96dc\u5ea6\u4ecd\u662f\u5f85\u89e3\u6311\u6230\uff0c\u4f46\u5e73\u884c\u8655\u7406\u7684\u512a\u52e2\u5df2\u6bcb\u5eb8\u7f6e\u7591\u2014\u2014\u5c24\u5176\u5728\u9999\u6e2f\u7684\u4f3a\u670d\u5668\u79df\u7528\u8207\u4f3a\u670d\u5668\u4ee3\u7ba1\u9019\u985e\u9ad8\u6548\u80fd\u751f\u614b\u4e2d\u3002\u96a8\u8457\u6578\u64da\u91cf\u8207\u6f14\u7b97\u6cd5\u8907\u96dc\u5ea6\u7684\u6301\u7e8c\u589e\u9577\uff0c\u638c\u63e1GPU\u6574\u5408\u6280\u8853\u7684\u7d44\u7e54\uff0c\u5c07\u5728\u63d0\u4f9b\u5feb\u901f\u3001\u7cbe\u6e96\u3001\u53ef\u64f4\u5145\u641c\u5c0b\u9ad4\u9a57\u7684\u7af6\u722d\u4e2d\u5360\u64da\u7d55\u5c0d\u512a\u52e2\u3002CPU\u4e3b\u5c0e\u7684\u641c\u5c0b\u6642\u4ee3\u5df2\u843d\u5e55\uff0c\u5e73\u884c\u904b\u7b97\u7684\u672a\u4f86\u5df2\u81f3\uff0c\u800cGPU\u6b63\u662f\u9a45\u52d5\u4e0b\u4e00\u4ee3\u641c\u5c0b\u6280\u8853\u7684\u6838\u5fc3\u5f15\u64ce\u3002<\/p>\n \u7121\u8ad6\u4f60\u662f\u5728\u6700\u4f73\u5316\u5927\u578b\u4f01\u696d\u641c\u5c0b\u5f15\u64ce\uff0c\u9084\u662f\u5efa\u69cb\u5782\u76f4\u9818\u57df\u7684\u5c0f\u773e\u641c\u5c0b\u5de5\u5177\uff0c\u7406\u89e3GPU\u52a0\u901f\u90fd\u5df2\u5f9e\u300c\u53ef\u9078\u9805\u300d\u8b8a\u70ba\u300c\u5fc5\u9078\u9805\u300d\u3002\u900f\u904e\u904b\u7528\u9019\u985e\u6280\u8853\uff0c\u4f60\u80fd\u6ee1\u8db3\u4f7f\u7528\u8005\u8207\u6578\u64da\u4e0d\u65b7\u589e\u9577\u7684\u9700\u6c42\uff0c\u78ba\u4fdd\u641c\u5c0b\u7cfb\u7d71\u5728\u65e5\u76ca\u8907\u96dc\u7684\u6578\u4f4d\u74b0\u5883\u4e2d\u4fdd\u6301\u7af6\u722d\u529b\u3002<\/p>\n [\/vc_column_text][\/vc_column][\/vc_row]<\/p>\n<\/div>","protected":false},"excerpt":{"rendered":" [vc_row el_class=”blog-detail-section”][vc_ […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":1,"featured_media":27241,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[61],"tags":[8825,8465,7753,3306,2681],"class_list":["post-27245","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-hong-kong-dedicated-server-tc","tag-search-engine-technology-tc","tag-heterogeneous-computing-tc","tag-distributed-systems-tc","tag-gpu-acceleration-tc","tag-parallel-computing-tc"],"acf":[],"yoast_head":"\n