\u8981\u7406\u89e3GPU\u4e3a\u4f55\u80fd\u6210\u4e3a\u201c\u6e38\u620f\u89c4\u5219\u6539\u53d8\u8005\u201d\uff0c\u9996\u5148\u9700\u5bf9\u6bd4\u5176\u4e0e\u4f20\u7edfCPU\u7684\u8bbe\u8ba1\u5dee\u5f02\u3002CPU\u4e13\u4e3a\u4e32\u884c\u5904\u7406\u4f18\u5316\uff0c\u4ec5\u914d\u5907\u5c11\u91cf\u9ad8\u9891\u6838\u5fc3\uff0c\u64c5\u957f\u5904\u7406\u590d\u6742\u63a7\u5236\u903b\u8f91\uff1b\u800cGPU\u5219\u4ee5\u5e76\u884c\u4e3a\u6838\u5fc3\u8bbe\u8ba1\uff0c\u96c6\u6210\u6570\u5343\u4e2a\u4f4e\u9891\u6838\u5fc3\uff0c\u80fd\u9ad8\u6548\u6267\u884c\u201c\u5bf9\u591a\u7ec4\u6570\u636e\u91cd\u590d\u540c\u4e00\u6307\u4ee4\u201d\u7684\u8fd0\u7b97\u3002\u8fd9\u79cd\u7279\u6027\u4f7f\u5176\u6210\u4e3a\u73b0\u4ee3\u641c\u7d22\u5f15\u64ce\u4e2d\u201c\u8ba1\u7b97\u5bc6\u96c6\u578b\u3001\u6570\u636e\u5e76\u884c\u5316\u201d\u4efb\u52a1\u7684\u7406\u60f3\u9009\u62e9\u3002<\/p>\n

\u4ee5\u4e0b\u662f\u4e24\u8005\u7684\u67b6\u6784\u5bf9\u6bd4\uff1a<\/p>\n\n\n\n\n\n\n\n

\u6307\u6807<\/th>\n	CPU<\/th>\n	GPU<\/th>\n<\/tr>\n
\u6838\u5fc3\u6570\u91cf<\/td>\n	8\u201364<\/td>\n	2,000\u20134,000+<\/td>\n<\/tr>\n
\u65f6\u949f\u9891\u7387<\/td>\n	3\u20135 GHz<\/td>\n	1\u20132 GHz<\/td>\n<\/tr>\n
\u5185\u5b58\u5e26\u5bbd<\/td>\n	50\u2013100 GB\/s<\/td>\n	400\u20131,000 GB\/s<\/td>\n<\/tr>\n
\u64c5\u957f\u7684\u4efb\u52a1<\/td>\n	\u590d\u6742\u903b\u8f91\u63a7\u5236<\/td>\n	\u5927\u89c4\u6a21\u5e76\u884c\u8ba1\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\u9a71\u52a8\u7684\u641c\u7d22\u5f15\u64ce\u8fd0\u4f5c\u6838\u5fc3\u9636\u6bb5<\/strong><\/h2>\n \u641c\u7d22\u5f15\u64ce\u7684\u8fd0\u4f5c\u5305\u542b\u6293\u53d6\u3001\u7d22\u5f15\u3001\u6392\u5e8f\u3001\u673a\u5668\u5b66\u4e60\u56db\u5927\u6838\u5fc3\u9636\u6bb5\uff0cGPU\u901a\u8fc7\u4e13\u95e8\u7684\u5e76\u884c\u5316\u7b56\u7565\uff0c\u4e3a\u6bcf\u4e2a\u9636\u6bb5\u63d0\u4f9b\u6027\u80fd\u589e\u76ca\u3002<\/p>\n 1. \u7f51\u9875\u6293\u53d6\uff1a\u6570\u636e\u91c7\u96c6\u7684\u5e76\u884c\u5316\u5347\u7ea7<\/strong><\/h3>\n \u6293\u53d6\u9636\u6bb5\u9700\u4ece\u4e92\u8054\u7f51\u83b7\u53d6\u7f51\u9875\u6570\u636e\uff0c\u4f46\u53cd\u722c\u673a\u5236\u7684\u9650\u5236\u4e0e\u6d77\u91cfURL\u7684\u5904\u7406\u538b\u529b\uff0c\u4e00\u76f4\u662f\u8be5\u9636\u6bb5\u7684\u74f6\u9888\u3002GPU\u901a\u8fc7\u5206\u5e03\u5f0f\u5e76\u884cURL\u5904\u7406\uff0c\u5f7b\u5e95\u6539\u53d8\u4e86\u8fd9\u4e00\u5c40\u9762\uff1a<\/p>\n \n \u57fa\u4e8e\u96c6\u7fa4\u7684\u4efb\u52a1\u8c03\u5ea6\uff1aGPU\u96c6\u7fa4\u5e76\u884c\u7ba1\u7406\u6570\u767e\u4e07\u4e2aURL\uff0c\u5c06\u6293\u53d6\u3001\u6e32\u67d3\u3001\u89e3\u6790\u4efb\u52a1\u5206\u914d\u5230\u6570\u5343\u4e2a\u6838\u5fc3\u4e0a\u540c\u6b65\u6267\u884c\u3002<\/li>\n JavaScript\u6e32\u67d3\u52a0\u901f\uff1a\u73b0\u4ee3\u7f51\u7ad9\u5927\u91cf\u4f9d\u8d56\u5ba2\u6237\u7aef\u811a\u672c\uff0c\u4f20\u7edf\u722c\u866b\u96be\u4ee5\u9ad8\u6548\u5904\u7406\uff0c\u800cGPU\u53ef\u5378\u8f7dJS\u6267\u884c\u4efb\u52a1\uff0c\u5728\u5b9e\u9645\u573a\u666f\u4e2d\u80fd\u5c06\u6e32\u67d3\u5ef6\u8fdf\u964d\u4f4e40%\u3002<\/li>\n \u7f51\u7edcI\/O\u4f18\u5316\uff1a\u5728\u4e2d\u56fd\u9999\u6e2f\u670d\u52a1\u5668\u79df\u7528\u8fd9\u7c7b\u9ad8\u5e26\u5bbd\u73af\u5883\u4e2d\uff0cGPU\u4e0e\u4f4e\u5ef6\u8fdf\u7f51\u7edc\u534f\u540c\uff0c\u6700\u5927\u7a0b\u5ea6\u51cf\u5c11\u722c\u866b\u8282\u70b9\u95f4\u7684\u6570\u636e\u4f20\u8f93\u5f00\u9500\u3002<\/li>\n<\/ul>\n 2. \u7d22\u5f15\u6784\u5efa\uff1a\u89c4\u6a21\u5316\u6784\u5efa\u9ad8\u6548\u6570\u636e\u7ed3\u6784<\/strong><\/h3>\n \u7d22\u5f15\u9636\u6bb5\u9700\u5c06\u6293\u53d6\u5230\u7684\u6570\u636e\u8f6c\u6362\u4e3a\u53ef\u641c\u7d22\u7ed3\u6784\uff0c\u8fc7\u7a0b\u4e2d\u6d89\u53ca\u5927\u91cf\u6587\u672c\u5206\u6790\u4e0e\u6570\u636e\u8f6c\u6362\u64cd\u4f5c\u3002GPU\u51ed\u501f\u5e76\u884c\u5904\u7406\u80fd\u529b\uff0c\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u4e0e\u5012\u6392\u7d22\u5f15\u6784\u5efa\u4e2d\u8868\u73b0\u7a81\u51fa\uff1a<\/p>\n \n NLP\u6d41\u6c34\u7ebf\u52a0\u901f\uff1a\u5206\u8bcd\u3001\u8bcd\u5e72\u63d0\u53d6\u3001\u8bcd\u6027\u6807\u6ce8\u7b49\u6838\u5fc3NLP\u4efb\u52a1\u5728GPU\u6838\u5fc3\u4e0a\u5e76\u884c\u5904\u7406\uff0c\u6bcf\u79d2\u53ef\u5904\u7406\u6570\u767e\u4e07\u4e2a\u8bcd\u5143\uff08token\uff09\u3002<\/li>\n \u5012\u6392\u7d22\u5f15\u5206\u7247\uff1a\u5c06\u7d22\u5f15\u7247\u6bb5\u5206\u5e03\u5f0f\u5b58\u50a8\u5230GPU\u5185\u5b58\u4e2d\uff0c\u641c\u7d22\u5f15\u64ce\u53ef\u51cf\u5c11\u78c1\u76d8I\/O\u74f6\u9888\u3002\u5185\u5b58\u5e26\u5bbd\u4f18\u52bf\u80fd\u5b9e\u73b0\u201c\u8bcd-\u6587\u6863\u77e9\u9635\u201d\u7684\u4e9a\u6beb\u79d2\u7ea7\u8bbf\u95ee\uff0c\u8fd9\u5bf9\u5b9e\u65f6\u67e5\u8be2\u5904\u7406\u81f3\u5173\u91cd\u8981\u3002<\/li>\n \u538b\u7f29\u7b97\u6cd5\u4f18\u5316\uff1aGPU\u5bf9\u7d22\u5f15\u6570\u636e\u8fdb\u884c\u65e0\u635f\u538b\u7f29\u4f18\u5316\uff0c\u5728\u4fdd\u8bc1\u5b58\u50a8\u6548\u7387\u7684\u540c\u65f6\uff0c\u517c\u987e\u67e5\u8be2\u6267\u884c\u65f6\u7684\u5feb\u901f\u89e3\u538b\u9700\u6c42\u3002<\/li>\n<\/ol>\n 3. \u7ed3\u679c\u6392\u5e8f\uff1a\u5b9e\u65f6\u76f8\u5173\u6027\u8ba1\u7b97<\/strong><\/h3>\n \u6392\u5e8f\u9636\u6bb5\u9700\u6839\u636e\u67e5\u8be2\u76f8\u5173\u6027\u786e\u5b9a\u7ed3\u679c\u987a\u5e8f\uff0c\u5982\u4eca\u8be5\u8fc7\u7a0b\u5df2\u7531\u673a\u5668\u5b66\u4e60\u6a21\u578b\u4e3b\u5bfc\u3002GPU\u4e3a\u590d\u6742\u6a21\u578b\u63d0\u4f9b\u8fd1\u5b9e\u65f6\u7684\u63a8\u7406\u80fd\u529b\uff1a<\/p>\n \u8bcd\u9891-\u9006\u6587\u6863\u9891\u7387\uff08TF-IDF\uff09\u7b49\u5411\u91cf\u7a7a\u95f4\u6a21\u578b\uff0c\u4f9d\u8d56\u4e8e\u968f\u6570\u636e\u96c6\u89c4\u6a21\u7ebf\u6027\u589e\u957f\u7684\u77e9\u9635\u8fd0\u7b97\u3002GPU\u5229\u7528\u5e76\u884c\u6838\u5fc3\u540c\u6b65\u8ba1\u7b97\u4f59\u5f26\u76f8\u4f3c\u5ea6\u4e0e\u76f8\u5173\u6027\u5206\u6570\uff0c\u5c06\u8ba1\u7b97\u65f6\u95f4\u4ece\u6beb\u79d2\u7ea7\u538b\u7f29\u81f3\u5fae\u79d2\u7ea7\u3002\u5bf9\u4e8eBERT\u4e0eTransformer\u7cfb\u5217\u7b49\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\uff0cGPU\u652f\u6301\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u4e0e\u63a8\u7406\uff1a<\/p>\n \n FP16\u6d6e\u70b9\u8fd0\u7b97\uff1a\u91c7\u7528\u534a\u7cbe\u5ea6\u6d6e\u70b9\u6570\uff0cGPU\u5904\u7406\u795e\u7ecf\u7f51\u7edc\u5c42\u7684\u901f\u5ea6\u53ef\u8fbeCPU\u76842\u500d\uff0c\u4e14\u4e0d\u4f1a\u5bfc\u81f4\u663e\u8457\u7cbe\u5ea6\u635f\u5931\u3002<\/li>\n \u5f20\u91cf\u6838\u5fc3\uff08Tensor Core\uff09\u4f18\u5316\uff1a\u73b0\u4ee3GPU\u4e2d\u7684\u4e13\u7528\u77e9\u9635\u5904\u7406\u5355\u5143\uff0c\u53ef\u5c06\u6df1\u5ea6\u5b66\u4e60\u8fd0\u7b97\u901f\u5ea6\u63d0\u534750%\uff0c\u8fd9\u5bf9\u9ad8\u6d41\u91cf\u573a\u666f\u4e0b\u7684\u5b9e\u65f6\u6392\u5e8f\u81f3\u5173\u91cd\u8981\u3002<\/li>\n<\/ul>\n 4. \u673a\u5668\u5b66\u4e60\uff1a\u8bad\u7ec3\u4e0b\u4e00\u4ee3\u7b97\u6cd5<\/strong><\/h3>\n \u8bad\u7ec3\u5927\u89c4\u6a21\u6a21\u578b\u9700\u5c06\u8ba1\u7b97\u4efb\u52a1\u5206\u5e03\u5f0f\u90e8\u7f72\u5230\u6570\u5343\u4e2a\u8282\u70b9\u4e0a\u3002GPU\u7ed3\u5408Horovod\u3001TensorFlow\u7b49\u6846\u67b6\uff0c\u53ef\u5b9e\u73b0\u9ad8\u6548\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\uff1a<\/p>\n \n \u6570\u636e\u5e76\u884c\u5316\uff1a\u5c06\u8bad\u7ec3\u6570\u636e\u62c6\u5206\u5230GPU\u96c6\u7fa4\u4e2d\uff0c\u901a\u8fc7\u201c\u5168\u5f52\u7ea6\uff08all-reduce\uff09\u201d\u64cd\u4f5c\u540c\u6b65\u68af\u5ea6\uff0c\u76f8\u6bd4\u7eafCPU\u65b9\u6848\uff0c\u53ef\u5c06\u8bad\u7ec3\u8f6e\u6b21\uff08epoch\uff09\u65f6\u95f4\u7f29\u77ed70%\u3002<\/li>\n \u6a21\u578b\u5e76\u884c\u5316\uff1a\u5c06\u5927\u578b\u6a21\u578b\u62c6\u5206\u5230\u591a\u4e2aGPU\u4e0a\uff0c\u6bcf\u4e2aGPU\u5904\u7406\u90e8\u5206\u7f51\u7edc\u5c42\uff0c\u8fd9\u5bf9\u8bad\u7ec3\u6570\u5341\u4ebf\u53c2\u6570\u7684\u6a21\u578b\u5fc5\u4e0d\u53ef\u5c11\u3002<\/li>\n \u8d85\u53c2\u6570\u4f18\u5316\uff1a\u5728GPU\u96c6\u7fa4\u4e0a\u5e76\u884c\u6267\u884c\u7f51\u683c\u641c\u7d22\uff0c\u5de5\u7a0b\u5e08\u53ef\u5728\u51e0\u5c0f\u65f6\u5185\uff08\u800c\u975e\u51e0\u5929\uff09\u8bc4\u4f30\u6570\u5343\u79cd\u53c2\u6570\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 \u5b9e\u6218\u6027\u80fd\u63d0\u5347\uff1a\u5173\u952e\u6307\u6807\u4e0e\u771f\u5b9e\u573a\u666f<\/strong><\/h2>\n \u91c7\u7528GPU\u540e\uff0c\u5173\u952e\u6027\u80fd\u6307\u6807\u53ef\u5b9e\u73b0\u53ef\u91cf\u5316\u63d0\u5347\uff0c\u5c24\u5176\u5728\u4e2d\u56fd\u9999\u6e2f\u670d\u52a1\u5668\u6258\u7ba1\u8fd9\u7c7b\u5bf9\u4f4e\u5ef6\u8fdf\u3001\u9ad8\u541e\u5410\u91cf\u6709\u4e25\u683c\u8981\u6c42\u7684\u73af\u5883\u4e2d\uff0c\u4f18\u52bf\u66f4\u4e3a\u660e\u663e\u3002<\/p>\n 1. \u5ef6\u8fdf\u964d\u4f4e\uff1a\u4ece\u6beb\u79d2\u5230\u5fae\u79d2\u7684\u8de8\u8d8a<\/strong><\/h3>\n \u91c7\u7528GPU\u52a0\u901f\u6d41\u6c34\u7ebf\u7684\u641c\u7d22\u5f15\u64ce\uff0c\u636e\u62a5\u544a\u4e2d\u4f4d\u54cd\u5e94\u65f6\u95f4\u5df2\u4ece300\u6beb\u79d2\u964d\u81f3100\u6beb\u79d2\u4ee5\u4e0b\uff0c\u5c3e\u90e8\u5ef6\u8fdf\uff08p99\uff09\u6539\u5584\u5e45\u5ea6\u8fbe60%\u3002\u8fd9\u4e00\u63d0\u5347\u4e3b\u8981\u4f9d\u8d56\u4e8e\uff1a<\/p>\n \n \u5e76\u884c\u67e5\u8be2\u6267\u884c\uff1a\u5728GPU\u6838\u5fc3\u4e0a\u540c\u6b65\u5904\u7406\u591a\u4e2a\u7528\u6237\u67e5\u8be2\uff0c\u800c\u975e\u5728CPU\u7ebf\u7a0b\u4e2d\u6392\u961f\u7b49\u5f85\u3002<\/li>\n \u5185\u5b58\u9a7b\u7559\u7d22\u5f15\uff1a\u5c06\u9ad8\u9891\u8bbf\u95ee\u7684\u7d22\u5f15\u7247\u6bb5\u5b58\u50a8\u5728GPU\u5185\u5b58\u4e2d\uff0c\u907f\u514d\u70ed\u70b9\u6570\u636e\u7684\u78c1\u76d8\u8bbf\u95ee\u5ef6\u8fdf\u3002<\/li>\n<\/ul>\n 2. \u541e\u5410\u91cf\u63d0\u5347\uff1a\u5e94\u5bf9\u5cf0\u503c\u6d41\u91cf<\/strong><\/h3>\n GPU\u96c6\u7fa4\u64c5\u957f\u5904\u7406\u9ad8\u5e76\u53d1\u573a\u666f\uff0c\u8d1f\u8f7d\u6d4b\u8bd5\u663e\u793a\uff0c\u5355\u4e2a\u8282\u70b9\u53ef\u652f\u6301\u8d85\u8fc710\u4e07\u6b21\/\u79d2\uff08QPS\uff09\u7684\u67e5\u8be2\u8bf7\u6c42\u3002\u8fd9\u5bf9\u5904\u7406\u4ea7\u54c1\u53d1\u5e03\u3001\u7a81\u53d1\u65b0\u95fb\u7b49\u6d41\u91cf\u5cf0\u503c\u81f3\u5173\u91cd\u8981\u2014\u2014\u5728\u8fd9\u7c7b\u573a\u666f\u4e0b\uff0c\u4f20\u7edfCPU\u7cfb\u7edf\u5f80\u5f80\u4f1a\u51fa\u73b0\u9650\u6d41\u6216\u5b95\u673a\u3002<\/p>\n 3. \u80fd\u6548\u4f18\u52bf\uff1a\u53ef\u6301\u7eed\u6027\u63d0\u5347<\/strong><\/h3>\n \u5c3d\u7ba1GPU\u7684\u521d\u59cb\u529f\u8017\u8f83\u9ad8\uff0c\u4f46\u5176\u5353\u8d8a\u7684\u8ba1\u7b97\u5bc6\u5ea6\u4f7f\u5176\u5728\u89c4\u6a21\u5316\u5e94\u7528\u4e2d\u5177\u5907\u66f4\u4f18\u80fd\u6548\u3002\u4e2d\u56fd\u9999\u6e2f\u91c7\u7528GPU\u52a0\u901f\u670d\u52a1\u5668\u7684\u73b0\u4ee3\u6570\u636e\u4e2d\u5fc3\u62a5\u544a\u663e\u793a\uff0c\u76f8\u6bd4\u7eafCPU\u67b6\u6784\uff0c\u6bcf\u67e5\u8be2\u7684\u80fd\u8017\u964d\u4f4e50%\uff0c\u8fd9\u4e0e\u5168\u7403\u7eff\u8272\u8ba1\u7b97\u8d8b\u52bf\u9ad8\u5ea6\u5951\u5408\u3002<\/p>\n GPU\u6574\u5408\u9762\u4e34\u7684\u6311\u6218\uff1a\u7a81\u7834\u6280\u672f\u969c\u788d<\/strong><\/h2>\n \u5c3d\u7ba1\u4f18\u52bf\u663e\u8457\uff0cGPU\u4e5f\u5e26\u6765\u4e86\u72ec\u7279\u6311\u6218\uff0c\u9700\u8981\u5de5\u7a0b\u5e08\u9488\u5bf9\u6027\u89e3\u51b3\u3002<\/p>\n 1. \u786c\u4ef6\u9650\u5236\u4e0e\u6563\u70ed\u95ee\u9898<\/strong><\/h3>\n \n \u5185\u5b58\u5bb9\u91cf\u9650\u5236\uff1a\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e0e\u5bc6\u96c6\u578b\u7d22\u5f15\u7ed3\u6784\u53ef\u80fd\u8d85\u51faGPU\u5185\u5b58\u5bb9\u91cf\uff0c\u89e3\u51b3\u65b9\u6848\u5305\u62ec\u5185\u5b58\u6c60\u5316\u3001\u5378\u8f7d\u5230\u9ad8\u901fNVMe\u5b58\u50a8\uff0c\u4ee5\u53ca\u91c7\u7528\u5206\u5c42\u5185\u5b58\u7ba1\u7406\u3002<\/li>\n \u6563\u70ed\u7ba1\u7406\uff1a\u9ad8\u5bc6\u5ea6GPU\u673a\u67b6\u4f1a\u4ea7\u751f\u5927\u91cf\u70ed\u91cf\uff0c\u9700\u91c7\u7528\u6db2\u51b7\u5faa\u73af\u7b49\u5148\u8fdb\u6563\u70ed\u65b9\u6848\uff0c\u76f8\u6bd4\u98ce\u51b7\u53ef\u5c06\u8fd0\u884c\u6e29\u5ea6\u964d\u4f4e20\u201330\u00b0C\u3002<\/li>\n<\/ul>\n 2. \u8f6f\u4ef6\u590d\u6742\u5ea6\uff1a\u5f02\u6784\u7cfb\u7edf\u7f16\u7a0b<\/strong><\/h3>\n \u4e3aCPU-GPU\u6df7\u5408\u67b6\u6784\u5f00\u53d1\u7a0b\u5e8f\uff0c\u9700\u638c\u63e1CUDA\u3001OpenCL\u7b49\u5f02\u6784\u8ba1\u7b97\u6846\u67b6\u7684\u4e13\u4e1a\u77e5\u8bc6\u3002\u5de5\u7a0b\u5e08\u9700\u5e73\u8861CPU\uff08\u8d1f\u8d23\u63a7\u5236\u903b\u8f91\uff09\u4e0eGPU\uff08\u8d1f\u8d23\u6570\u636e\u5e76\u884c\uff09\u7684\u4efb\u52a1\u5206\u914d\uff0c\u8fd9\u4e00\u8fc7\u7a0b\u5305\u62ec\uff1a<\/p>\n \n \u5185\u6838\uff08Kernel\uff09\u4f18\u5316\uff1a\u8c03\u6574GPU\u5185\u6838\u4ee5\u51cf\u5c11\u542f\u52a8\u5f00\u9500\u3001\u63d0\u9ad8\u6838\u5fc3\u5360\u7528\u7387\uff0c\u8fd9\u901a\u5e38\u9700\u8981\u501f\u52a9NVIDIA Nsight\u7b49\u5de5\u5177\u8fdb\u884c\u8fed\u4ee3\u5206\u6790\u3002<\/li>\n \u5185\u5b58\u540c\u6b65\uff1a\u901a\u8fc7\u56fa\u5b9a\u5185\u5b58\uff08pinned memory\uff09\u3001\u5f02\u6b65\u62f7\u8d1d\u7b49\u6280\u672f\uff0c\u9ad8\u6548\u7ba1\u7406CPU\u4e0eGPU\u95f4\u7684\u6570\u636e\u4f20\u8f93\u3002<\/li>\n<\/ol>\n 3. \u6210\u672c\u4e0e\u8d44\u6e90\u7ba1\u7406<\/strong><\/h3>\n GPU\u786c\u4ef6\u7684\u521d\u59cb\u6295\u5165\u8f83\u9ad8\uff0c\u4f01\u4e1a\u7ea7GPU\u7684\u6210\u672c\u901a\u5e38\u662f\u9ad8\u7aefCPU\u76845\u201310\u500d\u3002\u4e3a\u8bc1\u660e\u6295\u8d44\u5408\u7406\u6027\uff0c\u7ec4\u7ec7\u9700\u91c7\u53d6\u4ee5\u4e0b\u63aa\u65bd\uff1a<\/p>\n \n \u8d44\u6e90\u7f16\u6392\uff1a\u4f7f\u7528\u652f\u6301GPU\u8c03\u5ea6\u63d2\u4ef6\u7684Kubernetes\uff0c\u52a8\u6001\u5206\u914dGPU\u6838\u5fc3\u5230\u591a\u4e2a\u5de5\u4f5c\u8d1f\u8f7d\uff0c\u786e\u4fdd\u8d44\u6e90\u5229\u7528\u7387\u8fbe\u523080%\u4ee5\u4e0a\u3002<\/li>\n \u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5\uff1a\u5b9a\u671f\u5ba1\u8ba1GPU\u4f7f\u7528\u6307\u6807\uff08\u5982\u8ba1\u7b97\u5229\u7528\u7387\u3001\u5185\u5b58\u5e26\u5bbd\u3001PCIe\u541e\u5410\u91cf\uff09\uff0c\u8bc6\u522b\u672a\u5145\u5206\u5229\u7528\u7684\u8d44\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\u6574\u5408\u9762\u4e34\u7684\u6311\u6218\uff1a\u7a81\u7834\u6280\u672f\u969c\u788d<\/strong><\/h2>\n \u5c3d\u7ba1\u4f18\u52bf\u663e\u8457\uff0cGPU\u4e5f\u5e26\u6765\u4e86\u72ec\u7279\u6311\u6218\uff0c\u9700\u8981\u5de5\u7a0b\u5e08\u9488\u5bf9\u6027\u89e3\u51b3\u3002<\/p>\n 1. \u786c\u4ef6\u9650\u5236\u4e0e\u6563\u70ed\u95ee\u9898<\/strong><\/h3>\n \n \u5185\u5b58\u5bb9\u91cf\u9650\u5236\uff1a\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e0e\u5bc6\u96c6\u578b\u7d22\u5f15\u7ed3\u6784\u53ef\u80fd\u8d85\u51faGPU\u5185\u5b58\u5bb9\u91cf\uff0c\u89e3\u51b3\u65b9\u6848\u5305\u62ec\u5185\u5b58\u6c60\u5316\u3001\u5378\u8f7d\u5230\u9ad8\u901fNVMe\u5b58\u50a8\uff0c\u4ee5\u53ca\u91c7\u7528\u5206\u5c42\u5185\u5b58\u7ba1\u7406\u3002<\/li>\n \u6563\u70ed\u7ba1\u7406\uff1a\u9ad8\u5bc6\u5ea6GPU\u673a\u67b6\u4f1a\u4ea7\u751f\u5927\u91cf\u70ed\u91cf\uff0c\u9700\u91c7\u7528\u6db2\u51b7\u5faa\u73af\u7b49\u5148\u8fdb\u6563\u70ed\u65b9\u6848\uff0c\u76f8\u6bd4\u98ce\u51b7\u53ef\u5c06\u8fd0\u884c\u6e29\u5ea6\u964d\u4f4e20\u201330\u00b0C\u3002<\/li>\n<\/ul>\n 2. \u8f6f\u4ef6\u590d\u6742\u5ea6\uff1a\u5f02\u6784\u7cfb\u7edf\u7f16\u7a0b<\/strong><\/h3>\n \u4e3aCPU-GPU\u6df7\u5408\u67b6\u6784\u5f00\u53d1\u7a0b\u5e8f\uff0c\u9700\u638c\u63e1CUDA\u3001OpenCL\u7b49\u5f02\u6784\u8ba1\u7b97\u6846\u67b6\u7684\u4e13\u4e1a\u77e5\u8bc6\u3002\u5de5\u7a0b\u5e08\u9700\u5e73\u8861CPU\uff08\u8d1f\u8d23\u63a7\u5236\u903b\u8f91\uff09\u4e0eGPU\uff08\u8d1f\u8d23\u6570\u636e\u5e76\u884c\uff09\u7684\u4efb\u52a1\u5206\u914d\uff0c\u8fd9\u4e00\u8fc7\u7a0b\u5305\u62ec\uff1a<\/p>\n \n \u5185\u6838\uff08Kernel\uff09\u4f18\u5316\uff1a\u8c03\u6574GPU\u5185\u6838\u4ee5\u51cf\u5c11\u542f\u52a8\u5f00\u9500\u3001\u63d0\u9ad8\u6838\u5fc3\u5360\u7528\u7387\uff0c\u8fd9\u901a\u5e38\u9700\u8981\u501f\u52a9NVIDIA Nsight\u7b49\u5de5\u5177\u8fdb\u884c\u8fed\u4ee3\u5206\u6790\u3002<\/li>\n \u5185\u5b58\u540c\u6b65\uff1a\u901a\u8fc7\u56fa\u5b9a\u5185\u5b58\uff08pinned memory\uff09\u3001\u5f02\u6b65\u62f7\u8d1d\u7b49\u6280\u672f\uff0c\u9ad8\u6548\u7ba1\u7406CPU\u4e0eGPU\u95f4\u7684\u6570\u636e\u4f20\u8f93\u3002<\/li>\n<\/ol>\n 3. \u6210\u672c\u4e0e\u8d44\u6e90\u7ba1\u7406<\/strong><\/h3>\n GPU\u786c\u4ef6\u7684\u521d\u59cb\u6295\u5165\u8f83\u9ad8\uff0c\u4f01\u4e1a\u7ea7GPU\u7684\u6210\u672c\u901a\u5e38\u662f\u9ad8\u7aefCPU\u76845\u201310\u500d\u3002\u4e3a\u8bc1\u660e\u6295\u8d44\u5408\u7406\u6027\uff0c\u7ec4\u7ec7\u9700\u91c7\u53d6\u4ee5\u4e0b\u63aa\u65bd\uff1a<\/p>\n \n \u8d44\u6e90\u7f16\u6392\uff1a\u4f7f\u7528\u652f\u6301GPU\u8c03\u5ea6\u63d2\u4ef6\u7684Kubernetes\uff0c\u52a8\u6001\u5206\u914dGPU\u6838\u5fc3\u5230\u591a\u4e2a\u5de5\u4f5c\u8d1f\u8f7d\uff0c\u786e\u4fdd\u8d44\u6e90\u5229\u7528\u7387\u8fbe\u523080%\u4ee5\u4e0a\u3002<\/li>\n \u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5\uff1a\u5b9a\u671f\u5ba1\u8ba1GPU\u4f7f\u7528\u6307\u6807\uff08\u5982\u8ba1\u7b97\u5229\u7528\u7387\u3001\u5185\u5b58\u5e26\u5bbd\u3001PCIe\u541e\u5410\u91cf\uff09\uff0c\u8bc6\u522b\u672a\u5145\u5206\u5229\u7528\u7684\u8d44\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\u7d22\u6280\u672f\u4e2d\u7684\u672a\u6765\u8d8b\u52bf<\/strong><\/h2>\n \u968f\u7740\u786c\u4ef6\u4e0e\u8f6f\u4ef6\u751f\u6001\u7684\u6301\u7eed\u53d1\u5c55\uff0c\u4ee5\u4e0b\u8d8b\u52bf\u6b63\u5851\u9020GPU\u52a0\u901f\u641c\u7d22\u5f15\u64ce\u7684\u672a\u6765\u3002<\/p>\n 1. \u786c\u4ef6\u521b\u65b0<\/strong><\/h3>\n \n \u5148\u8fdb\u5236\u7a0b\u6280\u672f\uff1a\u91c7\u75283nm\u30012nm\u5236\u7a0b\u7684\u4e0b\u4e00\u4ee3GPU\uff0c\u5c06\u5177\u5907\u66f4\u9ad8\u6676\u4f53\u7ba1\u5bc6\u5ea6\uff0c\u5728\u66f4\u4f4e\u529f\u8017\u4e0b\u5b9e\u73b0\u66f4\u591a\u6838\u5fc3\u4e0e\u66f4\u5feb\u5185\u5b58\u3002<\/li>\n \u4e13\u7528\u52a0\u901f\u5668\u534f\u540c\uff1a\u4e0e\u5f20\u91cf\u5904\u7406\u5355\u5143\uff08TPU\uff09\u3001\u795e\u7ecf\u7f51\u7edc\u5904\u7406\u5355\u5143\uff08NPU\uff09\u3001\u73b0\u573a\u53ef\u7f16\u7a0b\u95e8\u9635\u5217\uff08FPGA\uff09\u534f\u540c\u8bbe\u8ba1\uff0c\u5f62\u6210\u9488\u5bf9\u7279\u5b9a\u4efb\u52a1\uff08\u5982\u77e5\u8bc6\u56fe\u8c31\u641c\u7d22\u7684\u56fe\u5904\u7406\uff09\u4f18\u5316\u7684\u6df7\u5408\u67b6\u6784\u3002<\/li>\n<\/ol>\n 2. \u8f6f\u4ef6\u5347\u7ea7<\/strong><\/h3>\n \u81ea\u52a8\u5316\u662f\u964d\u4f4eGPU\u4f7f\u7528\u95e8\u69db\u7684\u5173\u952e\uff1a<\/p>\n \n \u81ea\u52a8\u4f18\u5316\u5de5\u5177\uff1aTensorRT\u3001ONNX Runtime\u7b49\u6846\u67b6\u53ef\u81ea\u52a8\u4e3aGPU\u4f18\u5316\u6a21\u578b\u63a8\u7406\uff0c\u5c06\u624b\u52a8\u8c03\u4f18\u5de5\u4f5c\u91cf\u51cf\u5c1180%\u3002<\/li>\n \u65e0\u670d\u52a1\u5668GPU\u6846\u67b6\uff1a\u652f\u6301\u6309\u9700\u5206\u914dGPU\u8d44\u6e90\u7684\u5e73\u53f0\uff0c\u53ef\u5b9e\u73b0\u8d44\u6e90\u52a8\u6001\u6269\u5c55\uff0c\u6309\u5b9e\u9645\u4f7f\u7528\u65f6\u957f\u8ba1\u8d39\u2014\u2014\u8fd9\u5bf9\u4e2d\u5c0f\u578b\u56e2\u961f\u5c24\u4e3a\u53cb\u597d\u3002<\/li>\n<\/ul>\n 3. \u8fb9\u7f18\u8ba1\u7b97\u4e0e\u8fd1\u57df\u4f18\u5316<\/strong><\/h3>\n \u5728\u4e2d\u56fd\u9999\u6e2f\u8fd9\u7c7b\u9700\u5411\u4e1c\u5357\u4e9a\u7528\u6237\u63d0\u4f9b\u4f4e\u5ef6\u8fdf\u670d\u52a1\u7684\u5730\u533a\uff0c\u8fb9\u7f18GPU\u8282\u70b9\u6b63\u6210\u4e3a\u6838\u5fc3\u96c6\u7fa4\u7684\u8865\u5145\u3002\u8fd9\u4e9b\u8fb9\u7f18\u8282\u70b9\u5904\u7406\u672c\u5730\u5316\u67e5\u8be2\uff0c\u53ef\u5c06\u533a\u57df\u7528\u6237\u7684\u5f80\u8fd4\u5ef6\u8fdf\u964d\u4f4e30\u201340%\uff0c\u540c\u65f6\u51cf\u8f7b\u6838\u5fc3\u6570\u636e\u4e2d\u5fc3\u7684\u5e26\u5bbd\u538b\u529b\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 \u7ed3\u8bed\uff1a\u62e5\u62b1\u641c\u7d22\u7684\u5e76\u884c\u8ba1\u7b97\u672a\u6765<\/strong><\/h2>\n GPU\u5df2\u4ece\u201c\u5c0f\u4f17\u52a0\u901f\u5668\u201d\u8f6c\u53d8\u4e3a\u73b0\u4ee3\u641c\u7d22\u5f15\u64ce\u67b6\u6784\u7684\u6838\u5fc3\u7ec4\u4ef6\uff0c\u4e3a\u901f\u5ea6\u3001\u6269\u5c55\u6027\u4e0e\u673a\u5668\u5b66\u4e60\u80fd\u529b\u5e26\u6765\u7a81\u7834\u6027\u63d0\u5347\u3002\u5c3d\u7ba1\u786c\u4ef6\u7ba1\u7406\u4e0e\u8f6f\u4ef6\u590d\u6742\u5ea6\u4ecd\u662f\u5f85\u89e3\u6311\u6218\uff0c\u4f46\u5e76\u884c\u5904\u7406\u7684\u4f18\u52bf\u5df2\u6bcb\u5eb8\u7f6e\u7591\u2014\u2014\u5c24\u5176\u5728\u4e2d\u56fd\u9999\u6e2f\u7684\u670d\u52a1\u5668\u79df\u7528\u4e0e\u670d\u52a1\u5668\u6258\u7ba1\u8fd9\u7c7b\u9ad8\u6027\u80fd\u751f\u6001\u4e2d\u3002\u968f\u7740\u6570\u636e\u91cf\u4e0e\u7b97\u6cd5\u590d\u6742\u5ea6\u7684\u6301\u7eed\u589e\u957f\uff0c\u638c\u63e1GPU\u6574\u5408\u6280\u672f\u7684\u7ec4\u7ec7\uff0c\u5c06\u5728\u63d0\u4f9b\u5feb\u901f\u3001\u7cbe\u51c6\u3001\u53ef\u6269\u5c55\u641c\u7d22\u4f53\u9a8c\u7684\u7ade\u4e89\u4e2d\u5360\u636e\u7edd\u5bf9\u4f18\u52bf\u3002CPU\u4e3b\u5bfc\u7684\u641c\u7d22\u65f6\u4ee3\u5df2\u843d\u5e55\uff0c\u5e76\u884c\u8ba1\u7b97\u7684\u672a\u6765\u5df2\u81f3\uff0c\u800cGPU\u6b63\u662f\u9a71\u52a8\u4e0b\u4e00\u4ee3\u641c\u7d22\u6280\u672f\u7684\u6838\u5fc3\u5f15\u64ce\u3002<\/p>\n \u65e0\u8bba\u4f60\u662f\u5728\u4f18\u5316\u5927\u578b\u4f01\u4e1a\u641c\u7d22\u5f15\u64ce\uff0c\u8fd8\u662f\u6784\u5efa\u5782\u76f4\u9886\u57df\u7684\u5c0f\u4f17\u641c\u7d22\u5de5\u5177\uff0c\u7406\u89e3GPU\u52a0\u901f\u90fd\u5df2\u4ece\u201c\u53ef\u9009\u9879\u201d\u53d8\u4e3a\u201c\u5fc5\u9009\u9879\u201d\u3002\u901a\u8fc7\u9a7e\u9a6d\u8fd9\u7c7b\u6280\u672f\uff0c\u4f60\u80fd\u6ee1\u8db3\u7528\u6237\u4e0e\u6570\u636e\u4e0d\u65ad\u589e\u957f\u7684\u9700\u6c42\uff0c\u786e\u4fdd\u641c\u7d22\u7cfb\u7edf\u5728\u65e5\u76ca\u590d\u6742\u7684\u6570\u5b57\u73af\u5883\u4e2d\u4fdd\u6301\u7ade\u4e89\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_column][vc_column_text css=””] \u5728\u4e0d\u65ad\u6f14\u8fdb\u7684\u641c\u7d22\u6280\u672f\u9886\u57df\uff0c\u5bf9\u66f4\u5feb\u3001\u66f4\u9ad8\u6548\u6570\u636e\u5904\u7406\u7684\u8ffd\u6c42\uff0c\u50ac\u751f\u4e86\u4e00\u9879\u5173\u952e\u8f6c\u53d8\uff1a\u5c06GPU\u6574\u5408\u5230\u641c\u7d22\u5f15\u64ce\u67b6\u6784\u4e2d\u3002\u968f\u7740\u6570\u636e\u91cf\u6fc0\u589e\u2014\u2014\u6bcf\u5929\u65b0\u589e\u6570\u5341\u4ebf\u4e2a\u9700\u7d22\u5f15\u7684\u7f51\u9875\u2014\u2014\u4e14\u673a\u5668\u5b66\u4e60\u6a21\u578b\u590d\u6742\u5ea6\u6301\u7eed\u63d0\u5347\uff0c\u4f20\u7edf\u57fa\u4e8eCPU\u7684\u7cfb\u7edf\u5df2\u96be\u4ee5\u8ddf\u4e0a\u8282\u594f\u3002\u800cGPU\u51ed\u501f\u5176\u72ec\u7279\u7684\u5e76\u884c\u5904\u7406\u80fd\u529b\uff0c\u6210\u4e3a\u4e86\u53d8\u9769\u6027\u89e3\u51b3\u65b9\u6848\uff0c\u5f7b\u5e95\u91cd\u5851\u4e86\u641c\u7d22\u5f15\u64ce\u5904\u7406\u6293\u53d6\u3001\u7d22\u5f15\u3001\u6392\u5e8f\u53ca\u673a\u5668\u5b66\u4e60\u4efb\u52a1\u7684\u65b9\u5f0f\u3002\u672c\u6587\u5c06\u6df1\u5165\u5256\u6790GPU\u52a0\u901f\u641c\u7d22\u5f15\u64ce\u7684\u6280\u672f\u539f\u7406\uff0c\u63a2\u8ba8\u5176\u5b9e\u9645\u5e94\u7528\u4f18\u52bf\uff0c\u5e76\u89e3\u7b54\u5de5\u7a0b\u5e08\u5728\u5b9e\u65bd\u8fd9\u7c7b\u7cfb\u7edf\u65f6\u9762\u4e34\u7684\u6311\u6218\uff0c\u5c24\u5176\u805a\u7126\u4e2d\u56fd\u9999\u6e2f\u7684\u670d\u52a1\u5668\u79df\u7528\u4e0e\u670d\u52a1\u5668\u6258\u7ba1\u8bbe\u65bd\u7b49\u9ad8\u6027\u80fd\u73af\u5883\u3002 [\/vc_column_text][\/vc_column][\/vc_row][vc_row el_class=”blog-detail-section”][vc_column][vc_column_text css=””] \u67b6\u6784\u4f18\u52bf\uff1a\u8ba1\u7b97\u8303\u5f0f\u4e2d\u7684CPU\u4e0eGPU\u5bf9\u6bd4 \u8981\u7406\u89e3GPU\u4e3a\u4f55\u80fd\u6210\u4e3a\u201c\u6e38\u620f\u89c4\u5219\u6539\u53d8\u8005\u201d\uff0c\u9996\u5148\u9700\u5bf9\u6bd4\u5176\u4e0e\u4f20\u7edfCPU\u7684\u8bbe\u8ba1\u5dee\u5f02\u3002CPU\u4e13\u4e3a\u4e32\u884c\u5904\u7406\u4f18\u5316\uff0c\u4ec5\u914d\u5907\u5c11\u91cf\u9ad8\u9891\u6838\u5fc3\uff0c\u64c5\u957f\u5904\u7406\u590d\u6742\u63a7\u5236\u903b\u8f91\uff1b\u800cGPU\u5219\u4ee5\u5e76\u884c\u4e3a\u6838\u5fc3\u8bbe\u8ba1\uff0c\u96c6\u6210\u6570\u5343\u4e2a\u4f4e\u9891\u6838\u5fc3\uff0c\u80fd\u9ad8\u6548\u6267\u884c\u201c\u5bf9\u591a\u7ec4\u6570\u636e\u91cd\u590d\u540c\u4e00\u6307\u4ee4\u201d\u7684\u8fd0\u7b97\u3002\u8fd9\u79cd\u7279\u6027\u4f7f\u5176\u6210\u4e3a\u73b0\u4ee3\u641c\u7d22\u5f15\u64ce\u4e2d\u201c\u8ba1\u7b97\u5bc6\u96c6\u578b\u3001\u6570\u636e\u5e76\u884c\u5316\u201d\u4efb\u52a1\u7684\u7406\u60f3\u9009\u62e9\u3002 \u4ee5\u4e0b\u662f\u4e24\u8005\u7684\u67b6\u6784\u5bf9\u6bd4\uff1a \u6307\u6807 CPU GPU \u6838\u5fc3\u6570\u91cf 8\u201364 2,000\u20134,000+ \u65f6\u949f\u9891\u7387 3\u20135 GHz 1\u20132 GHz \u5185\u5b58\u5e26\u5bbd 50\u2013100 GB\/s 400\u20131,000 GB\/s \u64c5\u957f\u7684\u4efb\u52a1 \u590d\u6742\u903b\u8f91\u63a7\u5236 \u5927\u89c4\u6a21\u5e76\u884c\u8ba1\u7b97 [\/vc_column_text][\/vc_column][\/vc_row][vc_row el_class=”blog-detail-section”][vc_column][vc_column_text css=””] GPU\u9a71\u52a8\u7684\u641c\u7d22\u5f15\u64ce\u8fd0\u4f5c\u6838\u5fc3\u9636\u6bb5 \u641c\u7d22\u5f15\u64ce\u7684\u8fd0\u4f5c\u5305\u542b\u6293\u53d6\u3001\u7d22\u5f15\u3001\u6392\u5e8f\u3001\u673a\u5668\u5b66\u4e60\u56db\u5927\u6838\u5fc3\u9636\u6bb5\uff0cGPU\u901a\u8fc7\u4e13\u95e8\u7684\u5e76\u884c\u5316\u7b56\u7565\uff0c\u4e3a\u6bcf\u4e2a\u9636\u6bb5\u63d0\u4f9b\u6027\u80fd\u589e\u76ca\u3002 1. \u7f51\u9875\u6293\u53d6\uff1a\u6570\u636e\u91c7\u96c6\u7684\u5e76\u884c\u5316\u5347\u7ea7 \u6293\u53d6\u9636\u6bb5\u9700\u4ece\u4e92\u8054\u7f51\u83b7\u53d6\u7f51\u9875\u6570\u636e\uff0c\u4f46\u53cd\u722c\u673a\u5236\u7684\u9650\u5236\u4e0e\u6d77\u91cfURL\u7684\u5904\u7406\u538b\u529b\uff0c\u4e00\u76f4\u662f\u8be5\u9636\u6bb5\u7684\u74f6\u9888\u3002GPU\u901a\u8fc7\u5206\u5e03\u5f0f\u5e76\u884cURL\u5904\u7406\uff0c\u5f7b\u5e95\u6539\u53d8\u4e86\u8fd9\u4e00\u5c40\u9762\uff1a \u57fa\u4e8e\u96c6\u7fa4\u7684\u4efb\u52a1\u8c03\u5ea6\uff1aGPU\u96c6\u7fa4\u5e76\u884c\u7ba1\u7406\u6570\u767e\u4e07\u4e2aURL\uff0c\u5c06\u6293\u53d6\u3001\u6e32\u67d3\u3001\u89e3\u6790\u4efb\u52a1\u5206\u914d\u5230\u6570\u5343\u4e2a\u6838\u5fc3\u4e0a\u540c\u6b65\u6267\u884c\u3002 JavaScript\u6e32\u67d3\u52a0\u901f\uff1a\u73b0\u4ee3\u7f51\u7ad9\u5927\u91cf\u4f9d\u8d56\u5ba2\u6237\u7aef\u811a\u672c\uff0c\u4f20\u7edf\u722c\u866b\u96be\u4ee5\u9ad8\u6548\u5904\u7406\uff0c\u800cGPU\u53ef\u5378\u8f7dJS\u6267\u884c\u4efb\u52a1\uff0c\u5728\u5b9e\u9645\u573a\u666f\u4e2d\u80fd\u5c06\u6e32\u67d3\u5ef6\u8fdf\u964d\u4f4e40%\u3002 \u7f51\u7edcI\/O\u4f18\u5316\uff1a\u5728\u4e2d\u56fd\u9999\u6e2f\u670d\u52a1\u5668\u79df\u7528\u8fd9\u7c7b\u9ad8\u5e26\u5bbd\u73af\u5883\u4e2d\uff0cGPU\u4e0e\u4f4e\u5ef6\u8fdf\u7f51\u7edc\u534f\u540c\uff0c\u6700\u5927\u7a0b\u5ea6\u51cf\u5c11\u722c\u866b\u8282\u70b9\u95f4\u7684\u6570\u636e\u4f20\u8f93\u5f00\u9500\u3002 2. \u7d22\u5f15\u6784\u5efa\uff1a\u89c4\u6a21\u5316\u6784\u5efa\u9ad8\u6548\u6570\u636e\u7ed3\u6784 \u7d22\u5f15\u9636\u6bb5\u9700\u5c06\u6293\u53d6\u5230\u7684\u6570\u636e\u8f6c\u6362\u4e3a\u53ef\u641c\u7d22\u7ed3\u6784\uff0c\u8fc7\u7a0b\u4e2d\u6d89\u53ca\u5927\u91cf\u6587\u672c\u5206\u6790\u4e0e\u6570\u636e\u8f6c\u6362\u64cd\u4f5c\u3002GPU\u51ed\u501f\u5e76\u884c\u5904\u7406\u80fd\u529b\uff0c\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u4e0e\u5012\u6392\u7d22\u5f15\u6784\u5efa\u4e2d\u8868\u73b0\u7a81\u51fa\uff1a NLP\u6d41\u6c34\u7ebf\u52a0\u901f\uff1a\u5206\u8bcd\u3001\u8bcd\u5e72\u63d0\u53d6\u3001\u8bcd\u6027\u6807\u6ce8\u7b49\u6838\u5fc3NLP\u4efb\u52a1\u5728GPU\u6838\u5fc3\u4e0a\u5e76\u884c\u5904\u7406\uff0c\u6bcf\u79d2\u53ef\u5904\u7406\u6570\u767e\u4e07\u4e2a\u8bcd\u5143\uff08token\uff09\u3002 \u5012\u6392\u7d22\u5f15\u5206\u7247\uff1a\u5c06\u7d22\u5f15\u7247\u6bb5\u5206\u5e03\u5f0f\u5b58\u50a8\u5230GPU\u5185\u5b58\u4e2d\uff0c\u641c\u7d22\u5f15\u64ce\u53ef\u51cf\u5c11\u78c1\u76d8I\/O\u74f6\u9888\u3002\u5185\u5b58\u5e26\u5bbd\u4f18\u52bf\u80fd\u5b9e\u73b0\u201c\u8bcd-\u6587\u6863\u77e9\u9635\u201d\u7684\u4e9a\u6beb\u79d2\u7ea7\u8bbf\u95ee\uff0c\u8fd9\u5bf9\u5b9e\u65f6\u67e5\u8be2\u5904\u7406\u81f3\u5173\u91cd\u8981\u3002 \u538b\u7f29\u7b97\u6cd5\u4f18\u5316\uff1aGPU\u5bf9\u7d22\u5f15\u6570\u636e\u8fdb\u884c\u65e0\u635f\u538b\u7f29\u4f18\u5316\uff0c\u5728\u4fdd\u8bc1\u5b58\u50a8\u6548\u7387\u7684\u540c\u65f6\uff0c\u517c\u987e\u67e5\u8be2\u6267\u884c\u65f6\u7684\u5feb\u901f\u89e3\u538b\u9700\u6c42\u3002 3. \u7ed3\u679c\u6392\u5e8f\uff1a\u5b9e\u65f6\u76f8\u5173\u6027\u8ba1\u7b97 \u6392\u5e8f\u9636\u6bb5\u9700\u6839\u636e\u67e5\u8be2\u76f8\u5173\u6027\u786e\u5b9a\u7ed3\u679c\u987a\u5e8f\uff0c\u5982\u4eca\u8be5\u8fc7\u7a0b\u5df2\u7531\u673a\u5668\u5b66\u4e60\u6a21\u578b\u4e3b\u5bfc\u3002GPU\u4e3a\u590d\u6742\u6a21\u578b\u63d0\u4f9b\u8fd1\u5b9e\u65f6\u7684\u63a8\u7406\u80fd\u529b\uff1a \u8bcd\u9891-\u9006\u6587\u6863\u9891\u7387\uff08TF-IDF\uff09\u7b49\u5411\u91cf\u7a7a\u95f4\u6a21\u578b\uff0c\u4f9d\u8d56\u4e8e\u968f\u6570\u636e\u96c6\u89c4\u6a21\u7ebf\u6027\u589e\u957f\u7684\u77e9\u9635\u8fd0\u7b97\u3002GPU\u5229\u7528\u5e76\u884c\u6838\u5fc3\u540c\u6b65\u8ba1\u7b97\u4f59\u5f26\u76f8\u4f3c\u5ea6\u4e0e\u76f8\u5173\u6027\u5206\u6570\uff0c\u5c06\u8ba1\u7b97\u65f6\u95f4\u4ece\u6beb\u79d2\u7ea7\u538b\u7f29\u81f3\u5fae\u79d2\u7ea7\u3002\u5bf9\u4e8eBERT\u4e0eTransformer\u7cfb\u5217\u7b49\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\uff0cGPU\u652f\u6301\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u4e0e\u63a8\u7406\uff1a FP16\u6d6e\u70b9\u8fd0\u7b97\uff1a\u91c7\u7528\u534a\u7cbe\u5ea6\u6d6e\u70b9\u6570\uff0cGPU\u5904\u7406\u795e\u7ecf\u7f51\u7edc\u5c42\u7684\u901f\u5ea6\u53ef\u8fbeCPU\u76842\u500d\uff0c\u4e14\u4e0d\u4f1a\u5bfc\u81f4\u663e\u8457\u7cbe\u5ea6\u635f\u5931\u3002 \u5f20\u91cf\u6838\u5fc3\uff08Tensor Core\uff09\u4f18\u5316\uff1a\u73b0\u4ee3GPU\u4e2d\u7684\u4e13\u7528\u77e9\u9635\u5904\u7406\u5355\u5143\uff0c\u53ef\u5c06\u6df1\u5ea6\u5b66\u4e60\u8fd0\u7b97\u901f\u5ea6\u63d0\u534750%\uff0c\u8fd9\u5bf9\u9ad8\u6d41\u91cf\u573a\u666f\u4e0b\u7684\u5b9e\u65f6\u6392\u5e8f\u81f3\u5173\u91cd\u8981\u3002 4. \u673a\u5668\u5b66\u4e60\uff1a\u8bad\u7ec3\u4e0b\u4e00\u4ee3\u7b97\u6cd5 […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":1,"featured_media":27242,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[62],"tags":[8826,8467,7757,3310,2684],"class_list":["post-27246","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-hong-kong-dedicated-server-sc","tag-search-engine-technology-sc","tag-heterogeneous-computing-sc","tag-distributed-systems-sc","tag-gpu-acceleration-sc","tag-parallel-computing-sc"],"acf":[],"yoast_head":"\n