GPU \u548c GPGPU\uff1a\u6838\u5fc3\u5dee\u7570<\/strong><\/h2>\n

GPU \u7684\u4f5c\u7528<\/h3>\n

GPU \u662f\u70ba\u5e73\u884c\u5de5\u4f5c\u800c\u8a2d\u8a08\u7684\u8655\u7406\u5668\u3002\u5b83\u5305\u542b\u4e00\u584a\u6574\u5408\u6240\u6709\u8655\u7406\u55ae\u5143\u7684\u6676\u7247\uff08die\uff09\uff0c\u4ee5\u53ca\u70ba\u773e\u591a\u6838\u5fc3\u540c\u6642\u63d0\u4f9b\u8cc7\u6599\u7684\u986f\u793a\u8a18\u61b6\u9ad4\uff08VRAM\uff09\u3002\u6210\u5343\u4e0a\u842c\u500b\u5c0f\u6838\u5fc3\uff08\u4f8b\u5982 CUDA \u6838\u5fc3\u6216\u4e32\u6d41\u8655\u7406\u5668\uff09\u53ef\u4ee5\u540c\u6642\u8655\u7406\u591a\u8def\u8cc7\u6599\u6d41\u3002\u6563\u71b1\u7cfb\u7d71\u9632\u6b62\u56e0\u904e\u71b1\u800c\u964d\u983b\uff08thermal throttling\uff09\uff0c\u96fb\u6e90\u8abf\u7bc0\u5247\u78ba\u4fdd\u8f38\u9001\u91cf\u7a69\u5b9a\u3002\u9019\u4e9b\u90e8\u5206\u5171\u540c\u69cb\u6210\u4f60\u90e8\u7f72\u5728\u6a5f\u67b6\u4e2d\u7684 GPU \u786c\u9ad4\u3002<\/p>\n

GPU \u904b\u7b97\u6703\u5c07\u90e8\u5206\u8a08\u7b97\u4efb\u52d9\u5f9e CPU \u4e0a\u5378\u8f09\u4e0b\u4f86\uff0c\u800c GPU \u7684\u6578\u767e\u500b\u6838\u5fc3\u53ef\u4ee5\u900f\u904e\u5e73\u884c\u8655\u7406\u5728\u6578\u79d2\u5167\u5b8c\u6210\u6210\u5343\u4e0a\u842c\u9805\u4efb\u52d9\u3002\u5728\u5e73\u884c\u8655\u7406\u4e2d\uff0c\u8cc7\u6599\u96c6\u88ab\u5c0e\u5165 GPU \u7684\u5404\u500b\u8655\u7406\u6838\u5fc3\u4e26\u540c\u6642\u6c42\u89e3\uff0c\u5f9e\u800c\u986f\u8457\u63d0\u5347\u6548\u80fd\uff0c\u800c CPU \u5247\u8ca0\u8cac\u5176\u9918\u61c9\u7528\u7a0b\u5f0f\u3002<\/p>\n

GPU \u67b6\u69cb\u8207 CPU \u622a\u7136\u4e0d\u540c\u3002\u4e00\u9846\u4f3a\u670d\u5668 CPU \u901a\u5e38\u4ee5 4\u20135 GHz \u7684\u983b\u7387\u904b\u884c\u5927\u7d04 128\u2013192 \u500b\u6838\u5fc3\uff0c\u4e26\u914d\u6709\u8f03\u6df1\u7684\u5feb\u53d6\uff1b\u800c GPU \u5247\u628a\u6210\u5343\u4e0a\u842c\u689d\u7b97\u8853\u7ba1\u7dda\u5206\u5e03\u5230\u773e\u591a\u4e32\u6d41\u591a\u8655\u7406\u5668\uff08SM\uff09\u4e0a\uff0c\u4ee5\u8f03\u4f4e\u983b\u7387\u904b\u884c\u3002\u5b83\u900f\u904e\u5927\u898f\u6a21\u591a\u57f7\u884c\u7dd2\u8207\u4f86\u81ea HBM3 \u6216 HBM3e \u7684\u9ad8\u8a18\u61b6\u9ad4\u983b\u5bec\u4f86\u96b1\u85cf\u5ef6\u9072\u3002\u9019\u7a2e\u8a2d\u8a08\u89e3\u91cb\u4e86\u70ba\u4f55 GPU \u5728\u91cd\u8907\u4e14\u8cc7\u6599\u5bc6\u96c6\u578b\u7684\u4efb\u52d9\u4e0a\u6548\u80fd\u64f4\u5c55\u660e\u986f\u3002<\/p>\n

GPGPU \u7684\u6db5\u7fa9<\/h3>\n

GPGPU \u6307\u7684\u662f\u5c07\u5716\u5f62\u8655\u7406\u5668\u7528\u65bc\u901a\u7528\u904b\u7b97\u3002\u9019\u500b\u6982\u5ff5\u5728 2006 \u5e74\u958b\u59cb\u6210\u5f62\uff0c\u7576\u6642 CUDA \u5141\u8a31\u958b\u767c\u8005\u5c07 GPU \u7528\u65bc\u975e\u5716\u5f62\u5de5\u4f5c\u3002\u4e4b\u5f8c\u7684\u4e00\u4e9b\u91cc\u7a0b\u7891\uff0c\u5982 2017 \u5e74\u63a8\u51fa\u7684 Tensor Core \u4ee5\u53ca 2022 \u5e74\u8d77\u9762\u5411 Transformer \u7684\u6700\u4f73\u5316\u8a2d\u8a08\uff0c\u63a8\u52d5\u4e86\u901a\u7528 GPU \u904b\u7b97\u5728\u6a5f\u5668\u5b78\u7fd2\u4e2d\u7684\u666e\u53ca\u3002<\/p>\n

\u4f60\u53ef\u4ee5\u900f\u904e\u591a\u7a2e\u7a0b\u5f0f\u8a2d\u8a08\u6a21\u578b\u9032\u5165 GPGPU \u6a21\u5f0f\u3002CUDA \u70ba NVIDIA \u786c\u9ad4\u63d0\u4f9b\u4e86\u76f4\u63a5\u63a7\u5236\uff1bOpenCL \u63d0\u4f9b\u8de8\u5e73\u53f0\u6a19\u6e96\uff1bOpenACC \u548c OpenMP \u5247\u652f\u63f4\u4ee5\u6307\u793a\uff08directive\uff09\u70ba\u57fa\u790e\u7684\u5e73\u884c\u7a0b\u5f0f\u78bc\u3002\u50cf CuPy\u3001Numba \u9019\u985e Python \u51fd\u5f0f\u5eab\u5247\u964d\u4f4e\u4e86\u5165\u9580\u9580\u6abb\u3002<\/p>\n

\u9019\u9805\u5340\u5225\u5c0d\u4f60\u7684\u898f\u5283\u5f88\u91cd\u8981\u3002GPU \u662f\u90a3\u584a\u6676\u7247\uff0c\u800c GPGPU \u5247\u662f\u5c07\u5176\u7528\u65bc AI \u7684\u5be6\u52d9\u65b9\u5f0f\u3002\u9019\u7a2e\u5be6\u52d9\u5df2\u7d93\u64f4\u5c55\u5230 GPU \u89e3\u8026\uff08gpu disaggregation\uff09\uff1a\u4f60\u53ef\u4ee5\u5728\u7bc0\u9ede\u4e4b\u9593\u6c60\u5316\u52a0\u901f\u5361\u3002\u8cc7\u6599\u4e2d\u5fc3\u7d1a\u7684 GPU \u89e3\u8026\u8b93\u8a31\u591a\u5de5\u4f5c\u8ca0\u8f09\u5171\u4eab\u540c\u4e00\u5f35 GPU \u8cc7\u6e90\u7db2\u8def\u3002\u85c9\u7531 GPU \u89e3\u8026\uff0c\u4f60\u5c07\u904b\u7b97\u8207\u8a18\u61b6\u9ad4\u8cc7\u6e90\u5206\u96e2\uff0c\u5f9e\u800c\u63d0\u9ad8\u4f7f\u7528\u7387\u3002\u5718\u968a\u63a1\u7528 GPU \u89e3\u8026\u4f86\u8b93\u7b97\u529b\u5bb9\u91cf\u8207\u9700\u6c42\u5339\u914d\u3002\u6bcf\u4e00\u7a2e GPU \u89e3\u8026\u8a2d\u8a08\u90fd\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u4ee5\u5ef6\u9072\u63db\u53d6\u5f48\u6027\u3002\u9762\u5411\u8cc7\u6599\u4e2d\u5fc3\u7684\u5927\u898f\u6a21 GPU \u89e3\u8026\u9084\u80fd\u7c21\u5316\u5347\u7d1a\uff0c\u56e0\u70ba\u4f60\u53ef\u4ee5\u6c70\u63db\u8cc7\u6e90\u6c60\u800c\u4e0d\u662f\u6574\u53f0\u4f3a\u670d\u5668\u3002<\/p>\n

\u5c0d\u65bc\u65e5\u672c\u7684 AI \u4f3a\u670d\u5668\u4f86\u8aaa\uff0c\u9019\u9805\u8f49\u8b8a\u6703\u6539\u8b8a\u6210\u672c\u8207\u96fb\u529b\u898f\u5283\u3002\u4f60\u6703\u4f9d\u64da\u8a13\u7df4\u7684\u5cf0\u503c\u9700\u6c42\u4f86\u898f\u5283 GPU \u5bb9\u91cf\uff0c\u7136\u5f8c\u518d\u5c07\u9019\u90e8\u5206\u5bb9\u91cf\u5171\u4eab\u7d66\u63a8\u8ad6\u4efb\u52d9\u3002GPU \u8207 GPGPU \u7684\u7d44\u5408\u65e2\u7d66\u4f60\u786c\u9ad4\u57fa\u790e\uff0c\u4e5f\u63d0\u4f9b\u4e86\u71df\u904b\u6a21\u578b\u3002\u7406\u89e3 GPU \u8207 GPGPU \u5982\u4f55\u5354\u540c\uff0c\u6709\u52a9\u65bc\u4f60\u70ba\u6bcf\u7a2e\u5de5\u4f5c\u8ca0\u8f09\u9078\u64c7\u5408\u9069\u7684\u90e8\u7f72\u65b9\u5f0f\u3002<\/p>\n

GPGPU \u5982\u4f55\u652f\u6490 AI \u5de5\u4f5c\u8ca0\u8f09<\/strong><\/h2>\n

\u70ba\u6a21\u578b\u63d0\u4f9b\u5e73\u884c\u904b\u7b97<\/h3>\n

\u795e\u7d93\u7db2\u8def\u7684\u5404\u5c64\u672c\u8cea\u4e0a\u90fd\u53ef\u4ee5\u5316\u7d04\u70ba\u77e9\u9663\u4e58\u6cd5\uff1aY = X \u00d7 W + b\u3002\u4e00\u500b CUDA kernel \u6703\u70ba\u6bcf\u500b\u8f38\u51fa\u5143\u7d20\u5206\u914d\u4e00\u500b\u57f7\u884c\u7dd2\u3002\u6bcf\u500b\u57f7\u884c\u7dd2\u6703\u53d6 A \u7684\u4e00\u5217\u8207 B \u7684\u4e00\u884c\u505a\u5167\u7a4d\uff0c\u518d\u5728\u8ff4\u5708\u4e2d\u7d2f\u52a0\u6c42\u548c\u3002\u5c0d\u65bc\u4e00\u5c64\u96b1\u85cf\u5c64\u4f86\u8aaa\uff0c\u82e5\u6279\u6b21\u5927\u5c0f\u70ba 64\uff0c\u795e\u7d93\u5143\u6578\u70ba 128\uff0c\u5247\u6703\u555f\u52d5 64 \u00d7 128 = 8,192 \u500b\u57f7\u884c\u7dd2\uff0c\u5f9e\u800c\u5e73\u884c\u8a08\u7b97\u6240\u6709\u8f38\u51fa\u3002\u50cf ReLU \u9019\u6a23\u7684\u555f\u7528\u51fd\u5f0f\u4e5f\u9075\u5faa\u540c\u6a23\u7684\u6a21\u5f0f\uff0c\u6bcf\u500b\u57f7\u884c\u7dd2\u5c0d\u4e00\u500b\u5143\u7d20\u5957\u7528 max(0, x)\u3002<\/p>\n

\u9019\u7a2e\u5c0d\u61c9\u65b9\u5f0f\u89e3\u91cb\u4e86\u70ba\u4ec0\u9ebc GPGPU \u52a0\u901f\u975e\u5e38\u9069\u5408\u6a5f\u5668\u5b78\u7fd2\u3002GPU \u63a1\u7528 SIMD \u548c SIMT \u57f7\u884c\u6a21\u578b\uff0c\u56e0\u6b64\u540c\u4e00\u64cd\u4f5c\u53ef\u4ee5\u540c\u6642\u5957\u7528\u5230\u6210\u5343\u4e0a\u842c\u500b\u795e\u7d93\u5143\u4e0a\u3002CUDA \u6838\u5fc3\u8ca0\u8cac\u901a\u7528\u6d6e\u9ede\u904b\u7b97\uff0c\u800c Tensor Core \u5247\u4ee5\u6df7\u5408\u7cbe\u5ea6\u57f7\u884c\u77e9\u9663\u904b\u7b97\uff0c\u662f\u5c08\u9580\u70ba AI \u8a13\u7df4\u800c\u751f\u7684\u3002\u5c0d\u65bc\u6df1\u5ea6\u5b78\u7fd2\u4f86\u8aaa\uff0cTensor Core \u6578\u91cf\u8207 VRAM \u5bb9\u91cf\u5f80\u5f80\u6bd4 CUDA \u6838\u5fc3\u6578\u91cf\u66f4\u95dc\u9375\u3002<\/p>\n

GPU \u63d0\u4f9b\u4e86\u5927\u91cf\u9069\u5408\u77e9\u9663\u548c\u5411\u91cf\u904b\u7b97\u7684\u5e73\u884c\u904b\u7b97\u6838\u5fc3\uff0c\u76f8\u8f03\u50b3\u7d71 CPU\uff0c\u5728\u540c\u6a23\u6642\u9593\u5167\u80fd\u5b8c\u6210\u66f4\u591a\u8a08\u7b97\uff0c\u5927\u5e45\u52a0\u901f\u8a13\u7df4\u8207\u63a8\u8ad6\u6d41\u7a0b\u3002<\/p>\n

\u8a13\u7df4\u8207\u63a8\u8ad6\u7684\u5dee\u7570\u9700\u6c42<\/h3>\n

\u8a13\u7df4\u8207\u63a8\u8ad6\u5c0d\u786c\u9ad4\u7684\u8981\u6c42\u622a\u7136\u4e0d\u540c\u3002AI \u8a13\u7df4\u9700\u8981\u540c\u6642\u4fdd\u7559\u53c3\u6578\u3001\u6700\u4f73\u5316\u5668\u72c0\u614b\u3001\u68af\u5ea6\u548c\u4e2d\u4ecb\u555f\u7528\u503c\u3002\u4e00\u500b 70B \u6a21\u578b\u5728\u6df7\u5408\u7cbe\u5ea6\u4e0b\u8a13\u7df4\u5927\u7d04\u9700\u8981 840 GB \u8a18\u61b6\u9ad4\uff0c\u800c\u63a8\u8ad6\u5927\u7d04\u53ea\u9700\u8981 140 GB\uff0c\u5bb9\u91cf\u9700\u6c42\u7e2e\u6e1b\u7d04 6 \u500d\u3002\u4e00\u500b\u5be6\u7528\u7684 GPU \u8a13\u7df4\u7d93\u9a57\u503c\u662f\uff1a\u6bcf 1B \u53c3\u6578\u5927\u7d04\u9700\u8981 16 GB \u986f\u793a\u8a18\u61b6\u9ad4\u3002\u7528\u65bc\u8a13\u7df4\u7684 GPU \u9084\u9700\u8981\u985e\u4f3c NVLink\u3001InfiniBand \u9019\u985e\u9ad8\u983b\u5bec\u4e92\u9023\u4f86\u9032\u884c\u68af\u5ea6\u540c\u6b65\u3002<\/p>\n

\u63a8\u8ad6\u53ea\u57f7\u884c\u524d\u5411\u50b3\u64ad\uff0c\u4e0d\u505a\u53cd\u5411\u50b3\u64ad\u3002\u5176\u89e3\u78bc\u968e\u6bb5\u5f80\u5f80\u53d7\u5236\u65bc\u8a18\u61b6\u9ad4\u983b\u5bec\uff0c\u56e0\u70ba\u6bcf\u7522\u751f\u4e00\u500b token\uff0c\u90fd\u8981\u91cd\u8907\u8b80\u53d6\u6b0a\u91cd\u548c\u5b8c\u6574\u7684 KV \u5feb\u53d6\u3002\u5728\u9019\u88e1\uff0c\u8a18\u61b6\u9ad4\u983b\u5bec\u8207\u5bb9\u91cf\u7684\u91cd\u8981\u6027\u5f80\u5f80\u4e0d\u4e9e\u65bc\u904b\u7b97\u6307\u6a19\u3002\u4f60\u53ef\u4ee5\u5728\u540c\u4e00\u584a GPU \u4e0a\u57f7\u884c\u8a31\u591a\u63a8\u8ad6\u4efb\u52d9\uff0c\u6216\u8005\u9078\u64c7 CPU \u57f7\u884c\u3002\u8cc7\u6599\u4e2d\u5fc3\u7d1a GPU \u89e3\u8026\u53ef\u4ee5\u5e6b\u52a9\u4f60\u5728\u8a13\u7df4\u8207\u63a8\u8ad6\u4e4b\u9593\u5171\u4eab\u6c60\u5316\u7b97\u529b\u3002GPU \u89e3\u8026\u80fd\u63d0\u5347 AI \u5de5\u4f5c\u8ca0\u8f09\u7684\u7b97\u529b\u4f7f\u7528\u7387\uff0c\u4e5f\u8b93 GPU \u89e3\u8026\u5728\u5347\u7d1a\u6642\u66f4\u7c21\u55ae\u3002\u5718\u968a\u900f\u904e GPU \u89e3\u8026\u8b93\u7b97\u529b\u5bb9\u91cf\u8207\u5be6\u969b\u9700\u6c42\u5339\u914d\uff0c\u4e26\u5229\u7528 GPU \u89e3\u8026\u5728 AI \u8a13\u7df4\u548c\u63a8\u8ad6\u4e2d\u7dad\u6301 GPU \u6548\u80fd\u7684\u7a69\u5b9a\u53ef\u9810\u671f\u3002\u898f\u5283 GPU \u52a0\u901f\u6642\uff0c\u4ee5\u8a13\u7df4\u7684\u5cf0\u503c\u9700\u6c42\u70ba\u57fa\u6e96\uff0c\u7136\u5f8c\u5c07\u9019\u90e8\u5206\u5bb9\u91cf\u91cd\u8907\u7528\u65bc AI \u6a21\u578b\u7684\u63a8\u8ad6\u3002<\/p>\n

AI \u5de5\u4f5c\u8ca0\u8f09\u4e2d\u7684 CPU \u8207 GPU<\/strong><\/h2>\n

CPU vs GPU\uff1a\u57fa\u7dda\u5c0d\u6bd4<\/h3>\n

\u95dc\u65bc CPU vs GPU \u5728 AI \u4e2d\u5982\u4f55\u53d6\u6368\uff0c\u9996\u5148\u8981\u770b\u67b6\u69cb\u3002CPU \u901a\u5e38\u53ea\u6709\u5c11\u91cf\u8907\u96dc\u6838\u5fc3\uff0c\u5177\u5099\u5206\u652f\u9810\u6e2c\u8207\u4e82\u5e8f\u57f7\u884c\u80fd\u529b\uff0c\u64c5\u9577\u9700\u8981\u4f4e\u5ef6\u9072\u7684\u5e8f\u5217\u5316\u3001\u6c7a\u7b56\u5bc6\u96c6\u578b\u5de5\u4f5c\u3002GPU \u5247\u64c1\u6709\u6210\u5343\u4e0a\u842c\u500b\u7c21\u55ae\u6838\u5fc3\uff0c\u900f\u904e\u5c0d\u5927\u898f\u6a21\u8cc7\u6599\u96c6\u57f7\u884c\u76f8\u540c\u6307\u4ee4\u4f86\u7372\u5f97\u9ad8\u8f38\u9001\u91cf\u3002\u6bcf\u500b\u4e32\u6d41\u591a\u8655\u7406\u5668\uff08SM\uff09\u5c31\u6574\u5408\u4e86\u6578\u767e\u500b\u6838\u5fc3\uff0c\u4f7f\u5f97\u6578\u5341\u842c\u689d\u57f7\u884c\u7dd2\u53ef\u4ee5\u540c\u6642\u5728\u98db\u884c\u4e2d\u57f7\u884c\u3002<\/p>\n

\n\n\n\n\n<\/colgroup>\n\n\n\n\n\n\n

\u7dad\u5ea6<\/th>\n	CPU<\/th>\n	GPU<\/th>\n<\/tr>\n
\u6838\u5fc3\u8a2d\u8a08<\/td>\n	\u5c11\u91cf\u8907\u96dc\u6838\u5fc3\uff0c\u5177\u5099\u5206\u652f\u9810\u6e2c<\/td>\n	\u6210\u5343\u4e0a\u842c\u500b\u7c21\u55ae\u6578\u5b78\u6838\u5fc3<\/td>\n<\/tr>\n
\u5e73\u884c\u6027<\/td>\n	\u4f4e\u5ef6\u9072\u5e8f\u5217\u5316\u5de5\u4f5c<\/td>\n	\u900f\u904e SIMD \u548c SIMT \u63d0\u4f9b\u5de8\u5927\u8f38\u9001\u91cf<\/td>\n<\/tr>\n
\u8a18\u61b6\u9ad4<\/td>\n	\u63a7\u5236\u5668\u504f\u5411\u4f4e\u5ef6\u9072\uff0c\u5927\u7d04 350 GB\/s<\/td>\n	\u900f\u904e HBM3 \u6216 HBM3e \u63d0\u4f9b 3.35\u20138 TB\/s \u7684\u9ad8\u983b\u5bec<\/td>\n<\/tr>\n
\u5c08\u7528\u55ae\u5143<\/td>\n	\u6574\u6578\u3001\u6d6e\u9ede\u3001SIMD \u55ae\u5143<\/td>\n	Tensor Core\u3001RT Core\u3001\u7d0b\u7406\u55ae\u5143<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u9019\u7a2e\u8a18\u61b6\u9ad4\u983b\u5bec\u5dee\u7570\u9a45\u52d5\u4e86\u5927\u90e8\u5206\u6c7a\u7b56\u3002\u5728\u5169\u985e\u6676\u7247\u4e0a\uff0c\u5206\u5c64\u7684\u8a18\u61b6\u9ad4\u5c64\u7d1a\u90fd\u7528\u65bc\u8b93\u88ab\u53cd\u8986\u4f7f\u7528\u7684\u8cc7\u6599\u5118\u91cf\u9760\u8fd1\u57f7\u884c\u55ae\u5143\uff1b\u4e0d\u540c\u7684\u662f\uff0cGPU \u6bcf\u79d2\u53ef\u4ee5\u642c\u904b\u7684\u8cc7\u6599\u8981\u591a\u5f97\u591a\u3002<\/p>\n \u6548\u80fd\u3001\u9069\u914d\u5ea6\u8207\u6cd5\u898f\u9075\u5faa<\/h3>\n \u5373\u4fbf\u662f 70B \u7684\u5927\u6a21\u578b\uff0c\u4f60\u4e5f\u53ef\u4ee5\u900f\u904e 4bit \u91cf\u5316\u5728 CPU \u4e0a\u57f7\u884c\u3002\u4f46\u6b64\u6642\u7522\u751f\u901f\u5ea6\u5f80\u5f80\u6703\u4f4e\u65bc\u6bcf\u79d2 10 token\uff0c\u5728\u65e5\u5e38\u4f7f\u7528\u4e2d\u6703\u660e\u986f\u611f\u89ba\u504f\u6162\u3002\u5be6\u52d9\u4e0a\uff0cCPU \u66f4\u9069\u5408\u57f7\u884c 4bit \u91cf\u5316\u7684 8B \u6a21\u578b\uff0c\u4ee5\u53ca\u63a1\u7528 bfloat16 \u5230 8bit \u7cbe\u5ea6\u5340\u9593\u7684 4B \u6a21\u578b\u3002\u4e00\u500b\u6bcf\u5929\u53ea\u6709 5\u201310 \u6b21\u8acb\u6c42\u3001\u4f7f\u7528\u8005\u53ef\u63a5\u53d7 10\u201320 \u79d2\u7b49\u5f85\u6642\u9593\u7684\u5c0f\u4f01\u696d\u804a\u5929\u6a5f\u5668\u4eba\uff0c\u8dd1\u5728\u91cf\u5316\u5f8c\u7684 Mistral 7B \u4e0a\u5c31\u8db3\u5920\uff0c\u6210\u672c\u5927\u7d04\u80fd\u6bd4 GPU \u65b9\u6848\u4f4e 90%\u3002\u4f46\u82e5\u662f\u4e00\u500b\u6bcf\u5929\u9700\u8981\u8655\u7406 1,000+ \u8acb\u6c42\u4e26\u8981\u6c42 5 \u79d2\u4ee5\u5167\u56de\u61c9\u7684\u516c\u5171 API\uff0c\u5247\u9700\u8981 GPU \u4f3a\u670d\u5668\u6216\u5206\u6563\u5f0f\u53e2\u96c6\u3002\u4ee5\u4e0b\u60c5\u6cc1\u53ef\u4ee5\u512a\u5148\u8003\u616e\u53ea\u7528 CPU\uff1a\u4f60\u5728\u5efa\u7f6e\u5167\u90e8\u5de5\u5177\uff1b\u65e5\u8acb\u6c42\u91cf\u4f4e\u65bc\u7d04 100 \u6b21\uff1b\u4f7f\u7528\u8005\u80fd\u63a5\u53d7\u6578\u79d2\u5ef6\u9072\uff1b\u6210\u672c\u662f\u6838\u5fc3\u8003\u91cf\uff1b\u4ee5\u53ca\u6a21\u578b\u898f\u6a21\u5c0f\u65bc 10B \u53c3\u6578\u3002\u8cc7\u6599\u524d\u8655\u7406\u3001\u7279\u5fb5\u5de5\u7a0b\u4ee5\u53ca\u5728\u4f4e\u50f9\u7bc0\u9ede\u4e0a\u505a\u6a6b\u5411\u64f4\u5c55\u4e5f\u901a\u5e38\u66f4\u504f\u5411 CPU\u3002<\/p>\n \u5c0d\u65bc\u5927\u898f\u6a21 AI \u8a13\u7df4\u4f86\u8aaa\uff0cGPGPU \u5247\u662f\u525b\u9700\u3002\u4e00\u500b 70B \u6a21\u578b\u5728\u8a13\u7df4\u968e\u6bb5\u5927\u7d04\u9700\u8981 840 GB \u8a18\u61b6\u9ad4\uff0c\u800c\u63a8\u8ad6\u5247\u9700\u8981\u7d04 140 GB\u3002Sakura Internet \u7684 GPU \u670d\u52d9\u53ef\u4ee5\u628a 8 \u5f35 GPU \u7576\u4f5c\u4e00\u500b\u5927 GPU \u4f86\u4f7f\u7528\uff0c\u5171\u4eab\u4e00\u6574\u500b\u8a18\u61b6\u9ad4\u6c60\uff0c\u9069\u5408\u5927\u6a21\u578b\u8a13\u7df4\u548c\u63a8\u8ad6\u3002GMO GPU Cloud \u5247\u70ba\u5e0c\u671b\u5728\u96f2\u7aef\u57f7\u884c GPU \u5de5\u4f5c\u8ca0\u8f09\u7684\u5718\u968a\u63d0\u4f9b GPU \u5373\u670d\u52d9\uff08GPU-as-a-service\uff09\u3002\u8cc7\u6599\u4e2d\u5fc3\u7d1a GPU \u89e3\u8026\u53ef\u4ee5\u5728\u7bc0\u9ede\u4e4b\u9593\u6c60\u5316\u52a0\u901f\u5361\u8cc7\u6e90\uff0c\u5f9e\u800c\u63d0\u5347 GPU \u4f7f\u7528\u7387\u3002\u5718\u968a\u5229\u7528 GPU \u89e3\u8026\u8b93\u7b97\u529b\u5bb9\u91cf\u8207\u9700\u6c42\u66f4\u597d\u5339\u914d\uff0c\u4e26\u900f\u904e GPU \u89e3\u8026\u4f86\u7c21\u5316\u5347\u7d1a\u6d41\u7a0b\u3002\u5927\u898f\u6a21\u7684\u89e3\u8026 GPU \u8cc7\u6e90\u6c60\u9084\u80fd\u8b93\u4e3b\u6a5f\u4e4b\u9593\u7684 GPU \u8cc7\u6e90\u8abf\u914d\u66f4\u52a0\u5bb9\u6613\u3002\u5c07 GPU \u89e3\u8026\u898f\u5283\u570d\u7e5e\u5cf0\u503c\u9700\u6c42\u5c55\u958b\uff0c\u7136\u5f8c\u91cd\u8907\u5229\u7528\u9019\u90e8\u5206\u5bb9\u91cf\u3002GPU \u89e3\u8026\u80fd\u8b93 GPU \u5728 AI \u5de5\u4f5c\u8ca0\u8f09\u4e2d\u7684\u6548\u80fd\u4fdd\u6301\u53ef\u9810\u671f\u3002<\/p>\n \u6280\u8853\u7248\u5716\u5728\u4e0d\u65b7\u6f14\u9032\u3002NVIDIA \u7684 Vera Rubin \u5e73\u53f0\u9396\u5b9a\u5927\u898f\u6a21\u8a13\u7df4\u4ee5\u53ca\u667a\u6167\u9ad4\uff0f\u63a8\u7406\u5f37\u5316\u6a21\u578b\uff1b\u57fa\u65bc NPU \u7684\u4f3a\u670d\u5668\u5247\u66f4\u9069\u5408\u5728\u908a\u7de3\u5074\u505a\u53d7\u96fb\u529b\u9650\u5236\u7684\u63a8\u8ad6\u3002\u70ba\u6bcf\u7a2e\u9078\u9805\u627e\u5230\u5408\u9069\u7684\u9069\u7528\u60c5\u5883\uff0c\u6b0a\u8861\u96fb\u529b\u3001\u6563\u71b1\u3001\u8cc7\u6599\u5728\u5730\u5316\u8207\u51fa\u53e3\u7ba1\u5236\uff0c\u4e26\u96a8\u8457\u6a21\u578b\u548c\u6d41\u91cf\u8b8a\u5316\uff0c\u6301\u7e8c\u91cd\u5be9 CPU vs GPU \u7684\u9078\u64c7\u3002<\/p>\n \u5728\u65e5\u672c\u5982\u4f55\u9078\u64c7 GPU \u6216 GPGPU<\/strong><\/h2>\n \u4ee5\u5de5\u4f5c\u8ca0\u8f09\u70ba\u57fa\u790e\u7684\u6c7a\u7b56\u6846\u67b6<\/h3>\n \u9996\u5148\u8981\u770b\u4f60\u7684\u5de5\u4f5c\u8ca0\u8f09\u5f62\u614b\u3002GPGPU \u7684\u50f9\u503c\u5728\u65bc\uff1a\u4efb\u52d9\u80fd\u62c6\u5206\u6210\u6210\u5343\u4e0a\u842c\u689d\u7368\u7acb\u57f7\u884c\u7dd2\u3001\u8cc7\u6599\u5e73\u884c\u5ea6\u9ad8\u3001\u800c\u4e14\u6bcf\u500b\u8cc7\u6599\u9ede\u8981\u57f7\u884c\u8a31\u591a\u5e8f\u5217\u6307\u4ee4\u3002\u53ef\u9810\u6e2c\u3001\u898f\u5247\u7684\u8a18\u61b6\u9ad4\u5b58\u53d6\u6a21\u5f0f\u4ee5\u53ca\u7c21\u55ae\u3001\u91cd\u8907\u7684\u8a08\u7b97\u540c\u6a23\u504f\u5411 GPU\u3002\u800c\u4e00\u65e6\u5b58\u5728\u5f37\u8cc7\u6599\u76f8\u4f9d\u6027\u3001\u8a18\u61b6\u9ad4\u5b58\u53d6\u6a21\u5f0f\u9ad8\u5ea6\u4e0d\u898f\u5247\u3001\u6216\u8005\u5206\u652f\u908f\u8f2f\u5f88\u8907\u96dc\uff0cGPU \u7684\u512a\u52e2\u5c31\u6703\u6e1b\u5f31\u3002\u53ef\u4ee5\u7528\u4e0b\u8868\u505a\u7b2c\u4e00\u8f2a\u7be9\u9078\u3002<\/p>\n \n\n\n\n\n<\/colgroup>\n\n\n\n\n\n\n \u5de5\u4f5c\u8ca0\u8f09\u7279\u5fb5<\/th>\n \u6709\u5229\u65bc GPU \/ GPGPU<\/th>\n \u8b93 GPU \u5931\u53bb\u512a\u52e2\u7684\u60c5\u6cc1<\/th>\n<\/tr>\n \u5e73\u884c\u5ea6<\/td>\n \u6210\u5343\u4e0a\u842c\u689d\u7368\u7acb\u57f7\u884c\u7dd2<\/td>\n \u5f37\u8cc7\u6599\u76f8\u4f9d<\/td>\n<\/tr>\n \u7b97\u8853\u5f37\u5ea6<\/td>\n \u6bcf\u500b\u8cc7\u6599\u9ede\u9700\u57f7\u884c\u5927\u91cf\u6307\u4ee4<\/td>\n \u2014<\/td>\n<\/tr>\n \u8a18\u61b6\u9ad4\u5b58\u53d6<\/td>\n \u53ef\u9810\u6e2c\u3001\u898f\u5247\u7684\u5b58\u53d6\u6a21\u5f0f<\/td>\n \u9ad8\u5ea6\u4e0d\u898f\u5247\u7684\u5b58\u53d6\u6a21\u5f0f<\/td>\n<\/tr>\n \u63a7\u5236\u6d41\u7a0b<\/td>\n \u91cd\u8907\u6027\u5f37\u7684\u8a08\u7b97<\/td>\n \u8907\u96dc\u5206\u652f\u908f\u8f2f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n GPU \u4e5f\u5177\u6709\u76f8\u7576\u5f37\u7684\u9069\u914d\u6027\u3002\u5176\u901a\u7528\u7279\u6027\u652f\u63f4\u591a\u7a2e\u6846\u67b6\u548c\u901a\u7528\u904b\u7b97\uff0c\u800c\u5c08\u7528\u52a0\u901f\u5668\u7684\u9069\u7528\u7bc4\u570d\u5247\u66f4\u7a84\u3002\u8981\u5224\u65b7\u64f4\u5c55\u65b9\u5411\uff0c\u53ef\u4ee5\u91dd\u5c0d\u8af8\u5982\u7247\u4e0a\u7db2\u8def\uff08NoC\uff09\u8f38\u9001\u91cf\u3001\u5408\u4f75\u5b58\u53d6\u6bd4\u7387\uff08coalescing rate\uff09\u3001L1 \u5feb\u53d6\u672a\u547d\u4e2d\u7387\u7b49\u6307\u6a19\u5c0d\u61c9\u7528\u9032\u884c\u6548\u80fd\u5206\u6790\u3002\u5c07\u6bcf\u500b\u5be6\u6e2c\u503c\u8207\u5176\u8ff4\u6b78\u4fc2\u6578\u76f8\u4e58\u4e26\u6c42\u548c\uff1a\u548c\u70ba\u6b63\u9810\u6e2c\u4ee5\u7e31\u5411\u64f4\u5145\uff08scale-up\uff09\u66f4\u5408\u9069\uff0c\u548c\u70ba\u8ca0\u5247\u9810\u6e2c\u6a6b\u5411\u64f4\u5c55\uff08scale-out\uff09\u66f4\u6709\u6548\u3002<\/p>\n \u5be6\u7528\u5efa\u8b70<\/h3>\n \u5c0d IT \u7ba1\u7406\u8005\u4f86\u8aaa\uff0c\u7b2c\u4e00\u500b\u554f\u984c\u61c9\u8a72\u662f\u8cc7\u6599\u654f\u611f\u6027\u3002\u6a21\u578b\u63a5\u89f8\u5230\u7684\u4efb\u4f55\u8cc7\u6599\u4e2d\u662f\u5426\u5305\u542b\u53ef\u8b58\u5225\u7684\u65e5\u672c\u81ea\u7136\u4eba\u8cc7\u8a0a\uff1f\u5982\u679c\u662f\uff0c\u5247\u9700\u9075\u5faa\u300a\u500b\u4eba\u8cc7\u8a0a\u4fdd\u8b77\u6cd5\u300b\uff08APPI\uff09\uff0c\u5176\u4e2d\u7b2c 24 \u689d\u8981\u6c42\uff1a\u5728\u5c07\u53ef\u8b58\u5225\u8cc7\u6599\u50b3\u9001\u5230\u5883\u5916\u4e4b\u524d\uff0c\u5fc5\u9808\u53d6\u5f97\u540c\u610f\u6216\u5177\u5099\u5408\u898f\u7684\u6cd5\u5f8b\u57fa\u790e\u3002\u56e0\u6b64\uff0c\u61c9\u5c07\u6b64\u985e\u8cc7\u6599\u4fdd\u5b58\u5728\u65e5\u672c\u5883\u5167\u7684\u904b\u7b97\u8cc7\u6e90\u4e0a\u3002\u5c0d\u65bc\u4f01\u696d\u7d1a\u5ba2\u6236\u4f86\u8aaa\uff0cISMAP \u8a3b\u518a\u662f\u4e00\u9805\u91cd\u8981\u7684\u4fe1\u4efb\u8a0a\u865f\uff0c\u800c\u653f\u5e9c\u985e\u5de5\u4f5c\u8ca0\u8f09\u5be6\u52d9\u4e0a\u5247\u8981\u6c42\u4f7f\u7528\u5728\u653f\u5e9c\u96f2\uff08Government Cloud\uff09\u540d\u55ae\u4e0a\u7684\u3001\u5728\u570b\u5167\u8a3b\u518a\u7684\u57fa\u790e\u8a2d\u65bd\u3002<\/p>\n \u5c0d AI \u5de5\u7a0b\u5e2b\u4f86\u8aaa\uff0c\u61c9\u5c07 GPGPU \u512a\u5148\u4fdd\u7559\u7d66\u5927\u898f\u6a21 AI \u8a13\u7df4\u4ee5\u53ca\u5c0d\u5ef6\u9072\u9ad8\u5ea6\u654f\u611f\u7684\u63a8\u8ad6\u60c5\u5883\u3002\u5c0d\u65bc\u8f15\u91cf\u6d41\u91cf\u8207\u5c0f\u6a21\u578b\u7684\u5167\u90e8\u5de5\u5177\uff0cCPU \u4ecd\u7136\u8db3\u5920\u3002\u82e5\u4f60\u4e0d\u53d7\u8cc7\u6599\u5728\u5730\u5316\u9650\u5236\uff0c\u5e0c\u671b\u4ee5\u8f03\u4f4e\u6210\u672c\u5728\u96f2\u7aef\u57f7\u884c GPU \u5de5\u4f5c\u8ca0\u8f09\uff0c\u90a3\u9ebc\u5168\u7403\u6027\u7684\u65b0\u578b\u96f2\uff08neocloud\uff09\u5f80\u5f80\u80fd\u628a\u6210\u672c\u58d3\u5230\u6771\u4eac\u5340\u57df\u8d85\u5927\u898f\u6a21\u96f2\u5ee0\u5546\u7684\u5e7e\u5206\u4e4b\u4e00\u3002\u82e5\u6d89\u53ca\u53d7\u76e3\u7ba1\u8cc7\u6599\uff0c\u5247\u61c9\u512a\u5148\u9078\u64c7\u5728\u65e5\u672c\u5883\u5167\u7684 GPU \u4f3a\u670d\u5668\uff0c\u6216\u4f4d\u65bc\u6771\u4eac\u3001\u5927\u962a\u7b49\u5340\u57df\u7684\u8cc7\u6599\u4e2d\u5fc3\u3002<\/p>\n \u5c0d\u4f3a\u670d\u5668\u7ba1\u7406\u54e1\u4f86\u8aaa\uff0c\u898f\u5283\u6642\u8981\u5148\u770b\u96fb\u529b\u548c\u6563\u71b1\uff0c\u56e0\u70ba\u96fb\u529b\u4e0a\u9650\u6703\u9650\u5236\u4efb\u4f55\u5927\u898f\u6a21\u5efa\u7f6e\u3002\u5efa\u8b70\u62c6\u5206\u5de5\u4f5c\u8ca0\u8f09\uff1a\u628a\u5be6\u9a57\u578b\u5de5\u4f5c\u653e\u5728\u6210\u672c\u6700\u4f4e\u7684\u7b97\u529b\u4e0a\uff0c\u800c\u5c07\u5177\u6cd5\u898f\u9075\u5faa\u8981\u6c42\u7684\u8cc7\u6599\u7559\u5728\u570b\u5167\u7684 GPU \u8cc7\u6e90\u4e0a\u57f7\u884c\u3002\u96a8\u8457\u6d41\u91cf\u6210\u9577\uff0c\u6301\u7e8c\u91cd\u5be9\u8a13\u7df4\u8207\u63a8\u8ad6\u7684\u6bd4\u4f8b\uff0c\u4e26\u6839\u64da\u57fa\u7dda\u76e3\u63a7 GPU \u6548\u80fd\u8868\u73fe\u3002<\/p>\n \n <\/div>\n<\/div>\n GPU \u662f\u786c\u9ad4\u672c\u8eab\uff0c\u800c GPGPU \u5247\u662f\u5c07\u9019\u5957\u786c\u9ad4\u7528\u65bc\u5716\u5f62\u4ee5\u5916\u4efb\u52d9\u7684\u65b9\u5f0f\u3002\u5c0d\u65bc\u65e5\u672c\u7684 AI \u4f3a\u670d\u5668\u800c\u8a00\uff0cGPGPU \u662f\u5fc5\u7d93\u4e4b\u8def\uff0c\u56e0\u70ba\u5e73\u884c\u904b\u7b97\u662f\u5be6\u73fe\u9ad8\u6548\u8a13\u7df4\u8207\u63a8\u8ad6\u7684\u6838\u5fc3\u9a45\u52d5\u529b\u3002GPU \u8207 GPGPU \u7684\u7d44\u5408\u65e2\u63d0\u4f9b\u6676\u7247\uff0c\u4e5f\u63d0\u4f9b\u71df\u904b\u6a21\u578b\u3002<\/p>\n \u63a5\u4e0b\u4f86\u5c31\u53ef\u4ee5\u628a\u9019\u4e9b\u8a8d\u8b58\u8f49\u5316\u70ba\u884c\u52d5\u898f\u5283\u4e86\u3002\u4f9d\u64da\u5de5\u4f5c\u8ca0\u8f09\u9078\u64c7\u5408\u9069\u6a21\u5f0f\uff0c\u70ba\u8f15\u91cf\u4efb\u52d9\u4fdd\u7559\u4e00\u5957 CPU \u57fa\u7dda\u3002\u5728\u505a\u6c7a\u7b56\u524d\u6b0a\u8861\u96fb\u529b\u3001\u6563\u71b1\u8207\u6cd5\u898f\u9075\u5faa\uff0c\u4e26\u70ba\u672a\u4f86\u64f4\u5c55\u505a\u597d\u8a2d\u8a08\u3002\u904b\u7528 GPU \u89e3\u8026\u5728\u591a\u500b\u4f5c\u696d\u4e4b\u9593\u5171\u4eab\u7b97\u529b\u8cc7\u6e90\uff0c\u96a8\u8457\u6a21\u578b\u898f\u6a21\u6210\u9577\u4e0d\u65b7\u91cd\u5be9 GPU \u7684\u9078\u64c7\uff0c\u8b93 GPU \u89e3\u8026\u5e6b\u52a9\u4f60\u7dad\u6301\u9ad8\u4f7f\u7528\u7387\u3002\u9019\u6a23\u4e00\u4f86\uff0c\u4f60\u65e2\u80fd\u7dad\u6301\u53ef\u63a7\u7684 AI \u6210\u672c\uff0c\u4e5f\u80fd\u78ba\u4fdd AI \u6548\u679c\u7a69\u5b9a\u53ef\u9760\u3002<\/p>\n \u5e38\u898b\u554f\u984c<\/strong><\/h2>\n GPGPU \u662f\u5426\u9700\u8981\u5c08\u7528\u786c\u9ad4\uff1f<\/h3>\n \u4e0d\u9700\u8981\u3002\u4efb\u4f55\u73fe\u4ee3\u8cc7\u6599\u4e2d\u5fc3 GPU \u90fd\u652f\u63f4\u901a\u7528\u904b\u7b97\u3002\u4f60\u9700\u8981\u7684\u662f\u5408\u9069\u7684\u7a0b\u5f0f\u8a2d\u8a08\u6a21\u578b\uff08\u5982 CUDA \u6216 OpenCL\uff09\uff0c\u4ee5\u53ca\u5c0d\u61c9\u7684\u9a45\u52d5\u7a0b\u5f0f\u548c\u51fd\u5f0f\u5eab\u3002GPU \u8207 GPGPU \u7684\u7d44\u5408\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u662f\u4e00\u7a2e\u8edf\u9ad4\u5c64\u9762\u7684\u6c7a\u7b56\u3002\u53ea\u8981\u9078\u64c7\u4e00\u5f35\u986f\u793a\u8a18\u61b6\u9ad4\u8db3\u4ee5\u5bb9\u7d0d\u6a21\u578b\u7684\u986f\u793a\u5361\uff0c\u518d\u5b89\u88dd\u901a\u7528\u904b\u7b97\u5de5\u5177\u93c8\u5373\u53ef\u3002<\/p>\n \u4e00\u500b\u6a21\u578b\u9700\u8981\u591a\u5c11\u986f\u793a\u8a18\u61b6\u9ad4\uff1f<\/h3>\n \u53ef\u6309\u6bcf 1B \u53c3\u6578\u5927\u7d04 16 GB \u986f\u793a\u8a18\u61b6\u9ad4\u4f86\u4f30\u7b97\u8a13\u7df4\u9700\u6c42\u3002\u4e00\u500b 70B \u6a21\u578b\u5728\u8a13\u7df4\u968e\u6bb5\u5927\u7d04\u9700\u8981 840 GB\uff0c\u5728\u63a8\u8ad6\u968e\u6bb5\u7d04\u9700 140 GB\u3002\u5728\u898f\u5283\u4f3a\u670d\u5668\u4e4b\u524d\uff0c\u8acb\u7d50\u5408\u6279\u6b21\u5927\u5c0f\u548c\u7cbe\u5ea6\u8a2d\u5b9a\u4f86\u6821\u6e96\u9019\u4e9b\u4f30\u7b97\u6578\u503c\u3002<\/p>\n \u4ec0\u9ebc\u6642\u5019 CPU \u6703\u5728 AI \u4e2d\u512a\u65bc GPU\uff1f<\/h3>\n \u5728\u8f15\u91cf\u7d1a\u3001\u5167\u90e8\u4f7f\u7528\u7684\u5de5\u5177\u60c5\u5883\u4e2d\uff0cCPU \u5f80\u5f80\u66f4\u5360\u512a\u3002\u82e5\u65e5\u8acb\u6c42\u91cf\u4f4e\u65bc\u7d04 100 \u6b21\u3001\u6a21\u578b\u898f\u6a21\u5c0f\u65bc 10B \u53c3\u6578\u3001\u4f7f\u7528\u8005\u80fd\u63a5\u53d7\u6578\u79d2\u5ef6\u9072\uff0c\u90a3\u9ebc\u63a1\u7528 CPU \u57fa\u7dda\u66f4\u5408\u9069\u3002\u7576\u6d41\u91cf\u6216\u6a21\u578b\u898f\u6a21\u6210\u9577\u6642\uff0c\u518d\u9077\u79fb\u5230 GPU\u3002<\/p>\n GPU \u89e3\u8026\u5982\u4f55\u5e6b\u52a9\u6211\u7684\u90e8\u7f72\uff1f<\/h3>\n GPU \u89e3\u8026\u6703\u5728\u7bc0\u9ede\u4e4b\u9593\u6c60\u5316\u52a0\u901f\u5361\uff0c\u8b93\u591a\u9805\u4f5c\u696d\u5171\u4eab\u540c\u4e00\u5957\u8cc7\u6e90\u7db2\u8def\u3002\u9019\u65e2\u63d0\u9ad8\u4e86\u4f7f\u7528\u7387\uff0c\u4e5f\u7c21\u5316\u4e86\u5347\u7d1a\u6d41\u7a0b\u3002\u4f60\u53ef\u4ee5\u6309\u7167\u8a13\u7df4\u7684\u5cf0\u503c\u9700\u6c42\u4f86\u898f\u5283\u7e3d\u5bb9\u91cf\uff0c\u7136\u5f8c\u5c07\u9019\u90e8\u5206\u5bb9\u91cf\u91cd\u8907\u7528\u65bc\u63a8\u8ad6\u3002\u8cc7\u6599\u4e2d\u5fc3\u7d1a GPU \u89e3\u8026\u80fd\u8b93 GPU \u5728\u5404\u7a2e AI \u5de5\u4f5c\u8ca0\u8f09\u4e2d\u7684\u6548\u80fd\u4fdd\u6301\u53ef\u9810\u671f\u3002<\/p>\n \u54ea\u4e9b\u6cd5\u898f\u9075\u5faa\u8981\u6c42\u6703\u5f71\u97ff\u6211\u7684\u9078\u64c7\uff1f<\/h3>\n \u300a\u500b\u4eba\u8cc7\u8a0a\u4fdd\u8b77\u6cd5\u300b\uff08APPI\uff09\u7d04\u675f\u8457\u500b\u4eba\u8cc7\u6599\u7684\u8655\u7406\u3002\u6839\u64da\u7b2c 24 \u689d\uff0c\u5728\u5c07\u53ef\u8b58\u5225\u8cc7\u6599\u50b3\u9001\u5230\u5883\u5916\u4e4b\u524d\uff0c\u4f60\u5fc5\u9808\u53d6\u5f97\u540c\u610f\u6216\u5177\u5099\u5408\u6cd5\u4f9d\u64da\u3002\u727d\u6d89\u53d7\u76e3\u7ba1\u8cc7\u6599\u6642\uff0c\u8981\u5c07\u5176\u4fdd\u5b58\u5728\u65e5\u672c\u5883\u5167\u7684\u904b\u7b97\u8cc7\u6e90\u4e0a\u3002ISMAP \u8a3b\u518a\u5c0d\u4f01\u696d\u7d1a\u5ba2\u6236\u800c\u8a00\u662f\u4e00\u9805\u91cd\u8981\u7684\u4fe1\u4efb\u8a0a\u865f\u3002\u5728\u505a\u4efb\u4f55 GPU \u65b9\u6848\u6c7a\u7b56\u524d\uff0c\u90fd\u8981\u7d9c\u5408\u8003\u91cf\u8cc7\u6599\u5728\u5730\u5316\u8207\u51fa\u53e3\u7ba1\u5236\u8981\u6c42\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" GPU \u662f\u5e73\u884c\u8655\u7406\u786c\u9ad4\uff0cGPGPU \u5247\u662f\u5c07\u9019\u7a2e GPU \u7528\u65bc\u5716\u5f62\u4ee5\u5916\u7684\u901a\u7528\u904b\u7b97\u3002\u5c0d\u65bc\u65e5\u672c AI \u4f3a\u670d\u5668\u4f86\u8aaa\uff0cG […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":35348,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[8116],"tags":[13885,13886,13887,13888,3416],"class_list":["post-35352","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-japan-dedicated-server-tc","tag-gpgpu","tag-ai-servers-in-japan","tag-gpu-disaggregation","tag-data-center-compliance","tag-gpu-tc"],"acf":[],"yoast_head":"\n