\u7a0b\u5f0f\u78bc\u7d1a\u512a\u5316\u6280\u8853<\/strong><\/h2>\n

\u667a\u6167\u7684\u7de8\u78bc\u5be6\u8e10\u53ef\u4ee5\u986f\u8457\u63d0\u9ad8\u7f8e\u570bGPU\u4f3a\u670d\u5668\u7684\u8a13\u7df4\u6548\u7387\u3002\u4ee5\u4e0b\u662f\u5982\u4f55\u512a\u5316\u7a0b\u5f0f\u78bc\u4ee5\u7372\u5f97\u6700\u4f73\u6548\u80fd\uff1a<\/p>\n

\n
• \u6279\u91cf\u5927\u5c0f\u512a\u5316\uff1a\n
  \n
  • \u5f9e2\u7684\u51aa\u6b21\u65b9\u6279\u91cf\u5927\u5c0f\u958b\u59cb\uff0832\u300164\u3001128\uff09<\/li>\n
  • \u4f7f\u7528\u68af\u5ea6\u7d2f\u7a4d\u5be6\u73fe\u66f4\u5927\u7684\u6709\u6548\u6279\u91cf<\/li>\n
  • \u76e3\u63a7\u8a18\u61b6\u9ad4\u4f7f\u7528\u8207\u8a13\u7df4\u7a69\u5b9a\u6027<\/li>\n<\/ul>\n<\/li>\n
  • \u8a18\u61b6\u9ad4\u7ba1\u7406\uff1a\n
    \n
    • \u5be6\u73fe\u68af\u5ea6\u6aa2\u67e5\u9ede<\/li>\n
    • \u4f7f\u7528\u6df7\u5408\u7cbe\u5ea6\u8a13\u7df4\uff08FP16\/BF16\uff09<\/li>\n
    • \u5728\u8a13\u7df4\u8fed\u4ee3\u4e4b\u9593\u6e05\u9664\u5feb\u53d6<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n

      \u4ee5\u4e0b\u662f\u5be6\u73fe\u6df7\u5408\u7cbe\u5ea6\u8a13\u7df4\u7684\u5be6\u7528\u793a\u4f8b\uff1a<\/p>\n

      \r\nimport torch\r\nfrom torch.cuda.amp import autocast, GradScaler\r\n\r\nscaler = GradScaler()\r\nfor data in dataloader:\r\n    with autocast():\r\n        output = model(data)\r\n        loss = criterion(output)\r\n    scaler.scale(loss).backward()\r\n    scaler.step(optimizer)\r\n    scaler.update()\r\n<\/code><\/pre>\n
      \u6578\u64da\u7ba1\u9053\u512a\u5316<\/strong><\/h2>\n
      \u9ad8\u6548\u7684\u6578\u64da\u8655\u7406\u5c0d\u65bc\u7dad\u6301\u7f8e\u570bGPU\u4f3a\u670d\u5668\u7684\u6700\u4f73\u5229\u7528\u7387\u81f3\u95dc\u91cd\u8981\u3002\u8003\u616e\u9019\u4e9b\u9032\u968e\u6280\u8853\uff1a<\/p>\n
      \n
      \u6578\u64da\u8f09\u5165\uff1a\n
      \n
      \u4f7f\u7528NVIDIA DALI\u9032\u884cGPU\u52a0\u901f\u6578\u64da\u8f09\u5165<\/li>\n
      \u5be6\u73fe\u9810\u53d6\u6a5f\u5236<\/li>\n
      \u512a\u5316\u6578\u64da\u96c6\u683c\u5f0f\uff08TFRecord\u3001WebDataset\uff09<\/li>\n<\/ul>\n<\/li>\n
      \u5132\u5b58\u89e3\u6c7a\u65b9\u6848\uff1a\n
      \n
      \u4f7f\u7528NVMe SSD\u4ee5\u7372\u5f97\u66f4\u5feb\u7684I\/O<\/li>\n
      \u5be6\u73fe\u6578\u64da\u5206\u7247<\/li>\n
      \u5c0d\u5c0f\u578b\u6578\u64da\u96c6\u8003\u616e\u57fa\u65bcRAM\u7684\u6578\u64da\u96c6<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n
      \u6846\u67b6\u7279\u5b9a\u512a\u5316<\/strong><\/h2>\n
      \u4e0d\u540c\u7684\u6df1\u5ea6\u5b78\u7fd2\u6846\u67b6\u70ba\u7f8e\u570bGPU\u4f3a\u670d\u5668\u63d0\u4f9b\u7368\u7279\u7684\u512a\u5316\u6a5f\u6703\uff1a<\/p>\n
      \n
      PyTorch\u512a\u5316\uff1a\n
      \n
      \u555f\u7528JIT\u7de8\u8b6f<\/li>\n
      \u5c0dPyTorch 2.0+\u4f7f\u7528torch.compile()<\/li>\n
      \u5be6\u73feDistributedDataParallel<\/li>\n<\/ul>\n<\/li>\n
      TensorFlow\u512a\u5316\uff1a\n
      \n
      \u555f\u7528XLA\u7de8\u8b6f<\/li>\n
      \u4f7f\u7528tf.function\u88dd\u98fe\u5668<\/li>\n
      \u5be6\u73fetf.distribute\u7b56\u7565<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
      \u76e3\u63a7\u548c\u6548\u80fd\u8ffd\u8e64<\/strong><\/h2>\n
      \u5be6\u65bd\u5f37\u5927\u7684\u76e3\u63a7\u7cfb\u7d71\u78ba\u4fdd\u7f8e\u570bGPU\u4f3a\u670d\u5668\u7684\u6301\u7e8c\u512a\u5316\uff1a<\/p>\n
      \n
      \u95dc\u9375\u6307\u6a19\u8ffd\u8e64\uff1a\n
      \n
      GPU\u4f7f\u7528\u7387\uff08\u76ee\u6a19>90%\uff09<\/li>\n
      \u8a18\u61b6\u9ad4\u4f7f\u7528\u6a21\u5f0f<\/li>\n
      PCIe\u983b\u5bec\u5229\u7528\u7387<\/li>\n
      \u6eab\u5ea6\u6307\u6a19<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
      \u4f7f\u7528\u9019\u500b\u7c21\u55ae\u7684Python\u8173\u672c\u9032\u884c\u57fa\u672cGPU\u76e3\u63a7\uff1a<\/p>\n
      \r\nimport nvidia_smi\r\n\r\ndef monitor_gpu():\r\n    nvidia_smi.nvmlInit()\r\n    handle = nvidia_smi.nvmlDeviceGetHandleByIndex(0)\r\n    info = nvidia_smi.nvmlDeviceGetMemoryInfo(handle)\r\n    util = nvidia_smi.nvmlDeviceGetUtilizationRates(handle)\r\n    print(f\"\u8a18\u61b6\u9ad4\uff1a{info.used\/1024**2:.2f}MB\")\r\n    print(f\"\u4f7f\u7528\u7387\uff1a{util.gpu}%\")\r\n<\/code><\/pre>\n
      \u5e38\u898b\u6548\u80fd\u554f\u984c\u6545\u969c\u6392\u9664<\/strong><\/h2>\n
      \u89e3\u6c7a\u9019\u4e9b\u983b\u7e41\u51fa\u73fe\u7684\u74f6\u9838\u4ee5\u7dad\u6301\u7f8e\u570bGPU\u4f3a\u670d\u5668\u7684\u6700\u4f73\u8a13\u7df4\u901f\u5ea6\uff1a<\/p>\n
      \n
      \u8a18\u61b6\u9ad4\u554f\u984c\uff1a\n
      \n
      \u8a18\u61b6\u9ad4\u6ea2\u51fa\u932f\u8aa4<\/li>\n
      \u8a18\u61b6\u9ad4\u788e\u7247\u5316<\/li>\n
      \u5feb\u53d6\u6ea2\u51fa<\/li>\n<\/ul>\n<\/li>\n
      \u8655\u7406\u74f6\u9838\uff1a\n
      \n
      CPU\u74f6\u9838<\/li>\n
      I\/O\u9650\u5236<\/li>\n
      \u7db2\u8def\u983b\u5bec\u9650\u5236<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n
      \u6700\u4f73\u5be6\u8e10\u548c\u9762\u5411\u672a\u4f86<\/strong><\/h2>\n
      \u900f\u904e\u9019\u4e9b\u7b56\u7565\u7dad\u6301\u7f8e\u570bGPU\u4f3a\u670d\u5668\u7684\u9577\u671f\u512a\u5316\uff1a<\/p>\n
      \n
      \u5b9a\u671f\u7dad\u8b77\uff1a\n
      \n
      \u6bcf\u9031\u9a45\u52d5\u7a0b\u5f0f\u66f4\u65b0<\/li>\n
      \u6bcf\u6708\u6548\u80fd\u7a3d\u6838<\/li>\n
      \u5b63\u5ea6\u786c\u9ad4\u6aa2\u67e5<\/li>\n<\/ul>\n<\/li>\n
      \u672a\u4f86\u8003\u616e\uff1a\n
      \n
      \u898f\u5283\u53ef\u64f4\u5c55\u6027<\/li>\n
      \u53ca\u6642\u4e86\u89e3\u6700\u65b0GPU\u6280\u8853<\/li>\n
      \u8003\u616e\u96f2GPU\u4f3a\u670d\u5668\u79df\u7528\u66ff\u4ee3\u65b9\u6848<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
      \u7d50\u8ad6<\/strong><\/h2>\n
      \u512a\u5316\u7f8e\u570bGPU\u4f3a\u670d\u5668\u8a13\u7df4\u901f\u5ea6\u9700\u8981\u6574\u9ad4\u65b9\u6cd5\uff0c\u7d50\u5408\u786c\u9ad4\u5c08\u696d\u77e5\u8b58\u548c\u8edf\u9ad4\u6280\u5de7\u3002\u900f\u904e\u5be6\u65bd\u9019\u4e9b\u5148\u9032\u7684\u512a\u5316\u6280\u8853\uff0c\u60a8\u53ef\u4ee5\u986f\u8457\u63d0\u5347GPU\u4f3a\u670d\u5668\u6548\u80fd\u548c\u8a13\u7df4\u6548\u7387\u3002\u8acb\u8a18\u4f4f\uff0c\u7f8e\u570bGPU\u4f3a\u670d\u5668\u512a\u5316\u662f\u4e00\u500b\u9700\u8981\u5b9a\u671f\u76e3\u63a7\u548c\u66f4\u65b0\u7684\u6301\u7e8c\u904e\u7a0b\uff0c\u4ee5\u4fdd\u6301\u6700\u4f73\u6548\u80fd\u3002<\/p>\n
      \u7121\u8ad6\u60a8\u662f\u4f7f\u7528\u7f8e\u570bGPU\u4f3a\u670d\u5668\u79df\u7528\u670d\u52d9\u9084\u662f\u7ba1\u7406\u81ea\u5df1\u7684\u4f3a\u670d\u5668\u8a17\u7ba1\u8a2d\u7f6e\uff0c\u9019\u4e9b\u512a\u5316\u7b56\u7565\u90fd\u5c07\u5e6b\u52a9\u60a8\u5be6\u73fe\u6700\u5927\u8a13\u7df4\u901f\u5ea6\u548c\u6700\u4f73\u8cc7\u6e90\u5229\u7528\u3002\u5728\u512a\u5316\u5de5\u4f5c\u4e2d\u4fdd\u6301\u7a4d\u6975\u4e3b\u52d5\uff0c\u4e26\u96a8\u8457\u6280\u8853\u767c\u5c55\u52c7\u65bc\u5617\u8a66\u65b0\u6280\u8853\u3002<\/p>\n
       \u91cd\u9ede\u95dc\u9375\u8a5e\u5728US GPU Server<\/p>\n","protected":false},"excerpt":{"rendered":"
      \u5728\u7576\u4eca\u7684\u79d1\u6280\u74b0\u5883\u4e2d\uff0c\u512a\u5316\u7f8e\u570bGPU\u4f3a\u670d\u5668\u5728AI\u8a13\u7df4\u4e2d\u7684\u6548\u80fd\u8868\u73fe\u8b8a\u5f97\u81f3\u95dc\u91cd\u8981\u3002\u7121\u8ad6\u662f\u904b\u884c\u8907\u96dc\u7684\u795e\u7d93\u7db2\u8def\u9084\u662f\u8655\u7406\u6d77 […]<\/p>\n
      Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":29012,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[4815],"tags":[9549,9550,9551,4278,3354],"class_list":["post-29018","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-america-dedicated-server-tc","tag-us-gpu-server-optimization","tag-training-speed-improvement","tag-deep-learning-performance","tag-gpu-hosting-tc","tag-ai-acceleration-tc"],"acf":[],"yoast_head":"\n