\u7a7a\u6c23\u51b7\u537b\u512a\u5316<\/h3>\n

\u5be6\u65bd\u4ee5\u4e0b\u95dc\u9375\u6539\u9032\uff1a<\/p>\n

\n
• \u9ad8\u975c\u58d3\u98a8\u6247\uff1a\n
  \n
  • \u6700\u5c0f\u6c23\u6d41\uff1a250 CFM<\/li>\n
  • \u975c\u58d3\uff1a4.5mm H\u2082O<\/li>\n
  • PWM\u63a7\u5236\u7bc4\u570d\uff1a800-3000 RPM<\/li>\n<\/ul>\n<\/li>\n
  • \u5148\u9032\u5c0e\u71b1\u754c\u9762\u6750\u6599\uff1a\n
    \n
    • \u5c0e\u71b1\u7387\uff1a>12 W\/mK<\/li>\n
    • \u63a5\u5408\u5c64\u539a\u5ea6\uff1a<0.05mm<\/li>\n
    • \u66f4\u63db\u9031\u671f\uff1a6\u500b\u6708<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n

      \u9999\u6e2f\u6c23\u5019\u7684\u74b0\u5883\u63a7\u5236\u63aa\u65bd<\/strong><\/h2>\n

      \u9999\u6e2f\u7684\u7368\u7279\u6c23\u5019\u9700\u8981\u5c08\u9580\u7684\u74b0\u5883\u63a7\u5236\u3002\u5be6\u65bd\u61c9\u9075\u5faa\u4ee5\u4e0b\u7cbe\u78ba\u898f\u7bc4\uff1a<\/p>\n

      \n

      \u95dc\u9375\u53c3\u6578\uff1a<\/p>\n

      \n
      • \u6eab\u5ea6\u68af\u5ea6\u7ba1\u7406\uff1a\n
        \n
        • \u51b7\u901a\u9053\u76ee\u6a19\uff1a18\u00b0C \u00b11\u00b0C<\/li>\n
        • \u71b1\u901a\u9053\u6700\u9ad8\uff1a35\u00b0C<\/li>\n
        • \u5782\u76f4\u68af\u5ea6\uff1a<3\u00b0C\/\u7c73<\/li>\n<\/ul>\n<\/li>\n
        • \u6fd5\u5ea6\u63a7\u5236\u5354\u8b70\uff1a\n
          \n
          • \u76f8\u5c0d\u6fd5\u5ea6\uff1a45-55%<\/li>\n
          • \u9732\u9ede\uff1a\u6700\u4f4e5.5\u00b0C<\/li>\n
          • \u6c34\u5206\u8b8a\u5316\u7387\uff1a<5%\/\u5c0f\u6642<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/div>\n

            \u9ad8\u7d1a\u8edf\u9ad4\u512a\u5316\u6280\u8853<\/strong><\/h2>\n

            \u4f7f\u7528\u8907\u96dc\u63a7\u5236\u7cfb\u7d71\u5be6\u65bd\u4ee5\u4e0b\u57fa\u65bc\u8edf\u9ad4\u7684\u6563\u71b1\u7ba1\u7406\u89e3\u6c7a\u65b9\u6848\uff1a<\/p>\n

            \r\n#!\/bin\/bash\r\n\r\n# \u9ad8\u7d1aGPU\u96fb\u6e90\u7ba1\u7406\u8173\u672c\r\ndeclare -A TEMP_THRESHOLDS=(\r\n    [\"critical\"]=85\r\n    [\"high\"]=80\r\n    [\"medium\"]=75\r\n    [\"low\"]=70\r\n)\r\n\r\ndeclare -A POWER_LIMITS=(\r\n    [\"critical\"]=200\r\n    [\"high\"]=250\r\n    [\"medium\"]=300\r\n    [\"low\"]=350\r\n)\r\n\r\nmonitor_and_adjust() {\r\n    while true; do\r\n        for gpu in $(nvidia-smi --query-gpu=index --format=csv,noheader); do\r\n            temp=$(nvidia-smi -i $gpu --query-gpu=temperature.gpu --format=csv,noheader)\r\n            util=$(nvidia-smi -i $gpu --query-gpu=utilization.gpu --format=csv,noheader | cut -d' ' -f1)\r\n            \r\n            # \u57fa\u65bc\u6eab\u5ea6\u548c\u4f7f\u7528\u7387\u7684\u52d5\u614b\u529f\u7387\u8abf\u6574\r\n            if [ $temp -gt ${TEMP_THRESHOLDS[\"critical\"]} ]; then\r\n                nvidia-smi -i $gpu -pl ${POWER_LIMITS[\"critical\"]}\r\n                notify_admin \"GPU $gpu \u81e8\u754c\u6eab\u5ea6: ${temp}\u00b0C\"\r\n            elif [ $temp -gt ${TEMP_THRESHOLDS[\"high\"]} ]; then\r\n                nvidia-smi -i $gpu -pl ${POWER_LIMITS[\"high\"]}\r\n            elif [ $temp -gt ${TEMP_THRESHOLDS[\"medium\"]} ]; then\r\n                nvidia-smi -i $gpu -pl ${POWER_LIMITS[\"medium\"]}\r\n            fi\r\n            \r\n            log_metrics $gpu $temp $util\r\n        done\r\n        sleep 60\r\n    done\r\n}\r\n<\/code><\/pre>\n
            \u667a\u6167\u5de5\u4f5c\u8ca0\u8f09\u5206\u914d\u67b6\u69cb<\/strong><\/h2>\n
            \u73fe\u4ee3\u4f3a\u670d\u5668\u8a17\u7ba1\u8a2d\u65bd\u5fc5\u9808\u5be6\u65bd\u667a\u6167\u5de5\u4f5c\u8ca0\u8f09\u5206\u914d\u4ee5\u9632\u6b62\u71b1\u9ede\u3002\u4ee5\u4e0b\u662f\u91dd\u5c0d\u6563\u71b1\u7ba1\u7406\u512a\u5316\u7684Kubernetes\u914d\u7f6e\uff1a<\/p>\n
            \r\napiVersion: v1\r\nkind: Pod\r\nmetadata:\r\n  name: gpu-workload-thermal-aware\r\n  annotations:\r\n    scheduler.alpha.kubernetes.io\/node-selector: |\r\n      thermal-zone=optimal\r\nspec:\r\n  containers:\r\n  - name: gpu-container\r\n    resources:\r\n      limits:\r\n        nvidia.com\/gpu: 1\r\n      requests:\r\n        nvidia.com\/gpu: 1\r\n    env:\r\n    - name: NVIDIA_VISIBLE_DEVICES\r\n      value: \"all\"\r\n    - name: NVIDIA_DRIVER_CAPABILITIES\r\n      value: \"compute,utility,video\"\r\n    - name: GPU_TEMP_THRESHOLD\r\n      value: \"80\"\r\n    volumeMounts:\r\n    - name: nvidia-docker-runtime\r\n      mountPath: \/usr\/local\/nvidia\r\n    securityContext:\r\n      privileged: true\r\n<\/code><\/pre>\n
            \u5168\u9762\u76e3\u63a7\u57fa\u790e\u8a2d\u65bd<\/strong><\/h2>\n
            \u90e8\u7f72\u4ee5\u4e0b\u57fa\u672c\u76e3\u63a7\u7d44\u4ef6\uff1a<\/p>\n
            \n
            \u5373\u6642\u6307\u6a19\u63a1\u96c6\uff1a\n
            \n
            GPU\u6eab\u5ea6\u63a1\u6a23\u7387\uff1a1\u6b21\/\u79d2<\/li>\n
            \u529f\u8017\u76e3\u63a7\uff1a500\u6beb\u79d2\u9593\u9694<\/li>\n
            \u98a8\u6247\u901f\u5ea6\u8ffd\u8e64\uff1a\u52d5\u614b\u8abf\u7bc0<\/li>\n
            \u8a18\u61b6\u9ad4\u7d50\u6eab\u5ea6\u76e3\u63a7<\/li>\n<\/ul>\n<\/li>\n
            \u8b66\u5831\u95be\u503c\uff1a\n
            \n
            \u6eab\u5ea6\u8b66\u544a\uff1a>80\u00b0C<\/li>\n
            \u81e8\u754c\u8b66\u5831\uff1a>85\u00b0C<\/li>\n
            \u529f\u7387\u5cf0\u503c\uff1a>110% TDP<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
            \u61c9\u6025\u97ff\u61c9\u5354\u8b70\u77e9\u9663<\/strong><\/h2>\n
            \u5be6\u65bd\u4ee5\u4e0b\u5206\u7d1a\u97ff\u61c9\u7cfb\u7d71\uff1a<\/p>\n
            \n
            \u4e00\u7d1a\u97ff\u61c9\uff08\u6eab\u5ea6 >80\u00b0C\uff09\uff1a\n
            \n
            \u81ea\u52d5\u529f\u7387\u9650\u5236<\/li>\n
            \u63d0\u5347\u51b7\u537b\u7cfb\u7d71\u5bb9\u91cf<\/li>\n
            \u555f\u52d5\u8ca0\u8f09\u91cd\u5206\u914d<\/li>\n<\/ul>\n<\/li>\n
            \u4e8c\u7d1a\u97ff\u61c9\uff08\u6eab\u5ea6 >85\u00b0C\uff09\uff1a\n
            \n
            \u5de5\u4f5c\u8ca0\u8f09\u9077\u79fb\u81f3\u5099\u7528\u7cfb\u7d71<\/li>\n
            \u555f\u52d5\u7dca\u6025\u51b7\u537b<\/li>\n
            \u901a\u77e5\u6280\u8853\u652f\u63f4<\/li>\n<\/ul>\n<\/li>\n
            \u4e09\u7d1a\u97ff\u61c9\uff08\u6eab\u5ea6 >90\u00b0C\uff09\uff1a\n
            \n
            \u7acb\u5373\u66ab\u505c\u5de5\u4f5c\u8ca0\u8f09<\/li>\n
            \u7dca\u6025\u95dc\u6a5f\u7a0b\u5e8f<\/li>\n
            \u555f\u52d5\u4e8b\u6545\u97ff\u61c9\u5718\u968a<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
            \u9810\u9632\u6027\u7dad\u8b77\u8a08\u5283<\/strong><\/h2>\n
            \u9075\u5faa\u4ee5\u4e0b\u5168\u9762\u7dad\u8b77\u6642\u9593\u8868\uff1a<\/p>\n
            \n
            \u6bcf\u65e5\u4efb\u52d9\uff1a\n
            \n
            \u6eab\u5ea6\u65e5\u8a8c\u5206\u6790<\/li>\n
            \u51b7\u537b\u7cfb\u7d71\u6548\u80fd\u6aa2\u67e5<\/li>\n
            \u8b66\u5831\u7cfb\u7d71\u9a57\u8b49<\/li>\n<\/ul>\n<\/li>\n
            \u6bcf\u9031\u4efb\u52d9\uff1a\n
            \n
            \u71b1\u6210\u50cf\u6383\u63cf<\/li>\n
            \u6c23\u6d41\u6a21\u5f0f\u5206\u6790<\/li>\n
            \u7a4d\u7070\u6aa2\u67e5<\/li>\n<\/ul>\n<\/li>\n
            \u6bcf\u6708\u4efb\u52d9\uff1a\n
            \n
            \u51b7\u537b\u7cfb\u7d71\u7dad\u8b77<\/li>\n
            \u6ffe\u7db2\u66f4\u63db<\/li>\n
            \u6563\u71b1\u818f\u6aa2\u67e5<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
            \u5728\u9999\u6e2f\u7684\u4f3a\u670d\u5668\u8a17\u7ba1\u8a2d\u65bd\u4e2d\u7ba1\u7406Blackwell GPU\u6eab\u5ea6\u9700\u8981\u786c\u9ad4\u89e3\u6c7a\u65b9\u6848\u3001\u8edf\u9ad4\u512a\u5316\u548c\u4e3b\u52d5\u76e3\u63a7\u7684\u8907\u96dc\u7d44\u5408\u3002\u901a\u904e\u5be6\u65bd\u9019\u4e9b\u5168\u9762\u63aa\u65bd\uff0c\u6578\u64da\u4e2d\u5fc3\u53ef\u4ee5\u5728\u5177\u6709\u6311\u6230\u6027\u7684\u6c23\u5019\u689d\u4ef6\u4e0b\u4fdd\u6301\u6700\u4f73GPU\u6548\u80fd\uff0c\u540c\u6642\u78ba\u4fdd\u7cfb\u7d71\u7684\u9577\u671f\u53ef\u7528\u6027\u3002\u57fa\u65bc\u6548\u80fd\u6307\u6a19\u548c\u74b0\u5883\u8b8a\u5316\u5b9a\u671f\u66f4\u65b0\u9019\u4e9b\u5354\u8b70\u5c07\u78ba\u4fdd\u60a8\u7684\u6563\u71b1\u7ba1\u7406\u7b56\u7565\u6301\u7e8c\u6709\u6548\u3002<\/p>\n
            [\/vc_column_text][\/vc_column][\/vc_row]<\/p>\n<\/div>","protected":false},"excerpt":{"rendered":"
            [vc_row][vc_column][vc_column_text css=””]  […]<\/p>\n
            Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":19633,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[61],"tags":[4856,4857,4858,3716,645],"class_list":["post-19642","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-hong-kong-dedicated-server-tc","tag-nvidia-blackwell-gpu-tc","tag-gpu-overheating-tc","tag-gpu-temperature-control-tc","tag-server-cooling-tc","tag-hong-kong-data-center-tc"],"acf":[],"yoast_head":"\n