{"id":710,"date":"2025-01-29T12:02:22","date_gmt":"2025-01-29T12:02:22","guid":{"rendered":"https:\/\/janusai.pro\/?p=710"},"modified":"2025-01-29T12:02:40","modified_gmt":"2025-01-29T12:02:40","slug":"deepseek-v3-paper-details-how-to-bypass-the-cuda-monopoly","status":"publish","type":"post","link":"https:\/\/janusai.pro\/zh_cn\/deepseek-v3-paper-details-how-to-bypass-the-cuda-monopoly\/","title":{"rendered":"DeepSeek V3\u8bba\u6587\u8be6\u60c5\uff1a\u5982\u4f55\u7ed5\u8fc7 CUDA \u5784\u65ad\uff01"},"content":{"rendered":"<div style=\"margin-top: 0px; margin-bottom: 0px;\" class=\"sharethis-inline-share-buttons\" ><\/div>\n<p><a href=\"https:\/\/www.deepseek.com\/\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">DeepSeek V3<\/a> \u8bba\u6587\u8be6\u60c5\uff1a\u5982\u4f55\u7ed5\u8fc7 CUDA \u5784\u65ad\uff01<\/p>\n\n\n\n<p>DeepSeek \u6700\u8fd1\u53d1\u5e03\u7684\u4e24\u4e2a\u578b\u53f7 DeepSeek-V3 \u548c DeepSeek-R1\uff0c\u5176\u6027\u80fd\u53ef\u4e0e OpenAI \u7684\u7c7b\u4f3c\u578b\u53f7\u76f8\u5ab2\u7f8e\uff0c\u800c\u6210\u672c\u5374\u4f4e\u5f97\u591a\u3002<\/p>\n\n\n\n<p>\u636e\u56fd\u5916\u5a92\u4f53\u62a5\u9053\uff0c\u5728\u77ed\u77ed\u4e24\u4e2a\u6708\u7684\u65f6\u95f4\u91cc\uff0c\u4ed6\u4eec\u5728\u7531 2048 \u4e2a H800 GPU \u7ec4\u6210\u7684\u96c6\u7fa4\u4e0a\u8bad\u7ec3\u51fa\u4e86\u4e00\u4e2a\u62e5\u6709 6710 \u4ebf\u4e2a\u53c2\u6570\u7684 MoE \u8bed\u8a00\u6a21\u578b\uff0c\u5176\u6548\u7387\u662f\u9876\u7ea7\u4eba\u5de5\u667a\u80fd\u7684 10 \u500d\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e00\u7a81\u7834\u4e0d\u662f\u901a\u8fc7 CUDA \u5b9e\u73b0\u7684\uff0c\u800c\u662f\u901a\u8fc7\u5927\u91cf\u7ec6\u7c92\u5ea6\u4f18\u5316\u548c\u4f7f\u7528\u82f1\u4f1f\u8fbe\u7c7b\u4f3c\u6c47\u7f16\u7684 PTX\uff08\u5e76\u884c\u7ebf\u7a0b\u6267\u884c\uff09\u7f16\u7a0b\u5b9e\u73b0\u7684\u3002<\/p>\n\n\n\n<p><a href=\"https:\/\/www.deepseek.com\/\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">\u6df1\u5ea6\u641c\u7d22<\/a> \u4e0e OpenAI \u548c\u5176\u4ed6\u516c\u53f8\u5728\u786c\u4ef6\u9650\u5236\u4e0b\u4f9d\u9760\u86ee\u529b\u8ba1\u7b97\u80fd\u529b\u7684\u505a\u6cd5\u4e0d\u540c\uff0cOpenAI \u88ab\u8feb\u8d70\u4e0a\u4e86\u53e6\u4e00\u6761\u9053\u8def\u3002\u5b83\u91c7\u7528\u4e86\u4e00\u7cfb\u5217\u6280\u672f\u521b\u65b0\uff0c\u5728\u5b9e\u73b0\u6027\u80fd\u63d0\u5347\u7684\u540c\u65f6\u964d\u4f4e\u4e86\u6a21\u578b\u7684\u529f\u8017\u8981\u6c42\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/rmy9ct2fln.feishu.cn\/space\/api\/box\/stream\/download\/asynccode\/?code=ZDM1YTM0ODZkYmQzOWNkNzc2ZTBmNzUwY2ZjOWYxMjZfYnUyVHFsb05ya0c1M0hvMGRUbk9CN3FVekR1ZjlQMEZfVG9rZW46TUtzM2JudThpb1p3NHJ4SlZNeWNWdU10bnNnXzE3MzgxNTE4NjQ6MTczODE1NTQ2NF9WNA\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u4e00\u4e9b\u7f51\u6c11\u7684\u70ed\u60c5\u8bc4\u8bba<\/p>\n\n\n\n<p>\"\u5728\u8fd9\u4e2a\u4e16\u754c\u4e0a\uff0c\u5982\u679c\u6709\u4efb\u4f55\u4e00\u7fa4\u4eba\u4f1a\u75af\u72c2\u5730\u8bf4\u51fa'CUDA \u592a\u6162\u4e86'\u8fd9\u6837\u7684\u8bdd\uff01<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/rmy9ct2fln.feishu.cn\/space\/api\/box\/stream\/download\/asynccode\/?code=OGEyMmE3ZTJkODlkZDlmNjliZTI1MzI5YTE4ZWE3MjdfWWRBam5VTkVaV1ZsMFg3VzVTRjRDZlUzV2ZiSHZYT2RfVG9rZW46VGZsdWJrTzZHb243OUx4bEZsbmNmMFNzblFiXzE3MzgxNTE4NjQ6MTczODE1NTQ2NF9WNA\" alt=\"\"\/><\/figure>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/rmy9ct2fln.feishu.cn\/space\/api\/box\/stream\/download\/asynccode\/?code=NzI1ZTBlMjJkMDI2N2MyMDdkMGI4YmU5OTJjNGM0YzFfZW4xbjVERFdhdGVObHBDUWR3NVZjbDRSM2lrVDlWRGlfVG9rZW46Q2N5MWIxV2ltbzdmZU14VXI2amNuZDk2bmRkXzE3MzgxNTE4NjQ6MTczODE1NTQ2NF9WNA\" alt=\"\"\/><\/figure>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/rmy9ct2fln.feishu.cn\/space\/api\/box\/stream\/download\/asynccode\/?code=MDMyN2YyYjYwYTNkZDhmMmEyYWY2MjMzZGE3MGM1ZmFfM29veUZrRWdYODRGR0JVdWVVTnRoMzVwTWxjV09CT25fVG9rZW46SVE2dGJWek9Mb29jaTJ4ZnkzWWN5bUZWbnVnXzE3MzgxNTE4NjQ6MTczODE1NTQ2NF9WNA\" alt=\"\"\/><\/figure>\n\n\n\n<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_72 counter-hierarchy ez-toc-counter ez-toc-grey ez-toc-container-direction\">\n<div class=\"ez-toc-title-container\">\n<p class=\"ez-toc-title\" style=\"cursor:inherit\">\u76ee\u5f55<\/p>\n<span class=\"ez-toc-title-toggle\"><a href=\"#\" class=\"ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle\" aria-label=\"\u5207\u6362\u76ee\u5f55\"><span class=\"ez-toc-js-icon-con\"><span class=\"\"><span class=\"eztoc-hide\" style=\"display:none;\">\u5207\u6362<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #999;color:#999\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewbox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #999;color:#999\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewbox=\"0 0 24 24\" version=\"1.2\" baseprofile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/span><\/a><\/span><\/div>\n<nav><ul class='ez-toc-list ez-toc-list-level-1' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/janusai.pro\/zh_cn\/deepseek-v3-paper-details-how-to-bypass-the-cuda-monopoly\/#Genius_geeks_fine-tune_PTX_to_maximize_GPU_performance\" title=\"\u5929\u624d\u6781\u5ba2\u5fae\u8c03 PTX\uff0c\u6700\u5927\u9650\u5ea6\u63d0\u9ad8 GPU \u6027\u80fd\">\u5929\u624d\u6781\u5ba2\u5fae\u8c03 PTX\uff0c\u6700\u5927\u9650\u5ea6\u63d0\u9ad8 GPU \u6027\u80fd<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/janusai.pro\/zh_cn\/deepseek-v3-paper-details-how-to-bypass-the-cuda-monopoly\/#PTX_and_CUDA\" title=\"PTX \u548c CUDA\">PTX \u548c CUDA<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/janusai.pro\/zh_cn\/deepseek-v3-paper-details-how-to-bypass-the-cuda-monopoly\/#However_the_technical_barriers_remain\" title=\"\u7136\u800c\uff0c\u6280\u672f\u969c\u788d\u4f9d\u7136\u5b58\u5728\">\u7136\u800c\uff0c\u6280\u672f\u969c\u788d\u4f9d\u7136\u5b58\u5728<\/a><\/li><\/ul><\/nav><\/div>\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Genius_geeks_fine-tune_PTX_to_maximize_GPU_performance\"><\/span>\u5929\u624d\u6781\u5ba2\u5fae\u8c03 PTX\uff0c\u6700\u5927\u9650\u5ea6\u63d0\u9ad8 GPU \u6027\u80fd<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p>\u82f1\u4f1f\u8fbe PTX\uff08\u5e76\u884c\u7ebf\u7a0b\u6267\u884c\uff09\u662f\u4e13\u4e3a\u82f1\u4f1f\u8fbe GPU \u8bbe\u8ba1\u7684\u4e2d\u95f4\u6307\u4ee4\u96c6\u67b6\u6784\uff0c\u4ecb\u4e8e\u9ad8\u7ea7 GPU \u7f16\u7a0b\u8bed\u8a00\uff08\u5982 CUDA C\/C++\uff09\u6216\u5176\u4ed6\u8bed\u8a00\u524d\u7aef\u4e0e\u4f4e\u7ea7\u673a\u5668\u4ee3\u7801\uff08\u6d41\u5f0f\u6c47\u7f16\u6216 SASS\uff09\u4e4b\u95f4\u3002<\/p>\n\n\n\n<p>PTX \u662f\u4e00\u79cd\u4f4e\u7ea7\u6307\u4ee4\u96c6\u67b6\u6784\uff0c\u5c06 GPU \u4f5c\u4e3a\u6570\u636e\u5e76\u884c\u8ba1\u7b97\u8bbe\u5907\uff0c\u5b9e\u73b0\u4e86\u5bc4\u5b58\u5668\u5206\u914d\u548c\u7ebf\u7a0b\/\u7ebf\u7a0b\u675f\u7ea7\u8c03\u6574\u7b49\u7ec6\u7c92\u5ea6\u4f18\u5316\uff0c\u800c CUDA C\/C++ \u7b49\u8bed\u8a00\u65e0\u6cd5\u5b9e\u73b0\u8fd9\u4e9b\u4f18\u5316\u3002<\/p>\n\n\n\n<p>\u5c06 PTX \u8f6c\u6362\u4e3a SASS \u65f6\uff0c\u4f1a\u9488\u5bf9\u7279\u5b9a\u4e00\u4ee3\u82f1\u4f1f\u8fbe\u2122\uff08NVIDIA\u00ae\uff09GPU \u8fdb\u884c\u4f18\u5316\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/rmy9ct2fln.feishu.cn\/space\/api\/box\/stream\/download\/asynccode\/?code=MGIwZTQ0ZDdhMTgxYTBlMmEzZDE5OTczN2ZlZmEzNGFfc3I3T2U0UzNCOGdjd1ZHcktHd1hkd1RpcXlLbkxrU1FfVG9rZW46Vk05WWJ0a1Bob3NkYzl4bXpFc2N6anI3bktjXzE3MzgxNTE4NjQ6MTczODE1NTQ2NF9WNA\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u5728\u8bad\u7ec3 V3 \u6a21\u578b\u65f6\uff0cDeepSeek \u91cd\u65b0\u914d\u7f6e\u4e86\u82f1\u4f1f\u8fbe H800 GPU\uff1a<\/p>\n\n\n\n<p>\u5728 132 \u4e2a\u6d41\u5904\u7406\u5668\u5185\u6838\u4e2d\uff0c\u6709 20 \u4e2a\u88ab\u5206\u914d\u7528\u4e8e\u670d\u52a1\u5668\u95f4\u901a\u4fe1\uff0c\u4e3b\u8981\u7528\u4e8e\u6570\u636e\u538b\u7f29\u548c\u89e3\u538b\u7f29\uff0c\u4ee5\u7a81\u7834\u5904\u7406\u5668\u7684\u8fde\u63a5\u9650\u5236\uff0c\u63d0\u9ad8\u4e8b\u52a1\u5904\u7406\u901f\u5ea6\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u6027\u80fd\uff0cDeepSeek \u8fd8\u901a\u8fc7\u989d\u5916\u7684\u7ec6\u7c92\u5ea6\u7ebf\u7a0b\/\u7ebf\u7a0b\u675f\u7ea7\u522b\u8c03\u6574\uff0c\u5b9e\u65bd\u4e86\u5148\u8fdb\u7684\u6d41\u6c34\u7ebf\u7b97\u6cd5\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e9b\u4f18\u5316\u8fdc\u8fdc\u8d85\u51fa\u4e86\u4f20\u7edf\u7684 CUDA \u5f00\u53d1\u6c34\u5e73\uff0c\u4f46\u5374\u6781\u96be\u7ef4\u62a4\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u4f18\u5316\u6c34\u5e73\u6b63\u662f DeepSeek \u56e2\u961f\u6770\u51fa\u6280\u672f\u80fd\u529b\u7684\u5145\u5206\u4f53\u73b0\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/rmy9ct2fln.feishu.cn\/space\/api\/box\/stream\/download\/asynccode\/?code=MDk2ZDMyM2IzOGU5OWNmN2JhOTY2ZDZmMjhkOWYwZGFfUnFiV0hvbnQ0ZUFHSHg3WHpyMW5jYTRvMURPM1pDSTZfVG9rZW46QnZVNWJyUzBDb2FWeE54Ym4ybGNZNXlnbmFnXzE3MzgxNTE4NjQ6MTczODE1NTQ2NF9WNA\" alt=\"\"\/><\/figure>\n\n\n\n<p>V3 \u6587\u4ef6\u7279\u522b\u63d0\u5230\u4e86 PTX \u7684\u7ec6\u8282<\/p>\n\n\n\n<p>\u8fd9\u662f\u56e0\u4e3a\uff0c\u5728\u5168\u7403 GPU \u77ed\u7f3a\u548c\u7f8e\u56fd\u9650\u5236\u7684\u53cc\u91cd\u538b\u529b\u4e0b\uff0cDeepSeek \u7b49\u516c\u53f8\u4e0d\u5f97\u4e0d\u5bfb\u6c42\u521b\u65b0\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n\n\n\n<p>\u5e78\u8fd0\u7684\u662f\uff0c\u4ed6\u4eec\u5728\u8fd9\u4e00\u9886\u57df\u53d6\u5f97\u4e86\u91cd\u5927\u7a81\u7834\u3002<\/p>\n\n\n\n<p>\u4e00\u4f4d\u5f00\u53d1\u4eba\u5458\u8ba4\u4e3a\uff0c\"\u4f4e\u7ea7 GPU \u7f16\u7a0b\u662f\u6b63\u786e\u7684\u65b9\u5411\u3002\u4f18\u5316\u8d8a\u591a\uff0c\u6210\u672c\u5c31\u8d8a\u4f4e\uff0c\u6216\u8005\u8bf4\u6027\u80fd\u9884\u7b97\u5c31\u8d8a\u4f4e\uff0c\u5c31\u53ef\u4ee5\u7528\u4e8e\u5176\u4ed6\u8fdb\u5c55\uff0c\u800c\u65e0\u9700\u989d\u5916\u652f\u51fa\"\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/rmy9ct2fln.feishu.cn\/space\/api\/box\/stream\/download\/asynccode\/?code=MmEzYzA5ZTVmNjE4ZTlhMWE0NWU1ZTgyZTA2NmUxMDJfUWdNb21QeEFtUWlFSFA1aGFWZEZJMzlUNjdPT3J5NXRfVG9rZW46RWtaaGJ2UlBHbzk2VWF4TmxkeGNPeGdKblJnXzE3MzgxNTE4NjQ6MTczODE1NTQ2NF9WNA\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u8fd9\u4e00\u7a81\u7834\u5bf9\u5e02\u573a\u4ea7\u751f\u4e86\u91cd\u5927\u5f71\u54cd\uff0c\u4e00\u4e9b\u6295\u8d44\u8005\u8ba4\u4e3a\uff0c\u65b0\u6a21\u5f0f\u5c06\u51cf\u5c11\u5bf9\u9ad8\u6027\u80fd\u786c\u4ef6\u7684\u9700\u6c42\uff0c\u4ece\u800c\u53ef\u80fd\u5f71\u54cd\u82f1\u4f1f\u8fbe\u7b49\u516c\u53f8\u7684\u9500\u552e\u4e1a\u7ee9\u3002<\/p>\n\n\n\n<p>\u4e0d\u8fc7\uff0c\u5305\u62ec\u82f1\u7279\u5c14\u524d\u9996\u5e2d\u6267\u884c\u5b98\u5e15\u7279-\u76d6\u5c14\u8f9b\u683c\u5728\u5185\u7684\u4e1a\u5185\u8d44\u6df1\u4eba\u58eb\u8ba4\u4e3a\uff0c\u4eba\u5de5\u667a\u80fd\u5e94\u7528\u53ef\u4ee5\u5145\u5206\u5229\u7528\u6240\u6709\u53ef\u7528\u7684\u8ba1\u7b97\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u76d6\u5c14\u8f9b\u683c\u8ba4\u4e3a\uff0cDeepSeek \u7684\u8fd9\u4e00\u7a81\u7834\u662f\u5728\u9762\u5411\u5927\u4f17\u5e02\u573a\u7684\u4f4e\u6210\u672c\u8bbe\u5907\u4e2d\u5d4c\u5165\u4eba\u5de5\u667a\u80fd\u529f\u80fd\u7684\u65b0\u9014\u5f84\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/rmy9ct2fln.feishu.cn\/space\/api\/box\/stream\/download\/asynccode\/?code=NzgzZjM2ZTVlOWM0OWI1MDE5OTI1NTQwNWRjYTI5Y2NfZ25sc2tPNFJ1UHZwemp1WEVlclU1cloxZXI5aHJMbEZfVG9rZW46SHlGTGJnNHpHbzNzbnd4bkxPQ2N4T0RyblZkXzE3MzgxNTE4NjQ6MTczODE1NTQ2NF9WNA\" alt=\"\"\/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"PTX_and_CUDA\"><\/span>PTX \u548c CUDA<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p>\u90a3\u4e48\uff0cDeepSeek \u7684\u51fa\u73b0\u662f\u5426\u610f\u5473\u7740\u5f00\u53d1\u5c16\u7aef LLM \u4e0d\u518d\u9700\u8981\u5927\u89c4\u6a21 GPU \u96c6\u7fa4\uff1f<\/p>\n\n\n\n<p>\u8c37\u6b4c\u5728\u8ba1\u7b97\u8d44\u6e90\u65b9\u9762\u7684\u5de8\u5927\u6295\u8d44\u3001 <a href=\"https:\/\/openai.com\/\" target=\"_blank\" rel=\"noreferrer noopener nofollow\">OpenAI<\/a>Meta \u548c xAI \u6700\u7ec8\u4f1a\u88ab\u6d6a\u8d39\u6389\u5417\uff1f\u4eba\u5de5\u667a\u80fd\u5f00\u53d1\u4eba\u5458\u666e\u904d\u8ba4\u4e3a\u60c5\u51b5\u5e76\u975e\u5982\u6b64\u3002<\/p>\n\n\n\n<p>\u4f46\u53ef\u4ee5\u80af\u5b9a\u7684\u662f\uff0c\u5728\u6570\u636e\u5904\u7406\u548c\u7b97\u6cd5\u4f18\u5316\u65b9\u9762\u4ecd\u6709\u5de8\u5927\u7684\u6f5c\u529b\u53ef\u6316\uff0c\u672a\u6765\u4e00\u5b9a\u4f1a\u51fa\u73b0\u66f4\u591a\u521b\u65b0\u7684\u4f18\u5316\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p>\u968f\u7740 DeepSeek V3 \u578b\u53f7\u7684\u5f00\u6e90\uff0c\u5176\u6280\u672f\u62a5\u544a\u4e2d\u62ab\u9732\u4e86\u8be6\u7ec6\u7684\u7ec6\u8282\u3002<\/p>\n\n\n\n<p>\u62a5\u544a\u8bb0\u5f55\u4e86 DeepSeek \u8fdb\u884c\u7684\u6df1\u5ea6\u5e95\u5c42\u4f18\u5316\u3002\u7b80\u800c\u8a00\u4e4b\uff0c\u4f18\u5316\u7a0b\u5ea6\u53ef\u4ee5\u6982\u62ec\u4e3a \"\u4ed6\u4eec\u4ece\u5934\u5f00\u59cb\u91cd\u5efa\u4e86\u6574\u4e2a\u7cfb\u7edf\"\u3002<\/p>\n\n\n\n<p>\u5982\u4e0a\u6240\u8ff0\uff0c\u5728\u4f7f\u7528 H800 GPU \u8bad\u7ec3 V3 \u65f6\uff0cDeepSeek \u5b9a\u5236\u4e86 GPU \u7684\u6838\u5fc3\u8ba1\u7b97\u5355\u5143\uff08\u6d41\u5f0f\u591a\u5904\u7406\u5668\u5185\u6838\u6216 SM\uff09\uff0c\u4ee5\u6ee1\u8db3\u7279\u5b9a\u9700\u6c42\u3002<\/p>\n\n\n\n<p>\u5728\u603b\u5171 132 \u4e2a SM \u4e2d\uff0c\u4ed6\u4eec\u7279\u522b\u5206\u914d\u4e86 20 \u4e2a\u7528\u4e8e\u5904\u7406\u670d\u52a1\u5668\u4e4b\u95f4\u7684\u901a\u4fe1\u4efb\u52a1\uff0c\u800c\u4e0d\u662f\u8ba1\u7b97\u4efb\u52a1\u3002<\/p>\n\n\n\n<p>\u8fd9\u79cd\u5b9a\u5236\u662f\u5728 PTX\uff08\u5e76\u884c\u7ebf\u7a0b\u6267\u884c\uff09\u7ea7\u522b\u5b8c\u6210\u7684\uff0cPTX \u662f\u82f1\u4f1f\u8fbe\u2122\uff08NVIDIA\u00ae\uff09\u56fe\u5f62\u5904\u7406\u5668\u7684\u5e95\u5c42\u6307\u4ee4\u96c6\u3002<\/p>\n\n\n\n<p>PTX \u5728\u63a5\u8fd1\u6c47\u7f16\u8bed\u8a00\u7684\u7ea7\u522b\u4e0a\u8fd0\u884c\uff0c\u53ef\u8fdb\u884c\u7ec6\u7c92\u5ea6\u4f18\u5316\uff0c\u5982\u5bc4\u5b58\u5668\u5206\u914d\u548c\u7ebf\u7a0b\/\u7ebf\u7a0b\u675f\u7ea7\u522b\u8c03\u6574\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u7cbe\u7ec6\u63a7\u5236\u65e2\u590d\u6742\u53c8\u96be\u4ee5\u7ef4\u62a4\u3002<\/p>\n\n\n\n<p>\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u5f00\u53d1\u4eba\u5458\u901a\u5e38\u66f4\u559c\u6b22\u4f7f\u7528 CUDA \u7b49\u9ad8\u7ea7\u7f16\u7a0b\u8bed\u8a00\u7684\u539f\u56e0\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u8bed\u8a00\u53ef\u4e3a\u5927\u591a\u6570\u5e76\u884c\u7f16\u7a0b\u4efb\u52a1\u63d0\u4f9b\u8db3\u591f\u7684\u6027\u80fd\u4f18\u5316\uff0c\u800c\u4e14\u65e0\u9700\u8fdb\u884c\u5e95\u5c42\u4f18\u5316\u3002<\/p>\n\n\n\n<p>\u7136\u800c\uff0c\u5f53\u9700\u8981\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8 GPU \u8d44\u6e90\u7684\u6548\u7387\u5e76\u5b9e\u73b0\u7279\u5b9a\u7684\u4f18\u5316\u8981\u6c42\u65f6\uff0c\u5f00\u53d1\u4eba\u5458\u4e0d\u5f97\u4e0d\u6c42\u52a9\u4e8e PTX\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"However_the_technical_barriers_remain\"><\/span>\u7136\u800c\uff0c\u6280\u672f\u969c\u788d\u4f9d\u7136\u5b58\u5728<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p>\u5bf9\u6b64\uff0c\u7f51\u6c11 Ian Cutress \u8bf4\uff1a\"Deepseek \u4f7f\u7528 PTX \u5e76\u4e0d\u80fd\u6d88\u9664 CUDA \u7684\u6280\u672f\u969c\u788d\"\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/rmy9ct2fln.feishu.cn\/space\/api\/box\/stream\/download\/asynccode\/?code=YTFkNWFkMTNiYjQzNDZiMDI3ZmYxYjA3MzExYjE1MGRfemNRaFdmM1R4MTMwUWVWTUxxbHN2SjZYNEhvazBrZlNfVG9rZW46SFlEU2IwNEd3b29kMGl4cmVaOGNTcHFZbmxjXzE3MzgxNTE4NjQ6MTczODE1NTQ2NF9WNA\" alt=\"\"\/><\/figure>\n\n\n\n<p>CUDA \u662f\u4e00\u79cd\u9ad8\u7ea7\u8bed\u8a00\u3002\u5b83\u4f7f\u4f7f\u7528\u82f1\u4f1f\u8fbe\u2122\uff08NVIDIA\u00ae\uff09GPU \u5f00\u53d1\u5e93\u548c\u63a5\u53e3\u53d8\u5f97\u66f4\u5bb9\u6613\uff0c\u5e76\u652f\u6301\u5feb\u901f\u8fed\u4ee3\u5f00\u53d1\u3002<\/p>\n\n\n\n<p>CUDA \u53ef\u4ee5\u901a\u8fc7\u5fae\u8c03\u5e95\u5c42\u4ee3\u7801\uff08\u5373 PTX\uff09\u6765\u4f18\u5316\u6027\u80fd\uff0c\u800c\u4e14\u57fa\u672c\u5e93\u5df2\u7ecf\u5b8c\u6210\u3002\u76ee\u524d\uff0c\u5927\u591a\u6570\u751f\u4ea7\u7ea7\u8f6f\u4ef6\u90fd\u662f\u57fa\u4e8e CUDA \u6784\u5efa\u7684\u3002<\/p>\n\n\n\n<p>PTX \u66f4\u7c7b\u4f3c\u4e8e GPU \u53ef\u76f4\u63a5\u7406\u89e3\u7684\u6c47\u7f16\u8bed\u8a00\u3002\u5b83\u5728\u4f4e\u5c42\u6b21\u4e0a\u5de5\u4f5c\uff0c\u5141\u8bb8\u5fae\u89c2\u5c42\u9762\u7684\u4f18\u5316\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u9009\u62e9\u4f7f\u7528 PTX \u7f16\u7a0b\uff0c\u5c31\u610f\u5473\u7740\u65e0\u6cd5\u4f7f\u7528\u4e0a\u8ff0\u5185\u7f6e\u7684 CUDA \u5e93\u3002\u8fd9\u662f\u4e00\u9879\u975e\u5e38\u7e41\u7410\u7684\u4efb\u52a1\uff0c\u9700\u8981\u6df1\u5165\u4e86\u89e3\u786c\u4ef6\u548c\u8fd0\u884c\u65f6\u95ee\u9898\u7684\u4e13\u4e1a\u77e5\u8bc6\u3002<\/p>\n\n\n\n<p>\u4e0d\u8fc7\uff0c\u5982\u679c\u5f00\u53d1\u4eba\u5458\u5b8c\u5168\u4e86\u89e3\u81ea\u5df1\u5728\u505a\u4ec0\u4e48\uff0c\u4ed6\u4eec\u786e\u5b9e\u53ef\u4ee5\u5728\u8fd0\u884c\u65f6\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\u548c\u4f18\u5316\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/rmy9ct2fln.feishu.cn\/space\/api\/box\/stream\/download\/asynccode\/?code=MGU3N2MyY2Y5NDg0MzAxYjkzNzI4MDg3YTRjN2JiNjJfSTI4Um5wZkxwamJMNjRLdmx6TnFmcTlhVDhLbTEyYlhfVG9rZW46SVpVcWJ4TmRtbzdYRjF4RFk5SWN4OWdjbmRlXzE3MzgxNTE4NjQ6MTczODE1NTQ2NF9WNA\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u76ee\u524d\uff0c\u82f1\u4f1f\u8fbe\u751f\u6001\u7cfb\u7edf\u7684\u4e3b\u6d41\u4ecd\u7136\u662f\u4f7f\u7528 CUDA\u3002<\/p>\n\n\n\n<p>\u90a3\u4e9b\u5e0c\u671b\u4ece\u8ba1\u7b97\u8d1f\u8f7d\u4e2d\u83b7\u5f97\u989d\u5916 10-20% \u6027\u80fd\u6216\u80fd\u6548\u7684\u5f00\u53d1\u4eba\u5458\uff0c\u5982\u5728\u4e91\u4e2d\u90e8\u7f72\u6a21\u578b\u548c\u9500\u552e\u4ee4\u724c\u670d\u52a1\u7684\u516c\u53f8\uff0c\u786e\u5b9e\u5df2\u7ecf\u4ece CUDA \u7ea7\u522b\u4f18\u5316\u5230\u4e86 PTX \u7ea7\u522b\u3002\u4ed6\u4eec\u613f\u610f\u6295\u5165\u65f6\u95f4\uff0c\u56e0\u4e3a\u4ece\u957f\u8fdc\u6765\u770b\u8fd9\u662f\u503c\u5f97\u7684\u3002<\/p>\n\n\n\n<p>\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0cPTX \u901a\u5e38\u9488\u5bf9\u7279\u5b9a\u7684\u786c\u4ef6\u578b\u53f7\u8fdb\u884c\u4e86\u4f18\u5316\uff0c\u9664\u975e\u4e13\u95e8\u7f16\u5199\u4e86\u9002\u914d\u903b\u8f91\uff0c\u5426\u5219\u5f88\u96be\u5728\u4e0d\u540c\u7684\u786c\u4ef6\u4e4b\u95f4\u8fdb\u884c\u79fb\u690d\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u624b\u52a8\u8c03\u6574\u8ba1\u7b97\u5185\u6838\u9700\u8981\u6781\u5927\u7684\u6bc5\u529b\u3001\u52c7\u6c14\u548c\u4fdd\u6301\u51b7\u9759\u7684\u7279\u6b8a\u80fd\u529b\uff0c\u56e0\u4e3a\u7a0b\u5e8f\u53ef\u80fd\u6bcf 5000 \u4e2a\u5468\u671f\u5c31\u4f1a\u51fa\u73b0\u4e00\u6b21\u5185\u5b58\u8bbf\u95ee\u9519\u8bef\u3002<\/p>\n\n\n\n<p>\u5f53\u7136\uff0c\u5bf9\u4e8e\u90a3\u4e9b\u786e\u5b9e\u9700\u8981 PTX \u7684\u60c5\u51b5\uff0c\u4ee5\u53ca\u90a3\u4e9b\u6709\u8db3\u591f\u62a5\u916c\u6765\u5904\u7406\u8fd9\u4e9b\u95ee\u9898\u7684\u5f00\u53d1\u4eba\u5458\uff0c\u6211\u4eec\u8868\u793a\u5145\u5206\u7684\u7406\u89e3\u548c\u5c0a\u91cd\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u6240\u6709\u5176\u4ed6\u5f00\u53d1\u4eba\u5458\uff0c\u5efa\u8bae\u7ee7\u7eed\u4f7f\u7528 CUDA \u6216\u5176\u4ed6\u57fa\u4e8e CUDA \u7684\u9ad8\u7ea7\u53d8\u4f53\uff08\u6216 MLIR\uff09\u3002<\/p>","protected":false},"excerpt":{"rendered":"<p>DeepSeek V3\u8bba\u6587\u8be6\u60c5\uff1a\u5982\u4f55\u7ed5\u8fc7 CUDA \u5784\u65ad\uff01DeepSeek\u6700\u8fd1\u53d1\u5e03\u7684\u4e24\u4e2a\u6a21\u578bDeepSeek-V3\u548cDeepSeek-R1\uff0c\u4ee5\u66f4\u4f4e\u7684\u6210\u672c\u5b9e\u73b0\u4e86\u4e0eOpenAI\u7c7b\u4f3c\u6a21\u578b\u76f8\u5f53\u7684\u6027\u80fd\u3002\u636e\u56fd\u5916\u5a92\u4f53\u62a5\u9053\uff0c\u5728\u77ed\u77ed\u4e24\u4e2a\u6708\u7684\u65f6\u95f4\u91cc\uff0c\u4ed6\u4eec\u5c31\u5728\u4e00\u4e2a\u7531 2048 \u4e2a\u96c6\u7fa4\u7ec4\u6210\u7684\u6a21\u578b\u96c6\u7fa4\u4e0a\u8bad\u7ec3\u51fa\u4e86\u4e00\u4e2a\u62e5\u6709 6710 \u4ebf\u4e2a\u53c2\u6570\u7684 MoE \u8bed\u8a00\u6a21\u578b\u3002<\/p>","protected":false},"author":2,"featured_media":684,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_kadence_starter_templates_imported_post":false,"_kad_post_transparent":"","_kad_post_title":"","_kad_post_layout":"","_kad_post_sidebar_id":"","_kad_post_content_style":"","_kad_post_vertical_padding":"","_kad_post_feature":"","_kad_post_feature_position":"","_kad_post_header":false,"_kad_post_footer":false,"footnotes":""},"categories":[1],"tags":[],"class_list":["post-710","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/posts\/710","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/comments?post=710"}],"version-history":[{"count":1,"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/posts\/710\/revisions"}],"predecessor-version":[{"id":711,"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/posts\/710\/revisions\/711"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/media\/684"}],"wp:attachment":[{"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/media?parent=710"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/categories?post=710"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/janusai.pro\/zh_cn\/wp-json\/wp\/v2\/tags?post=710"}],"curies":[{"name":"\u53ef\u6e7f\u6027\u7c89\u5242","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}