{"id":1252,"date":"2025-06-20T18:47:00","date_gmt":"2025-06-20T10:47:00","guid":{"rendered":"http:\/\/www.applicationsofllm.com\/?p=1252"},"modified":"2025-09-17T19:51:24","modified_gmt":"2025-09-17T11:51:24","slug":"deepseek%e5%88%9a%e6%8f%90%e5%88%b0fp8%ef%bc%8c%e8%8b%b1%e4%bc%9f%e8%be%be%e5%b0%b1%e6%8a%8afp4%e7%b2%be%e5%ba%a6%e6%8e%a8%e5%90%91%e9%a2%84%e8%ae%ad%e7%bb%83%ef%bc%8c%e6%9b%b4%e5%bf%ab%e3%80%81","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/06\/20\/deepseek%e5%88%9a%e6%8f%90%e5%88%b0fp8%ef%bc%8c%e8%8b%b1%e4%bc%9f%e8%be%be%e5%b0%b1%e6%8a%8afp4%e7%b2%be%e5%ba%a6%e6%8e%a8%e5%90%91%e9%a2%84%e8%ae%ad%e7%bb%83%ef%bc%8c%e6%9b%b4%e5%bf%ab%e3%80%81\/","title":{"rendered":"DeepSeek\u521a\u63d0\u5230FP8\uff0c\u82f1\u4f1f\u8fbe\u5c31\u628aFP4\u7cbe\u5ea6\u63a8\u5411\u9884\u8bad\u7ec3\uff0c\u66f4\u5feb\u3001\u66f4\u4fbf\u5b9c"},"content":{"rendered":"\n<p>\u524d\u4e9b\u5929\uff0cDeepSeek \u5728\u53d1\u5e03 DeepSeek V3.1 \u7684\u6587\u7ae0\u8bc4\u8bba\u533a\u4e2d\uff0c\u63d0\u53ca\u4e86&nbsp;&nbsp;\u7684\u91cf\u5316\u8bbe\u8ba1\uff0c\u58f0\u79f0\u662f\u9488\u5bf9\u5373\u5c06\u53d1\u5e03\u7684\u4e0b\u4e00\u4ee3\u56fd\u4ea7\u82af\u7247\u8bbe\u8ba1\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"137\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-465-1024x137.jpg\" alt=\"\" class=\"wp-image-1253\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-465-1024x137.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-465-300x40.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-465-768x103.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-465.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u8fd9\u4ef6\u4e8b\u4e00\u4e0b\u5f15\u53d1\u4e86\u5de8\u5927\u53cd\u54cd\uff0c\u4e0d\u4ec5\u662f\u5173\u4e8e\u65b0\u4e00\u4ee3\u56fd\u4ea7\u82af\u7247\u8bbe\u8ba1\u3001\u5927\u6a21\u578b\u5728\u56fd\u4ea7\u82af\u7247\u8bad\u7ec3\u7684\u8bdd\u9898\uff0c\u4e5f\u987a\u52bf\u5f15\u53d1\u4e86\u5927\u5bb6\u5bf9\u5927\u6a21\u578b\u91cf\u5316\u7b56\u7565\u7684\u5173\u6ce8\u3002<\/p>\n\n\n\n<p>FP8\uff0c\u5176\u5168\u79f0\u4e3a 8-bit floating point\uff088 \u4f4d\u6d6e\u70b9\u6570\uff09\uff0c\u662f\u4e00\u79cd<strong>\u8d85\u4f4e\u7cbe\u5ea6<\/strong>\u7684\u6570\u636e\u8868\u793a\u683c\u5f0f\uff0c\u76f8\u8f83\u4e8e FP32\uff08\u5355\u7cbe\u5ea6\uff09\u6216 FP16\uff08\u534a\u7cbe\u5ea6\uff09\u7b49\u4f20\u7edf\u6d6e\u70b9\u683c\u5f0f\uff0cFP8 \u53ef\u4ee5\u5728\u5c3d\u91cf\u4fdd\u6301\u6570\u503c\u7a33\u5b9a\u6027\u548c\u6a21\u578b\u7cbe\u5ea6\u7684\u524d\u63d0\u4e0b\uff0c\u8fdb\u4e00\u6b65\u964d\u4f4e\u5b58\u50a8\u548c\u8ba1\u7b97\u5f00\u9500\uff08\u53c2\u89c1\u673a\u5668\u4e4b\u5fc3\u6587\u7ae0\uff1a)\u3002<\/p>\n\n\n\n<p>\u5728\u82f1\u4f1f\u8fbe\u4e4b\u5916\uff0c\u5fae\u8f6f\u3001Meta\u3001\u82f1\u7279\u5c14\u3001AMD \u7b49\u4e5f\u90fd\u5728\u7814\u7a76 FP8 \u8bad\u7ec3\u4e0e\u63a8\u7406\uff0c\u6709\u6210\u4e3a\u4e1a\u754c\u300c\u65b0\u9ec4\u91d1\u6807\u51c6\u300d\u7684\u8d8b\u52bf\u3002<\/p>\n\n\n\n<p>\u5982\u4eca\uff0cDeepSeek \u91c7\u7528\u975e\u4e3b\u6d41\u7684 FP8 \u91cf\u5316\u7b56\u7565\uff0c\u9690\u9690\u5c55\u73b0\u51fa\u56fd\u4ea7\u5927\u6a21\u578b\u4e0e\u56fd\u4ea7\u82af\u7247\u82af\u7247\u8f6f\u786c\u7ed3\u5408\u7684\u4f18\u5316\u7b56\u7565\u4e0e\u82f1\u4f1f\u8fbe\u7684\u9ad8\u517c\u5bb9\u7b56\u7565\u7684\u4e0d\u540c\u53d1\u5c55\u8def\u5f84\u3002<\/p>\n\n\n\n<p>UE8M0 FP8 \u5177\u6709\u9c9c\u660e\u7684\u6218\u7565\u610f\u4e49\u3002DeepSeek \u9009\u62e9\u5728\u6a21\u578b\u7aef\u7387\u5148\u91c7\u7528\u5e76\u516c\u5f00\u58f0\u660e\u4f7f\u7528 UE8M0 \u683c\u5f0f\uff0c\u5c06\u5176\u8bad\u7ec3\u4e0e scale \u7b56\u7565\u4e0e\u8be5\u7cbe\u5ea6\u7ed1\u5b9a\u3002\u8fd9\u7b49\u4e8e\u7531\u5927\u6a21\u578b\u7aef\u4e3b\u52a8\u63d0\u51fa\u6807\u51c6\uff0c\u8feb\u4f7f\u786c\u4ef6\u548c\u5de5\u5177\u94fe\u8fdb\u884c\u9002\u914d\uff0c\u52a0\u901f\u4e86\u56fd\u4ea7\u8f6f\u786c\u4ef6\u4e00\u4f53\u5316\u7684\u751f\u6001\u5efa\u8bbe\u3002<\/p>\n\n\n\n<p>\u4e0d\u77e5\u9053\u662f\u4e0d\u662f\u5de7\u5408\uff0c\u5728 DeepSeek \u4e3a\u56fd\u4ea7\u82af\u7247\u51c6\u5907\u7684 FP8 \u91cf\u5316\u7b56\u7565\u7684\u63d0\u51fa\u4e0d\u4e45\uff0c\u5c31\u5728\u4eca\u5929\uff0c\u82f1\u4f1f\u8fbe\u4e5f\u5728\u4f4e\u7cbe\u5ea6\u91cf\u5316\u9886\u57df\u518d\u6b21\u53d1\u529b\u3002\u53ea\u4e0d\u8fc7\u8fd9\u6b21\u4e0d\u662f FP8 \u91cf\u5316\u7684\u65b0\u8fdb\u5c55\uff0c\u800c\u662f\u5411 FP4 \u91cf\u5316\u8dc3\u8fdb\u3002<\/p>\n\n\n\n<p>\u82f1\u4f1f\u8fbe\u5c06<strong>\u5176\u6700\u65b0\u7684 NVFP4 \u7b56\u7565\u62d3\u5c55\u5230\u9884\u8bad\u7ec3\u9636\u6bb5<\/strong>\uff0c\u58f0\u79f0\u80fd\u591f\u4ee5\u5339\u914d 16 \u4f4d\u7684\u7cbe\u5ea6\u8fdb\u884c\u8bad\u7ec3\uff0c\u5e76\u4ee5 4 \u4f4d\u7684\u901f\u5ea6\u548c\u6548\u7387\u8fd0\u884c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"507\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-470-1024x507.jpg\" alt=\"\" class=\"wp-image-1258\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-470-1024x507.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-470-300x149.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-470-768x381.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-470.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u82f1\u4f1f\u8fbe\u79f0\uff1a\u300c\u5728\u9884\u8bad\u7ec3\u4e2d\u4f7f\u7528 NVFP4\uff0c\u53ef\u663e\u8457\u63d0\u5347\u5927\u89c4\u6a21 LLM \u8bad\u7ec3\u6548\u7387\u548c\u57fa\u7840\u8bbe\u65bd\u6548\u80fd\u3002\u8fd9\u4e0d\u4ec5\u662f\u4e00\u6b21\u6e10\u8fdb\u5f0f\u4f18\u5316\uff0c\u800c\u662f\u4e00\u79cd\u91cd\u65b0\u5b9a\u4e49\u5927\u89c4\u6a21\u6a21\u578b\u8bad\u7ec3\u65b9\u5f0f\u7684\u6839\u672c\u6027\u8f6c\u53d8\u3002\u300d<\/p>\n\n\n\n<p>\u5728\u300cAI \u5de5\u5382\u300d\u65f6\u4ee3\uff0c\u7b97\u529b\u662f\u8fdb\u6b65\u7684\u5f15\u64ce\uff0c\u6570\u503c\u7cbe\u5ea6\u5df2\u4e0d\u518d\u662f\u540e\u7aef\u7ec6\u8282\uff0c\u800c\u662f\u4e00\u79cd\u6218\u7565\u4f18\u52bf\u3002NVFP4 4 \u6bd4\u7279\u9884\u8bad\u7ec3\u4e3a\u6548\u7387\u4e0e\u53ef\u6269\u5c55\u6027\u8bbe\u5b9a\u4e86\u65b0\u7684\u6807\u51c6\uff0c\u63a8\u52a8\u9ad8\u6027\u80fd AI \u6a21\u578b\u5f00\u53d1\u8fdb\u5165\u5168\u65b0\u9636\u6bb5\u3002<\/p>\n\n\n\n<p>\u76ee\u524d\uff0cNVFP4 \u8bad\u7ec3\u4ecd\u5904\u4e8e\u7814\u7a76\u9636\u6bb5\uff0c\u6b63\u5728\u63a2\u7d22\u5e76\u9a8c\u8bc1 4 \u4f4d\u7cbe\u5ea6\u5728\u5927\u89c4\u6a21\u6a21\u578b\u9884\u8bad\u7ec3\u4e2d\u7684\u6f5c\u529b\u3002\u56f4\u7ed5 NVFP4 \u7684\u5408\u4f5c\u4e0e\u5b9e\u9a8c\u6b63\u79ef\u6781\u63a8\u8fdb\uff0c\u53c2\u4e0e\u65b9\u5305\u62ec AWS\u3001Cohere\u3001Google Cloud\u3001Kimi AI\u3001Microsoft AI\u3001Mistral\u3001OpenAI\u3001Perplexity\u3001Reflection\u3001Runway \u7b49\u9886\u5148\u7ec4\u7ec7\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u82f1\u4f1f\u8fbe\u5728\u66f4\u4f4e\u4f4d\u7684\u63a2\u7d22\uff0c\u8bc4\u8bba\u533a\u7684\u7f51\u53cb\u610f\u89c1\u4e0d\u4e00\uff0c\u6709\u4eba\u8ba4\u53ef NVFP4 \u5728\u63d0\u5347\u8bad\u7ec3\u901f\u5ea6\u4ee5\u53ca\u964d\u4f4e\u6210\u672c\u548c\u80fd\u8017\u65b9\u9762\u7684\u79ef\u6781\u4f5c\u7528\uff0c\u8ba4\u4e3a\u5176\u6709\u671b\u63a8\u52a8\u66f4\u591a\u884c\u4e1a\u8fdb\u5165\u9ad8\u6548\u3001\u53ef\u6301\u7eed\u7684 AI \u65f6\u4ee3\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"317\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-469-1024x317.jpg\" alt=\"\" class=\"wp-image-1257\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-469-1024x317.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-469-300x93.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-469-768x238.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-469.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u8fd8\u6709\u4eba\u63d0\u5230 NVFP4 \u4e0e Jetson Thor \u7684\u7ed3\u5408\u6709\u671b\u5bf9\u73b0\u5b9e\u4e16\u754c\u7684\u5e94\u7528\u4ea7\u751f\u6df1\u8fdc\u5f71\u54cd\u3002Jetson Thor \u662f\u82f1\u4f1f\u8fbe\u524d\u51e0\u65e5\u53d1\u5e03\u7684\u65b0\u4e00\u4ee3\u673a\u5668\u4eba\u4e13\u7528\u82af\u7247\uff0c\u901a\u8fc7\u5927\u5e45\u63d0\u5347\u7b97\u529b\uff0c\u53ef\u4ee5\u9002\u914d\u5177\u8eab\u667a\u80fd\u65b0\u7b97\u6cd5\uff0c\u652f\u6301\u4eba\u5f62\u673a\u5668\u4eba\u7b49\u5404\u79cd\u5f62\u6001\u3002<\/p>\n\n\n\n<p>\u4e8c\u8005\u53ef\u80fd\u7684\u7ed3\u5408\uff0c\u4e00\u65b9\u9762\u5728\u8bad\u7ec3\u7aef\u5e26\u6765\u66f4\u9ad8\u7684\u80fd\u6548\u4e0e\u901f\u5ea6\u4f18\u5316\uff0c\u53e6\u4e00\u65b9\u9762\u5728\u8fb9\u7f18\u3001\u63a8\u7406\u573a\u666f\u5145\u5206\u5229\u7528\u9ad8\u6027\u80fd\u4f4e\u529f\u8017\u7684\u8ba1\u7b97\u80fd\u529b\uff0c\u6700\u7ec8\u4ece\u8bad\u7ec3\u5230\u90e8\u7f72\u5f62\u6210\u9ad8\u6548\u7684\u5b8c\u6574\u95ed\u73af\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"246\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-467-1024x246.jpg\" alt=\"\" class=\"wp-image-1255\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-467-1024x246.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-467-300x72.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-467-768x184.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-467.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u4e0d\u8fc7\u4e5f\u6709\u4eba\u4e0d\u4e70\u8d26\uff0c\u9488\u5bf9\u82f1\u4f1f\u8fbe\u58f0\u79f0\u7684\u66f4\u73af\u4fdd\uff08greener\uff09\uff0c\u4ed6\u8ba4\u4e3a\uff0c\u867d\u7136\u65b0\u7684\u6570\u636e\u683c\u5f0f\u5e26\u6765\u4e86\u79cd\u79cd\u4f18\u5316\uff0c\u4f46\u5e76\u4e0d\u4ee3\u8868 AI \u7684\u603b\u4f53\u7b97\u529b\u9700\u6c42\u548c\u80fd\u8017\u4f1a\u56e0\u6b64\u51cf\u5c11\uff0c\u4e5f\u65e0\u6cd5\u4ece\u6839\u672c\u4e0a\u6539\u53d8 AI \u6301\u7eed\u6269\u5f20\u9020\u6210\u7684\u80fd\u6e90\u4e0e\u8d44\u6e90\u538b\u529b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"283\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-466-1024x283.jpg\" alt=\"\" class=\"wp-image-1254\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-466-1024x283.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-466-300x83.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-466-768x212.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-466.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>\u4ec0\u4e48\u662f 4 \u6bd4\u7279\u91cf\u5316\uff084-bit quantization\uff09\uff1f<\/strong><\/p>\n\n\n\n<p>4 \u6bd4\u7279\u91cf\u5316\u6307\u7684\u662f\u5c06\u6a21\u578b\u4e2d\u7684\u6743\u91cd\u548c\u6fc0\u6d3b\u503c\u7684\u7cbe\u5ea6\u964d\u4f4e\u5230\u4ec5\u4ec5 4 \u4f4d\u3002\u8fd9\u76f8\u6bd4\u5e38\u89c1\u7684 16 \u4f4d\u6216 32 \u4f4d\u6d6e\u70b9\u6570\u683c\u5f0f\uff0c\u662f\u4e00\u6b21\u5927\u5e45\u5ea6\u7684\u7cbe\u5ea6\u538b\u7f29\u3002<\/p>\n\n\n\n<p>\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u4f7f\u7528 4 \u6bd4\u7279\u91cf\u5316\u975e\u5e38\u5177\u6709\u6311\u6218\u6027\u3002\u56e0\u4e3a\u9700\u8981\u5728\u4fdd\u6301\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347\u7684\u540c\u65f6\uff0c\u8c28\u614e\u5730\u5904\u7406\u68af\u5ea6\u548c\u53c2\u6570\u66f4\u65b0\uff0c\u4ee5\u786e\u4fdd\u6a21\u578b\u7cbe\u5ea6\u4e0d\u4f1a\u4e22\u5931\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u8fbe\u5230\u8fd9\u4e2a\u76ee\u6807\uff0c\u82f1\u4f1f\u8fbe\u5fc5\u987b\u4f7f\u7528\u4e13\u95e8\u7684\u6280\u672f\u548c\u65b9\u6cd5\uff0c\u628a\u539f\u672c\u9ad8\u7cbe\u5ea6\u7684\u5f20\u91cf\uff08tensor\uff09\u6620\u5c04\u5230\u66f4\u5c0f\u7684\u91cf\u5316\u503c\u96c6\u5408\u4e2d\uff0c\u540c\u65f6\u4ecd\u7136\u7ef4\u6301\u6a21\u578b\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<p><strong>\u66f4\u5c11\u7684\u6bd4\u7279\u5982\u4f55\u91ca\u653e AI \u5de5\u5382\u7684\u66f4\u5927\u6f5c\u80fd<\/strong><\/p>\n\n\n\n<p>\u8fd1\u4e9b\u5e74\u6765\uff0cAI \u7684\u5de5\u4f5c\u8d1f\u8f7d\u5448\u73b0\u7206\u70b8\u5f0f\u589e\u957f \u2014\u2014 \u4e0d\u4ec5\u4ec5\u662f\u5728\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM, Large Language Model\uff09\u7684\u63a8\u7406\u90e8\u7f72\u4e2d\uff0c\u8fd8\u5305\u62ec\u57fa\u7840\u6a21\u578b\uff08foundation model\uff09\u5728\u9884\u8bad\u7ec3\u548c\u540e\u8bad\u7ec3\u9636\u6bb5\u7684\u89c4\u6a21\u6269\u5f20\u3002<\/p>\n\n\n\n<p>\u968f\u7740\u8d8a\u6765\u8d8a\u591a\u673a\u6784\u6269\u5c55\u8ba1\u7b97\u57fa\u7840\u8bbe\u65bd\uff0c\u7528\u6765\u8bad\u7ec3\u548c\u90e8\u7f72\u62e5\u6709\u6570\u5341\u4ebf\u53c2\u6570\u7684\u6a21\u578b\uff0c\u4e00\u4e2a\u6838\u5fc3\u6307\u6807\u9010\u6e10\u51f8\u663e\uff1a<strong>AI \u5de5\u5382\u80fd\u7ef4\u6301\u591a\u9ad8\u7684 token \u541e\u5410\u91cf\uff0c\u4ece\u800c\u89e3\u9501\u4e0b\u4e00\u9636\u6bb5\u7684\u6a21\u578b\u80fd\u529b\u3002<\/strong><\/p>\n\n\n\n<p>\u5728\u63a8\u7406\uff08inference\uff09\u73af\u8282\uff0c\u7cbe\u5ea6\u683c\u5f0f\u5df2\u7ecf\u7ecf\u5386\u4e86\u591a\u6b21\u9769\u65b0\uff1a\u4ece\u6700\u521d\u7684 FP32\uff0832 \u4f4d\u6d6e\u70b9\u6570\uff09\u5230 FP16\uff0c\u518d\u5230 FP8\uff0c\u6700\u8fd1\u751a\u81f3\u53d1\u5c55\u5230 NVIDIA \u53d1\u5e03\u7684 NVFP4\uff0c\u7528\u4e8e AI \u63a8\u7406\u3002\u5b9e\u8df5\u8868\u660e\uff0c\u50cf\u540e\u8bad\u7ec3\u91cf\u5316\uff08PTQ\uff09\u8fd9\u6837\u7684\u65b9\u6cd5\uff0c\u5df2\u7ecf\u80fd\u591f\u501f\u52a9 NVFP4 \u663e\u8457\u63d0\u5347\u63a8\u7406\u541e\u5410\u91cf\uff0c\u540c\u65f6\u4fdd\u6301\u51c6\u786e\u6027\u3002<\/p>\n\n\n\n<p>\u7136\u800c\uff0c\u5728\u66f4\u4e0a\u6e38\u7684\u9884\u8bad\u7ec3\u9636\u6bb5\uff0c\u6311\u6218\u4f9d\u7136\u5b58\u5728 \u2014\u2014 \u76ee\u524d\u5927\u591a\u6570\u57fa\u7840\u6a21\u578b\u4ecd\u4f9d\u8d56\u4e8e BF16 \u6216 FP8 \u6765\u7ef4\u6301\u7a33\u5b9a\u6027\u548c\u6536\u655b\u6027\u3002<\/p>\n\n\n\n<p>\u9884\u8bad\u7ec3\u6070\u6070\u662f AI \u5de5\u5382\u6d88\u8017\u6700\u591a\u8ba1\u7b97\u529b\u3001\u80fd\u8017\u548c\u65f6\u95f4\u7684\u73af\u8282\u3002\u7b97\u529b\u9884\u7b97\u6709\u9650\uff0cGPU \u65f6\u949f\u5468\u671f\u7a00\u7f3a\uff0c\u5f00\u53d1\u8005\u5fc5\u987b\u7cbe\u6253\u7ec6\u7b97 \u2014\u2014 \u4ece\u6bcf\u4e00\u4e2a\u6bd4\u7279\u3001\u6bcf\u4e00\u4e2a token\uff0c\u5230\u6bcf\u4e00\u4e2a\u8bad\u7ec3\u5468\u671f\u90fd\u8981\u8ba1\u7b97\u5728\u5185\u3002\u541e\u5410\u91cf\u5728\u8fd9\u91cc\u4e0d\u53ea\u662f\u4e00\u4e2a\u62bd\u8c61\u6307\u6807\uff0c\u5b83\u76f4\u63a5\u51b3\u5b9a\u4e86\uff1a\u80fd\u591f\u8bad\u7ec3\u591a\u5927\u89c4\u6a21\u7684\u6a21\u578b\uff0c\u53ef\u4ee5\u8fd0\u884c\u591a\u5c11\u5b9e\u9a8c\uff0c\u53c8\u80fd\u591a\u5feb\u8fce\u6765\u65b0\u7684\u7a81\u7834\u3002<\/p>\n\n\n\n<p>\u8fd9\u5c31\u662f<strong>&nbsp;4 \u4f4d\u7cbe\u5ea6\u771f\u6b63\u5177\u5907\u98a0\u8986\u6027\u610f\u4e49\u7684\u5730\u65b9\u3002<\/strong><\/p>\n\n\n\n<p>\u901a\u8fc7\u51cf\u5c11\u5185\u5b58\u9700\u6c42\u3001\u63d0\u5347\u7b97\u672f\u8fd0\u7b97\u541e\u5410\u91cf\u3001\u4f18\u5316\u901a\u4fe1\u6548\u7387\uff0c4 \u6bd4\u7279\u9884\u8bad\u7ec3\u80fd\u591f\u8ba9 AI \u5de5\u5382\u5728\u76f8\u540c\u7684\u786c\u4ef6\u6761\u4ef6\u4e0b\u5904\u7406\u66f4\u591a\u7684 token\u3002\u53ea\u8981\u914d\u5408\u5408\u9002\u7684\u91cf\u5316\u65b9\u6cd5\uff0c\u5b83\u7684\u7cbe\u5ea6\u8868\u73b0\u53ef\u4ee5\u4e0e FP8 \u6216 BF16 \u76f8\u5f53\uff0c\u540c\u65f6\u8fd8\u80fd\u663e\u8457\u63d0\u5347\u541e\u5410\u91cf\u3002<\/p>\n\n\n\n<p>\u8fd9\u610f\u5473\u7740\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6a21\u578b\u6536\u655b\u901f\u5ea6\u66f4\u5feb\uff1b<\/li>\n\n\n\n<li>\u5355\u4f4d\u7b97\u529b\u80fd\u8fd0\u884c\u66f4\u591a\u5b9e\u9a8c\uff1b<\/li>\n\n\n\n<li>\u53ef\u4ee5\u8bad\u7ec3\u51fa\u524d\u6240\u672a\u6709\u89c4\u6a21\u7684\u524d\u6cbf\u6a21\u578b\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6362\u53e5\u8bdd\u8bf4\uff0c<strong>\u66f4\u5c11\u7684\u6bd4\u7279\u4e0d\u4ec5\u4ec5\u662f\u8282\u7701\u6210\u672c\uff0c\u5b83\u8fd8\u62d3\u5c55\u4e86 AI \u5de5\u5382\u7684\u80fd\u529b\u8fb9\u754c\u3002<\/strong><\/p>\n\n\n\n<p><strong>NVFP4 \u9884\u8bad\u7ec3\u91cf\u5316\u65b9\u6848<\/strong><\/p>\n\n\n\n<p>\u4e3a\u4e86\u5b9e\u73b0 4 \u4f4d\u7cbe\u5ea6\u7684\u9884\u8bad\u7ec3\uff0c\u82f1\u4f1f\u8fbe\u5f00\u53d1\u4e86\u4e00\u5957\u4e13\u95e8\u7684 NVFP4 \u9884\u8bad\u7ec3\u65b9\u6848\uff0c\u89e3\u51b3\u4e86\u5927\u89c4\u6a21\u8bad\u7ec3\u4e2d\u52a8\u6001\u8303\u56f4\u3001\u68af\u5ea6\u6ce2\u52a8\u4ee5\u53ca\u6570\u503c\u7a33\u5b9a\u6027\u7684\u6838\u5fc3\u6311\u6218\u3002<\/p>\n\n\n\n<p>Blackwell \u662f NVIDIA \u9996\u4e2a\u539f\u751f\u652f\u6301 FP4 \u683c\u5f0f\u7684\u67b6\u6784\u3002GB200 \u548c GB300 \u4e0a\u5de8\u5927\u7684 FP4 FLOPs \u541e\u5410\u91cf\uff0c\u901a\u8fc7\u52a0\u901f\u4f4e\u7cbe\u5ea6\u77e9\u9635\u8fd0\u7b97\uff0c\u540c\u65f6\u4fdd\u6301\u5927\u6a21\u578b\u6536\u655b\u6240\u9700\u7684\u89c4\u6a21\u548c\u5e76\u884c\u6027\uff0c\u4ece\u800c\u5b9e\u73b0\u9ad8\u6548\u7684 4 \u6bd4\u7279\u8bad\u7ec3 \u2014\u2014 \u4f7f\u5176\u6210\u4e3a\u4e0b\u4e00\u4ee3\u57fa\u4e8e FP4 \u7684 AI \u5de5\u5382\u8fdb\u884c\u9884\u8bad\u7ec3\u7684\u7406\u60f3\u9009\u62e9\u3002<\/p>\n\n\n\n<p>\u4e0b\u56fe 1 \u663e\u793a\u4e86 Blackwell Ultra \u7684 GEMM \u6027\u80fd\u6d4b\u91cf\u7ed3\u679c\uff0c\u76f8\u6bd4 Hopper \u4ee3\u5b9e\u73b0\u4e86 7 \u500d\u52a0\u901f\u3002\u73b0\u4ee3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5728<strong>\u672c\u8d28\u4e0a\u4f9d\u8d56\u77e9\u9635\u4e58\u6cd5<\/strong>\uff0c\u5c24\u5176\u662f\u5728\u5176\u5168\u8fde\u63a5\u5c42\u6216\u7ebf\u6027\u5c42\u4e2d\uff0c\u77e9\u9635\u4e58\u6cd5\u662f\u6838\u5fc3\u8ba1\u7b97\u5143\u7d20\u3002\u56e0\u6b64\uff0c\u8fd9\u4e9b\u8fd0\u7b97\u7684\u6548\u7387\u81f3\u5173\u91cd\u8981\u3002 &nbsp;<\/p>\n\n\n\n<p>FP4 \u7cbe\u5ea6\u80fd\u591f\u66f4\u5feb\u3001\u66f4\u9ad8\u6548\u5730\u6267\u884c\u8fd9\u4e9b\u8fd0\u7b97\uff0c\u6240\u89c2\u5bdf\u5230\u7684 GEMM \u52a0\u901f\u610f\u5473\u7740\u6574\u4e2a\u9884\u8bad\u7ec3\u8fc7\u7a0b\u90fd\u663e\u8457\u52a0\u5feb\uff0c\u4ece\u800c\u7f29\u77ed\u8bad\u7ec3\u65f6\u95f4\uff0c\u5e76\u652f\u6301\u66f4\u5927\u89c4\u6a21\u6a21\u578b\u7684\u5feb\u901f\u5f00\u53d1\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"885\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-468-1024x885.jpg\" alt=\"\" class=\"wp-image-1256\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-468-1024x885.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-468-300x259.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-468-768x664.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-468.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>&nbsp; &nbsp; &nbsp; \u56fe 1\uff1a\u6d4b\u5f97\u7684 GEMM \u6027\u80fd\u663e\u793a\uff0cGB300 \u76f8\u6bd4 Hopper \u5b9e\u73b0\u4e86 7 \u500d\u52a0\u901f\uff0c\u901a\u8fc7\u66f4\u5feb\u7684 FP4 \u4f18\u5316\u77e9\u9635\u4e58\u6cd5\u52a0\u901f\u4e86\u6838\u5fc3 LLM \u8bad\u7ec3\u64cd\u4f5c\u3002<\/sup><\/p>\n\n\n\n<p>\u4e3a\u4e86\u5b9e\u73b0\u9ad8\u6548\u7684\u4f4e\u7cbe\u5ea6\u8bad\u7ec3\uff0cNVIDIA \u7684 NVFP4 \u9884\u8bad\u7ec3\u65b9\u6848\u91c7\u7528\u4e86\u591a\u9879\u5173\u952e\u6280\u672f\uff0c\u8fd9\u4e9b\u6280\u672f\u662f\u57fa\u4e8e\u6027\u80fd\u548c\u7cbe\u5ea6\u7cbe\u5fc3\u9009\u62e9\u7684\uff0c\u5305\u62ec\uff1a<\/p>\n\n\n\n<p><strong>1. \u5229\u7528 NVFP4 \u7684\u5fae\u5757\u7f29\u653e\u589e\u5f3a\u6570\u503c\u8868\u793a<\/strong><\/p>\n\n\n\n<p>Blackwell \u5f15\u5165\u4e86\u5bf9 NVFP4 \u7684\u539f\u751f Tensor Core \u652f\u6301\u3002NVFP4 \u662f\u4e00\u79cd 4 \u4f4d\u6570\u503c\u683c\u5f0f\uff0c\u53ef\u7528\u4e8e\u6743\u91cd\u548c\u6fc0\u6d3b\u503c\uff0c\u91c7\u7528\u5fae\u5757\u7f29\u653e\u6280\u672f \u2014\u2014 \u6bcf 16 \u4e2a 4 \u4f4d\u5143\u7d20\u5171\u4eab\u4e00\u4e2a\u516c\u5171\u7f29\u653e\u56e0\u5b50\u3002\u76f8\u6bd4 MXFP4 \u5c06\u5757\u5927\u5c0f\u8bbe\u4e3a 32 \u5143\u7d20\uff0cNVFP4 \u5c06\u5757\u5927\u5c0f\u7f29\u5c0f\u81f3 16 \u5143\u7d20\uff0c\u4ece\u800c\u51cf\u5c11\u5f02\u5e38\u503c\u7684\u5f71\u54cd\uff0c\u5b9e\u73b0\u66f4\u7cbe\u786e\u7684\u7f29\u653e\u3002\u66f4\u7ec6\u7c92\u5ea6\u7684\u7f29\u653e\u964d\u4f4e\u4e86\u91cf\u5316\u8bef\u5dee\uff0c\u63d0\u5347\u4e86\u6a21\u578b\u6574\u4f53\u7cbe\u5ea6\u3002<\/p>\n\n\n\n<p><strong>2. \u4f7f\u7528 E4M3 \u7f29\u653e\u56e0\u5b50\u7684 NVFP4 \u9ad8\u7cbe\u5ea6\u5757\u7f16\u7801<\/strong><\/p>\n\n\n\n<p>\u7f29\u653e\u56e0\u5b50\u7cbe\u5ea6\u5728\u91cf\u5316\u8d28\u91cf\u548c\u7cbe\u5ea6\u4e2d\u81f3\u5173\u91cd\u8981\u3002\u4e0d\u540c\u4e8e\u4ec5\u9650\u4e8e 2 \u7684\u5e42\uff08E8M0\uff09\u4e14\u6613\u4ea7\u751f\u9ad8\u820d\u5165\u8bef\u5dee\u7684 MXFP4\uff0cNVFP4 \u4f7f\u7528\u5e26\u989d\u5916\u5c3e\u6570\u4f4d\u7684\u9ad8\u7cbe\u5ea6 E4M3 \u7f29\u653e\u56e0\u5b50\u3002\u8fd9\u5141\u8bb8\u66f4\u7ec6\u7c92\u5ea6\u7684\u7f29\u653e\uff0c\u66f4\u6709\u6548\u5229\u7528\u6709\u9650\u7684\u91cf\u5316\u533a\u95f4\uff0c\u5e76\u5728\u5757\u5185\u66f4\u51c6\u786e\u5730\u8868\u793a\u6570\u503c\u3002<\/p>\n\n\n\n<p><strong>3. \u91cd\u5851\u5f20\u91cf\u5206\u5e03\u4ee5\u9002\u5e94\u4f4e\u7cbe\u5ea6\u683c\u5f0f<\/strong><\/p>\n\n\n\n<p>LLM \u9884\u8bad\u7ec3\u671f\u95f4\u7684\u68af\u5ea6\u548c\u6fc0\u6d3b\u503c\u901a\u5e38\u5b58\u5728\u5927\u5e45\u5f02\u5e38\u503c\uff0c\u8fd9\u4f1a\u5f71\u54cd\u4f4e\u7cbe\u5ea6\u91cf\u5316\u3002\u5bf9 GEMM \u8f93\u5165\u5e94\u7528 Hadamard \u53d8\u6362\uff0c\u53ef\u5c06\u5176\u5206\u5e03\u91cd\u5851\u4e3a\u66f4\u63a5\u8fd1\u9ad8\u65af\u5206\u5e03\uff0c\u4ece\u800c\u5e73\u6ed1\u5f02\u5e38\u503c\uff0c\u4f7f\u5f20\u91cf\u66f4\u5bb9\u6613\u88ab\u7cbe\u786e\u8868\u793a\u3002\u8fd9\u4e9b\u53d8\u6362\u5bf9\u6a21\u578b\u7ed3\u6784\u662f\u900f\u660e\u7684\uff0c\u53ef\u5728\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u7684\u7ebf\u6027\u5c42\u4e2d\u5e94\u7528\u3002<\/p>\n\n\n\n<p><strong>4. \u4f7f\u7528\u91cf\u5316\u6280\u672f\u4fdd\u6301\u6570\u636e\u4e00\u81f4\u6027<\/strong><\/p>\n\n\n\n<p>\u4e3a\u4e86\u786e\u4fdd\u8bad\u7ec3\u7a33\u5b9a\u9ad8\u6548\uff0c\u82f1\u4f1f\u8fbe\u91c7\u7528\u4fdd\u6301\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u4e00\u81f4\u6027\u7684\u91cf\u5316\u65b9\u6cd5\u3002\u8bf8\u5982\u9009\u62e9\u6027\u4e8c\u7ef4\u5757\u91cf\u5316\u7b49\u6280\u672f\uff0c\u6709\u52a9\u4e8e\u5728\u6574\u4e2a\u8bad\u7ec3\u5468\u671f\u4e2d\u4fdd\u6301\u5f20\u91cf\u8868\u793a\u7684\u5bf9\u9f50\u3002\u8fd9\u79cd\u4e00\u81f4\u6027\u5bf9\u4e8e\u6700\u5c0f\u5316\u4fe1\u53f7\u5931\u771f\u3001\u6539\u5584\u6536\u655b\u884c\u4e3a\u3001\u589e\u5f3a\u6574\u4f53\u9c81\u68d2\u6027\u81f3\u5173\u91cd\u8981\uff0c\u5c24\u5176\u662f\u5728 NVFP4 \u7b49\u4f4e\u7cbe\u5ea6\u683c\u5f0f\u4e0b\u3002<\/p>\n\n\n\n<p><strong>5. \u901a\u8fc7\u968f\u673a\u820d\u5165\u51cf\u5c11\u504f\u5dee<\/strong><\/p>\n\n\n\n<p>\u4e0e\u4f20\u7edf\uff08\u786e\u5b9a\u6027\uff09\u820d\u5165\u603b\u662f\u5411\u6700\u63a5\u8fd1\u7684\u53ef\u8868\u793a\u503c\u820d\u5165\u4e0d\u540c\uff0c\u968f\u673a\u820d\u5165\u4f1a\u6839\u636e\u6570\u503c\u5728\u4e24\u4e2a\u53ef\u8868\u793a\u503c\u4e4b\u95f4\u7684\u4f4d\u7f6e\uff0c\u6309\u6982\u7387\u5411\u4e0a\u6216\u5411\u4e0b\u820d\u5165\u3002\u8fd9\u4e00\u6b65\u9aa4\u5bf9\u4e8e\u51cf\u5c11\u820d\u5165\u504f\u5dee\u3001\u4fdd\u6301\u8bad\u7ec3\u671f\u95f4\u68af\u5ea6\u6d41\u52a8\u4ee5\u53ca\u6700\u7ec8\u63d0\u9ad8\u6a21\u578b\u7cbe\u5ea6\u81f3\u5173\u91cd\u8981\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"710\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-471-1024x710.jpg\" alt=\"\" class=\"wp-image-1259\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-471-1024x710.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-471-300x208.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-471-768x532.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-471.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>&nbsp; &nbsp; &nbsp; \u56fe 2\uff1a\u82f1\u4f1f\u8fbe\u7684 NVFP4 \u9884\u8bad\u7ec3\u6280\u672f\uff0c\u7528\u4ee5\u5b9e\u73b0\u9ad8\u6548\u4f4e\u7cbe\u5ea6\u8bad\u7ec3\u3002<\/sup><\/p>\n\n\n\n<p><strong>\u4e07\u4ebf\u7ea7 Token \u89c4\u6a21\u4e0b\u7684\u7cbe\u5ea6\u4e0e\u7a33\u5b9a\u6027<\/strong><\/p>\n\n\n\n<p>\u8981\u8ba9\u4f4e\u7cbe\u5ea6\u683c\u5f0f\u5728\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u4e2d\u5b9e\u7528\uff0c\u5fc5\u987b\u540c\u65f6\u4fdd\u8bc1\u6a21\u578b\u7cbe\u5ea6\u548c\u6536\u655b\u7a33\u5b9a\u6027\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u8bc4\u4f30 4 \u4f4d\u7cbe\u5ea6\u5728\u5927\u89c4\u6a21\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u53ef\u884c\u6027\uff0c\u82f1\u4f1f\u8fbe\u5728\u4e00\u4e2a 120 \u4ebf\u53c2\u6570\u7684\u6df7\u5408 Mamba-Transformer \u67b6\u6784\u6a21\u578b\uff0812B Hybrid Mamba-Transformer\uff09\u4e0a\u8fdb\u884c\u4e86 FP8 \u548c NVFP4 \u7684\u5b9e\u9a8c\u3002<\/p>\n\n\n\n<p>\u8be5\u6a21\u578b\u7c7b\u4f3c\u4e8e NVIDIA Nemotron Nano 2\uff0c\u5b83\u5728\u5305\u542b 10 \u4e07\u4ebf\u4e2a token \u7684\u8d85\u5927\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u91c7\u7528\u5206\u9636\u6bb5\u6570\u636e\u6df7\u5408\u7b56\u7565\uff1a\u5728\u8bad\u7ec3\u7684 70% \u9636\u6bb5\u5207\u6362\u5230\u4e0d\u540c\u7684\u6570\u636e\u96c6\u6df7\u5408\uff0c\u5728\u9884\u8bad\u7ec3\u7684 90% \u9636\u6bb5\u8fdb\u884c\u7b2c\u4e09\u9636\u6bb5\u6570\u636e\u5207\u6362\u3002<\/p>\n\n\n\n<p>\u8be5 12B Hybrid Mamba-Transformer \u6a21\u578b\u7684\u4e00\u4e2a\u7248\u672c\u6700\u521d\u4f7f\u7528 8 \u7cbe\u5ea6\uff08FP8\uff09\u8fdb\u884c\u8bad\u7ec3\u3002\u4e4b\u524d\u7684\u7814\u7a76\u8868\u660e\uff0cFP8 \u7684\u7cbe\u5ea6\u4e0e 16 \u4f4d\u7cbe\u5ea6\u975e\u5e38\u63a5\u8fd1\uff0c\u56e0\u6b64 FP8 \u88ab\u4f5c\u4e3a\u82f1\u4f1f\u8fbe\u7684\u57fa\u7ebf\u8fdb\u884c\u5bf9\u6bd4\u3002<\/p>\n\n\n\n<p>\u968f\u540e\uff0c\u82f1\u4f1f\u8fbe\u6210\u529f\u5730\u4ece\u96f6\u5f00\u59cb\u4f7f\u7528 NVFP4 \u8bad\u7ec3\u540c\u6837\u7684 12B \u6a21\u578b\uff0c\u8bc1\u660e<strong>\u8fd9\u79cd\u65b0\u7684\u4f4e\u7cbe\u5ea6\u683c\u5f0f\u53ef\u4ee5\u652f\u6301\u4e07\u4ebf\u7ea7 Token \u89c4\u6a21\u7684\u5b8c\u6574\u9884\u8bad\u7ec3<\/strong>\u3002\u5e76\u4e14\uff0cNVFP4 \u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u8868\u73b0\u51fa\u7a33\u5b9a\u7684\u6536\u655b\u6027\uff0c\u6ca1\u6709\u901a\u5e38\u56f0\u6270\u8d85\u4f4e\u7cbe\u5ea6\u8bad\u7ec3\u7684\u4e0d\u7a33\u5b9a\u6027\u6216\u53d1\u6563\u95ee\u9898\u3002<\/p>\n\n\n\n<p>\u4e0b\u56fe 3 \u663e\u793a\uff0cNVFP4 \u7684\u9a8c\u8bc1\u635f\u5931\u66f2\u7ebf\u5728\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e0e\u9ad8\u7cbe\u5ea6\u57fa\u7ebf\uff08\u5373 FP8\uff09\u7684\u635f\u5931\u66f2\u7ebf\u9ad8\u5ea6\u4e00\u81f4\u3002\u4e0a\u8ff0\u91cf\u5316\u6280\u672f\u786e\u4fdd\u5373\u4f7f\u5728\u5927\u5e45\u964d\u4f4e\u4f4d\u5bbd\u7684\u60c5\u51b5\u4e0b\uff0c4 \u6bd4\u7279\u9884\u8bad\u7ec3\u7684\u52a8\u6001\u8868\u73b0\u4ecd\u4e0e\u9ad8\u7cbe\u5ea6\u8bad\u7ec3\u975e\u5e38\u63a5\u8fd1\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"745\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-473-1024x745.jpg\" alt=\"\" class=\"wp-image-1261\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-473-1024x745.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-473-300x218.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-473-768x559.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-473.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>\u56fe 3\uff1a\u5728\u5bf9 120 \u4ebf\u53c2\u6570\u7684 Hybrid Mamba-Transformer \u6a21\u578b\u8fdb\u884c\u9884\u8bad\u7ec3\u65f6\uff0c\u5bf9\u6bd4\u4f7f\u7528 FP8 \u4e0e NVFP4 \u7cbe\u5ea6\u5728 10 \u4e07\u4ebf tokens \u4e0b\u7684\u9a8c\u8bc1\u635f\u5931\u7ed3\u679c\u663e\u793a\uff0cNVFP4 \u7684\u635f\u5931\u66f2\u7ebf\u5728\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e0e FP8\uff08\u57fa\u7ebf\uff09\u7684\u66f2\u7ebf\u9ad8\u5ea6\u543b\u5408\u3002<\/sup><\/p>\n\n\n\n<p>\u968f\u540e\uff0c\u82f1\u4f1f\u8fbe\u4f7f\u7528 NVFP4 \u9884\u8bad\u7ec3 120 \u4ebf\u53c2\u6570\u7684 Hybrid Mamba-Transformer \u6a21\u578b\uff0c\u5e76\u4e0e\u66f4\u9ad8\u7cbe\u5ea6\u7684 FP8 \u57fa\u7ebf\u5728\u591a\u4e2a\u4e0b\u6e38\u4efb\u52a1\u4e0e\u667a\u80fd\u9886\u57df\u8fdb\u884c\u4e86\u5bf9\u6bd4\u3002<\/p>\n\n\n\n<p>\u5982\u4e0b\u56fe 4 \u6240\u793a\uff0c\u5728\u6240\u6709\u9886\u57df\u4e2d\uff0cNVFP4 \u7684\u51c6\u786e\u7387\u8868\u73b0\u5747\u4e0e FP8 \u76f8\u5f53\uff0c\u751a\u81f3\u5728\u4ee3\u7801\u9886\u57df\u5b9e\u73b0\u4e86\u53cd\u8d85\uff0c\u5c55\u73b0\u4e86\u5176\u6709\u6548\u6027\u3002\u8be5\u7ed3\u679c\u8fdb\u4e00\u6b65\u5f3a\u5316\u4e86\u6700\u521d\u7684\u5047\u8bbe\uff1a<strong>\u5373\u4f7f\u5728\u4e07\u4ebf token \u89c4\u6a21\u4e0b\uff0cNVFP4 \u4f9d\u7136\u662f\u5927\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u7684\u7a33\u5065\u9009\u62e9\uff0c\u9a8c\u8bc1\u4e86\u5176\u5728\u9ad8\u6548\u5927\u89c4\u6a21\u524d\u6cbf\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u6f5c\u529b\u3002<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"623\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-472-1024x623.jpg\" alt=\"\" class=\"wp-image-1260\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-472-1024x623.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-472-300x182.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-472-768x467.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-472.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>\u56fe 4\uff1a\u5206\u522b\u4f7f\u7528 FP8 \u7cbe\u5ea6\uff08\u57fa\u7ebf\uff09\u548c NVFP4 \u7cbe\u5ea6\uff0c\u5bf9 120 \u4ebf\u53c2\u6570\u7684 Hybrid Mamba-Transformer \u6a21\u578b\u8fdb\u884c\u9884\u8bad\u7ec3\uff0c\u6b64\u65f6\u7684\u4e0b\u6e38\u4efb\u52a1\u51c6\u786e\u7387\u5bf9\u6bd4\u3002<\/sup><\/p>\n\n\n\n<p><strong>\u806a\u660e\u8bad\u7ec3\uff0c\u800c\u4e0d\u662f\u4e00\u5473\u52a0\u7801<\/strong><\/p>\n\n\n\n<p>\u6839\u636e\u82f1\u4f1f\u8fbe\u7684\u8bf4\u6cd5\uff0cNVFP4 \u683c\u5f0f\u6b63\u5728\u91cd\u65b0\u5b9a\u4e49 AI \u8bad\u7ec3\u7684\u683c\u5c40\uff0c\u5e76\u53ef\u4ee5\u4e3a\u5b9e\u73b0\u901f\u5ea6\u3001\u6548\u7387\u548c\u6709\u76ee\u7684\u521b\u65b0\u8bbe\u7acb\u65b0\u7684\u6807\u6746\u3002\u901a\u8fc7\u5b9e\u73b0 4 \u6bd4\u7279\u9884\u8bad\u7ec3\uff0cNVFP4 \u8ba9 AI \u5de5\u5382\u66f4\u5feb\u3001\u66f4\u53ef\u6301\u7eed\u5730\u6269\u5c55\uff0c\u4e3a\u5168\u65b0\u7684\u751f\u6210\u5f0f AI \u65f6\u4ee3\u6253\u4e0b\u57fa\u7840\u3002<\/p>\n\n\n\n<p>\u53e6\u5916\uff0c\u4f5c\u4e3a\u4e00\u79cd\u52a8\u6001\u4e14\u4e0d\u65ad\u6f14\u8fdb\u7684\u6280\u672f\uff0cNVFP4 \u5c06\u6301\u7eed\u4e3a\u524d\u6cbf\u6a21\u578b\u56e2\u961f\u521b\u9020\u65b0\u7684\u673a\u9047\uff0c\u63a8\u52a8\u8282\u80fd\u9ad8\u6548\u548c\u9ad8\u6027\u80fd\u7684 AI \u53d1\u5c55\u3002\u51ed\u501f\u8ba1\u7b97\u6548\u7387\u7684\u7a81\u7834\uff0c4 \u6bd4\u7279\u9884\u8bad\u7ec3\u5c06\u8d4b\u80fd\u66f4\u5148\u8fdb\u7684\u67b6\u6784\u3001\u66f4\u5927\u89c4\u6a21\u7684\u8bad\u7ec3\u548c token \u5904\u7406\uff0c\u4ece\u800c\u4e3a\u672a\u6765\u7684\u667a\u80fd\u7cfb\u7edf\u6ce8\u5165\u65b0\u7684\u52a8\u529b\u3002<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u524d\u4e9b\u5929\uff0cDeepSeek \u5728\u53d1\u5e03 DeepSeek V3.1 \u7684\u6587\u7ae0\u8bc4\u8bba\u533a\u4e2d\uff0c\u63d0\u53ca\u4e86&nbsp;&nbsp;\u7684 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":1262,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[],"class_list":["post-1252","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-7"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1252","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=1252"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1252\/revisions"}],"predecessor-version":[{"id":1263,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1252\/revisions\/1263"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/1262"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=1252"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=1252"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=1252"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}