{"id":2791,"date":"2024-02-28T17:23:16","date_gmt":"2024-02-28T09:23:16","guid":{"rendered":"http:\/\/blog.cloud500.net\/?p=2791"},"modified":"2024-02-28T17:23:16","modified_gmt":"2024-02-28T09:23:16","slug":"groq-%e7%9a%84%e6%8e%a8%e7%90%86%e7%bb%8f%e6%b5%8e%e5%ad%a6","status":"publish","type":"post","link":"http:\/\/blog.cloud500.ai\/?p=2791","title":{"rendered":"Groq \u7684\u63a8\u7406\u7ecf\u6d4e\u5b66"},"content":{"rendered":"\n<p>\u8fd1\u671f\uff0cAI \u786c\u4ef6\u521d\u521b\u516c\u53f8 Groq \u5f15\u8d77\u4e86\u5e7f\u6cdb\u5173\u6ce8\uff0c\u4ed6\u4eec\u5728\u63a8\u7406 API \u4e0a\u6240\u5c55\u793a\u7684\u9886\u5148\u5f00\u6e90\u6a21\u578b Mistral 8x7b \u7ed9\u4eba\u7559\u4e0b\u4e86\u6781\u4e3a\u6df1\u523b\u7684\u5370\u8c61\u3002\u76f8\u6bd4\u5176\u4ed6\u63a8\u7406\u670d\u52a1\uff0c\u4ed6\u4eec\u7684\u541e\u5410\u91cf\u591a 4 \u500d\uff0c\u4f46\u6536\u8d39\u5374\u4e0d\u5230 Mistral \u81ea\u8eab\u7684 1\/3\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"1080\" height=\"590\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-40.webp\" alt=\"\" class=\"wp-image-2792\"\/><\/figure><\/div>\n\n\n<p>Groq \u9488\u5bf9\u4e2a\u4f53\u5e8f\u5217\u5c55\u73b0\u51fa\u4e86\u60ca\u4eba\u7684\u6027\u80fd\u4f18\u52bf\uff0c\u8fd9\u5c06\u4f7f\u8bf8\u5982\u201c\u601d\u7ef4\u94fe\u201d\u7b49\u6280\u672f\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u66f4\u6613\u4e8e\u4f7f\u7528\u3002\u6b64\u5916\uff0c\u968f\u7740 AI \u7cfb\u7edf\u53d8\u5f97\u66f4\u52a0\u81ea\u4e3b\u5316\uff0cLLM \u7684\u8f93\u51fa\u901f\u5ea6\u5bf9\u4e8e agent \u6765\u8bf4\u53d8\u5f97\u5c24\u4e3a\u91cd\u8981\u3002\u540c\u6837\uff0c\u4ee3\u7801\u751f\u6210\u5668\uff08codegen\uff09\u4e5f\u9700\u8981\u5927\u5927\u964d\u4f4e token \u8f93\u51fa\u5ef6\u8fdf\u3002\u5b9e\u65f6 Sora \u98ce\u683c\u7684\u6a21\u578b\u53ef\u80fd\u4f1a\u6210\u4e3a\u4e00\u79cd\u4ee4\u4eba\u96be\u4ee5\u7f6e\u4fe1\u7684\u5a31\u4e50\u65b9\u5f0f\uff0c\u4f46\u5982\u679c\u5ef6\u8fdf\u8fc7\u9ad8\uff0c\u8fd9\u4e9b\u670d\u52a1\u5bf9\u7ec8\u7aef\u5e02\u573a\u7684\u5ba2\u6237\u6765\u8bf4\u53ef\u80fd\u6839\u672c\u4e0d\u53ef\u884c\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0cGroq \u7684\u786c\u4ef6\u548c\u63a8\u7406\u670d\u52a1\u88ab\u5927\u8086\u7092\u4f5c\uff0c\u88ab\u89c6\u4e3a\u5bf9 AI \u884c\u4e1a\u5177\u6709\u9769\u547d\u6027\u610f\u4e49\u3002\u867d\u7136\u5bf9\u4e8e\u67d0\u4e9b\u5e02\u573a\u548c\u5e94\u7528\u800c\u8a00\uff0cGroq \u65e0\u7591\u662f\u4e00\u4e2a\u53d8\u9769\u8005\uff0c\u4f46\u901f\u5ea6\u53ea\u662f\u201c\u7b49\u5f0f\u201d\u7684\u4e00\u90e8\u5206\u3002\u4f9b\u5e94\u94fe\u7684\u591a\u5143\u5316\u4e5f\u662f\u5bf9 Groq \u6709\u5229\u7684\u53e6\u4e00\u4e2a\u56e0\u7d20\uff1a\u4ed6\u4eec\u7684\u82af\u7247\u5b8c\u5168\u5728\u7f8e\u56fd\u8fdb\u884c\u5236\u9020\u548c\u5c01\u88c5\uff0c\u800c Nvidia\u3001Google \u548c AMD \u7b49\u516c\u53f8\u7684 AI \u82af\u7247\u5219\u9700\u8981\u4ece\u97e9\u56fd\u91c7\u8d2d\u5185\u5b58\uff0c\u5e76\u4f9d\u8d56\u53f0\u6e7e\u5148\u8fdb\u7684\u82af\u7247\u5c01\u88c5\u6280\u672f\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e9b\u5bf9\u4e8e Groq \u6765\u8bf4\u662f\u79ef\u6781\u7684\u56e0\u7d20\uff0c\u4f46\u8bc4\u4f30\u786c\u4ef6\u662f\u5426\u5177\u6709\u9769\u547d\u6027\u7684\u4e3b\u8981\u6307\u6807\u662f\u6027\u80fd\u4e0e\u603b\u62e5\u6709\u6210\u672c\u4e4b\u6bd4\u3002\u8fd9\u4e00\u70b9\uff0c\u8c37\u6b4c\u6700\u6709\u53d1\u8a00\u6743\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote\">\n<p>AI \u65f6\u4ee3\u5df2\u7ecf\u6765\u4e34\uff0c\u80fd\u591f\u7406\u89e3 AI \u9a71\u52a8\u8f6f\u4ef6\u7684\u6210\u672c\u7ed3\u6784\u4e0e\u4f20\u7edf\u8f6f\u4ef6\u6709\u7740\u663e\u8457\u7684\u504f\u5dee\u81f3\u5173\u91cd\u8981\u3002\u82af\u7247\u5fae\u67b6\u6784\u548c\u7cfb\u7edf\u67b6\u6784\u5bf9\u8fd9\u4e9b\u521b\u65b0\u578b\u8f6f\u4ef6\u7684\u5f00\u53d1\u548c\u53ef\u6269\u5c55\u6027\u8d77\u7740\u81f3\u5173\u91cd\u8981\u7684\u4f5c\u7528\u3002\u4e0e\u5f00\u53d1\u6210\u672c\u76f8\u5bf9\u8f83\u9ad8\u7684\u524d\u51e0\u4ee3\u8f6f\u4ef6\u76f8\u6bd4\uff0cAI \u8f6f\u4ef6\u7684\u90e8\u7f72\u5bf9\u8d44\u672c\u652f\u51fa\u3001\u8fd0\u8425\u652f\u51fa\u4ee5\u53ca\u968f\u4e4b\u800c\u6765\u7684\u6bdb\u5229\u6da6\u4ea7\u751f\u4e86\u66f4\u5927\u7684\u5f71\u54cd\uff0c\u56e0\u4e3a\u5b83\u4eec\u5bf9\u786c\u4ef6\u57fa\u7840\u8bbe\u65bd\u6709\u66f4\u9ad8\u7684\u8981\u6c42\u3002\u56e0\u6b64\uff0c\u8981\u60f3\u90e8\u7f72 AI \u8f6f\u4ef6\uff0c\u5fc5\u987b\u9ad8\u5ea6\u5173\u6ce8\u4f18\u5316 AI \u57fa\u7840\u8bbe\u65bd\u3002\u62e5\u6709\u57fa\u7840\u8bbe\u65bd\u4f18\u52bf\u7684\u4f01\u4e1a\u4e5f\u5c06\u5728\u90e8\u7f72\u548c\u6269\u5c55 AI \u5e94\u7528\u65b9\u9762\u5177\u6709\u7ade\u4e89\u4f18\u52bf\u3002<br>\u51fa\u81ea\uff1a\u300a\u8c37\u6b4c AI \u57fa\u7840\u8bbe\u65bd\u7684\u4f18\u52bf\uff1a\u7cfb\u7edf\u6bd4\u5fae\u67b6\u6784\u66f4\u91cd\u8981\u300b<\/p>\n<\/blockquote>\n\n\n\n<p>\u8c37\u6b4c\u5728\u57fa\u7840\u8bbe\u65bd\u65b9\u9762\u7684\u4f18\u52bf\u4f53\u73b0\u5728\u8c37\u6b4c Gemini 1.5 \u4e0e OpenAI GPT-4 Turbo \u76f8\u6bd4\uff0cGemini 1.5 \u7684\u670d\u52a1\u6210\u672c\u660e\u663e\u8f83\u4f4e\uff0c\u540c\u65f6\u5728\u8bb8\u591a\u4efb\u52a1\u4e2d\uff0c\u5c24\u5176\u662f\u957f\u5e8f\u5217\u4ee3\u7801\u4e2d\uff0c\u8c37\u6b4c Gemini 1.5 \u8868\u73b0\u8981\u66f4\u597d\u3002\u867d\u7136\u5728\u5355\u4e2a\u63a8\u7406\u7cfb\u7edf\u4e2d\u8c37\u6b4c\u4f7f\u7528\u7684\u82af\u7247\u6570\u91cf\u8981\u591a\u5f97\u591a\uff0c\u4f46\u4ece\u6027\u80fd\/\u603b\u62e5\u6709\u6210\u672c\u7684\u89d2\u5ea6\u6765\u770b\uff0c\u8c37\u6b4c\u7684\u4f18\u52bf\u66f4\u4e3a\u663e\u8457\u3002<\/p>\n\n\n\n<p>\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6027\u80fd\u4e0d\u4ec5\u4ec5\u662f\u6307\u5355\u4e2a\u7528\u6237\u6bcf\u79d2\u5904\u7406\u7684\u539f\u59cb token \u6570\uff0c\u5373\u5ef6\u8fdf\u4f18\u5316\u3002\u5728\u8bc4\u4f30\u603b\u4f53\u62e5\u6709\u6210\u672c\uff08TCO\uff09\u65f6\uff0c\u8fd8\u5fc5\u987b\u8003\u8651\u5230\u786c\u4ef6\u4e0a\u5e76\u53d1\u670d\u52a1\u7684\u7528\u6237\u6570\u91cf\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u6539\u8fdb\u8fb9\u7f18\u8ba1\u7b97\u7684\u786c\u4ef6\u4ee5\u6ee1\u8db3 LLM \u63a8\u7406\u9700\u6c42\u65f6\u4f1a\u5fc5\u987b\u6743\u8861\u7684\u4e3b\u8981\u539f\u56e0\u3002\u5927\u591a\u6570\u8fb9\u7f18\u8ba1\u7b97\u7cfb\u7edf\u90fd\u65e0\u6cd5\u5f25\u8865\u4e3a\u8fd0\u884c LLM \u800c\u589e\u52a0\u7684\u786c\u4ef6\u6210\u672c\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u8fb9\u7f18\u7cfb\u7edf\u65e0\u6cd5\u5c06\u6210\u672c\u5728\u5927\u91cf\u7528\u6237\u4e4b\u95f4\u8fdb\u884c\u644a\u9500\u3002\u800c\u5bf9\u4e8e\u4ee5\u6781\u5927 batch sizes \u4e3a\u4f17\u591a\u7528\u6237\u63d0\u4f9b\u670d\u52a1\u3001\u4f18\u5316 IE \u541e\u5410\u91cf\u548c\u6210\u672c\u6765\u8bf4\uff0cGPU \u624d\u662f\u6700\u4f73\u9009\u62e9\u3002<\/p>\n\n\n\n<p>\u6b63\u5982\u5728\u300a\u63a8\u7406\u7684\u7ade\u4e89\u300b\u4e2d\u6240\u8ba8\u8bba\u7684\uff0c\u8bb8\u591a\u516c\u53f8\u5728\u4f7f\u7528 Mixtral API \u63a8\u7406\u670d\u52a1\u65f6\u786e\u5b9e\u5904\u4e8e\u4e8f\u635f\u72b6\u6001\u3002\u4e00\u4e9b\u516c\u53f8\u8fd8\u5236\u5b9a\u4e86\u975e\u5e38\u4f4e\u7684\u8d39\u7387\u9650\u5236\uff0c\u4ee5\u9650\u5236\u4ed6\u4eec\u7684\u4e8f\u635f\u3002\u6211\u4eec\u5728\u62a5\u544a\u4e2d\u6df1\u5165\u8ba8\u8bba\u4e86\u91cf\u5316\u548c\u5176\u4ed6\u786c\u4ef6 GPU \u9009\u62e9\uff08\u5982 MI300X\uff09\uff0c\u4f46\u4e3b\u8981\u7ed3\u8bba\u662f\uff0c\u90a3\u4e9b\u63d0\u4f9b\u672a\u4fee\u6539\u6a21\u578b\uff08\u4f7f\u7528 FP16 \u7cbe\u5ea6\uff09\u670d\u52a1\u7684\u516c\u53f8\u81f3\u5c11\u8981\u5c06 batch size \u8bbe\u7f6e\u6210 64 \u4ee5\u4e0a\u624d\u80fd\u5b9e\u73b0\u76c8\u5229\u3002\u6211\u4eec\u8ba4\u4e3a\uff0cMistral\u3001Together \u548c Fireworks \u6b63\u5728\u4ee5\u4fdd\u672c\u6216\u5fae\u5229\u7684\u72b6\u6001\u63d0\u4f9b Mistral \u670d\u52a1\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"1080\" height=\"652\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-41.webp\" alt=\"\" class=\"wp-image-2793\"\/><\/figure><\/div>\n\n\n<p>\u4f46\u5176\u4ed6\u63d0\u4f9b Mixtral API \u7684\u516c\u53f8\u60c5\u51b5\u5219\u622a\u7136\u4e0d\u540c\u3002\u4ed6\u4eec\u8981\u4e48\u5728\u91cf\u5316\u95ee\u9898\u4e0a\u6492\u8c0e\uff0c\u8981\u4e48\u4f9d\u8d56 VC \u7684\u8d44\u91d1\u6765\u83b7\u53d6\u7528\u6237\u3002Groq \u5728\u5b9a\u4ef7\u65b9\u9762\u91c7\u53d6\u4e86\u5927\u80c6\u7684\u5c1d\u8bd5\uff0c\u4ee5 0.27 \u7f8e\u5143\u6bcf\u767e\u4e07 token \u7684\u8d85\u4f4e\u4ef7\u683c\u4e0e\u8fd9\u4e9b\u516c\u53f8\u7ade\u4e89\u3002<\/p>\n\n\n\n<p>\u5b83\u4eec\u7684\u5b9a\u4ef7\u662f\u5426\u4e5f\u50cf Together \u548c Fireworks \u4e00\u6837\u662f\u57fa\u4e8e\u6027\u80fd \/ TCO \u800c\u8ba1\u7b97\u7684\u5462\uff1f<\/p>\n\n\n\n<p>\u4ea6\u6216\u662f\u901a\u8fc7\u8865\u8d34\u6765\u8fdb\u884c\u7092\u4f5c\uff1f\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0cGroq \u7684\u4e0a\u4e00\u8f6e\u878d\u8d44\u662f\u5728 2021 \u5e74\uff0c\u53bb\u5e74\u7684 SAFE \u989d\u5ea6\u4e3a 5000 \u4e07\u7f8e\u5143\uff0c\u76ee\u524d\u4ed6\u4eec\u6b63\u5728\u8fdb\u884c\u65b0\u4e00\u8f6e\u7684\u878d\u8d44\u3002<\/p>\n\n\n\n<p>\u63a5\u4e0b\u6765\u672c\u6587\u4f1a\u5bf9 Groq \u7684\u82af\u7247\u3001\u7cfb\u7edf\u548c\u6210\u672c\u8fdb\u884c\u6df1\u5165\u7814\u7a76\uff0c\u5e76\u63a2\u8ba8\u4ed6\u4eec\u662f\u5982\u4f55\u5b9e\u73b0\u51fa\u8272\u6027\u80fd\u7684\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"1080\" height=\"352\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-42.webp\" alt=\"\" class=\"wp-image-2794\"\/><\/figure><\/div>\n\n\n<p>Groq \u7684\u82af\u7247\u91c7\u7528\u5b8c\u5168\u786e\u5b9a\u6027\u7684 VLIW \u67b6\u6784\uff0c\u6ca1\u6709\u7f13\u51b2\u533a\uff0c\u5728 Global Foundries \u7684 14 \u7eb3\u7c73\u5de5\u827a\u8282\u70b9\u4e0a\u8fbe\u5230\u4e86\u7ea6 725 mm2 \u7684\u82af\u7247\u5c3a\u5bf8\u3002\u8be5\u82af\u7247\u6ca1\u6709\u5916\u90e8\u5185\u5b58\uff0c\u5728\u5904\u7406\u8fc7\u7a0b\u4e2d\u5c06\u6240\u6709\u6743\u91cd\u3001KVCache \u548c\u6fc0\u6d3b\u7b49\u6570\u636e\u90fd\u4fdd\u7559\u5728\u82af\u7247\u5185\u90e8\u3002\u7531\u4e8e\u6bcf\u4e2a\u82af\u7247\u53ea\u6709 230MB \u7684 SRAM\uff0c\u56e0\u6b64\u5355\u4e2a\u82af\u7247\u5b9e\u9645\u4e0a\u65e0\u6cd5\u5bb9\u7eb3\u4efb\u4f55\u6709\u7528\u7684\u6a21\u578b\u3002\u56e0\u6b64\uff0c\u5b83\u4eec\u5fc5\u987b\u5229\u7528\u591a\u4e2a\u82af\u7247\u6765\u9002\u914d\u6a21\u578b\uff0c\u5e76\u5c06\u5b83\u4eec\u4ee5\u7f51\u7edc\u5316\u7684\u65b9\u5f0f\u8fde\u63a5\u8d77\u6765\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"1080\" height=\"704\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-43.webp\" alt=\"\" class=\"wp-image-2795\"\/><\/figure><\/div>\n\n\n<p>\u5c31 Mixtral \u6a21\u578b\u800c\u8a00\uff0cGroq \u9700\u8981\u8fde\u63a5 8 \u4e2a\u673a\u67b6\uff0c\u6bcf\u4e2a\u673a\u67b6\u6709 9 \u53f0\u670d\u52a1\u5668\uff0c\u6bcf\u53f0\u670d\u52a1\u5668\u542b\u6709 8 \u4e2a\u82af\u7247\u3002\u8fd9\u610f\u5473\u7740\u603b\u5171\u9700\u8981 576 \u4e2a\u82af\u7247\u624d\u80fd\u5efa\u7acb\u63a8\u7406\u5355\u5143\uff0c\u5e76\u4e3a Mixtral \u6a21\u578b\u63d0\u4f9b\u670d\u52a1\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cNvidia \u5355\u4e2a H100 \u5c31\u80fd\u6ee1\u8db3\u4f4e\u6279\u91cf\u6a21\u578b\u7684\u9700\u6c42\uff0c\u5e76\u4e14\u4e24\u4e2a\u82af\u7247\u5c31\u6709\u8db3\u591f\u7684\u5185\u5b58\u652f\u6301\u66f4\u5927\u7684 batch size\u3002<\/p>\n\n\n\n<p>\u7528\u4e8e\u5236\u9020 Groq \u82af\u7247\u7684\u6676\u5706\u6210\u672c\u53ef\u80fd\u4e0d\u5230\u6bcf\u7247 6,000 \u7f8e\u5143\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cNvidia \u7684 H100 \u82af\u7247\u5c3a\u5bf8\u4e3a 814 mm2\uff0c\u91c7\u7528\u53f0\u79ef\u7535\u7684 5 \u7eb3\u7c73\u5b9a\u5236\u53d8\u4f53 4N\uff0c\u8fd9\u4e9b\u6676\u5706\u7684\u6210\u672c\u63a5\u8fd1\u6bcf\u7247 16,000 \u7f8e\u5143\u3002\u7136\u800c\uff0c\u4e0e Nvidia \u7684\u67b6\u6784\u76f8\u6bd4\uff0cGroq \u7684\u67b6\u6784\u5728\u5b9e\u73b0\u9ad8\u826f\u7387\u65b9\u9762\u4f3c\u4e4e\u5b58\u5728\u4e00\u4e9b\u6311\u6218\uff0c\u800c Nvidia \u7684\u53c2\u6570\u826f\u7387\u975e\u5e38\u9ad8\uff0c\u4ed6\u4eec\u5728\u5927\u591a\u6570 H100 SKUs \u4e2d\u7981\u7528\u4e86\u7ea6 15% \u7684\u6676\u7247\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0cNvidia \u4ece SK Hynix \u8d2d\u4e70\u4e86\u6bcf\u9897 H100 \u82af\u7247\u6240\u9700\u7684 80GB \u7684 HBM\uff0c\u6bcf\u4e2a\u6210\u672c\u7ea6\u4e3a 1,150 \u7f8e\u5143\u3002Nvidia \u8fd8\u9700\u8981\u5411\u53f0\u79ef\u7535\u652f\u4ed8 CoWoS \u8d39\u7528\uff0c\u5e76\u627f\u62c5\u7531\u6b64\u5e26\u6765\u7684\u826f\u7387\u635f\u5931\uff0c\u800c Groq \u5219\u4e0d\u9700\u8981\u4efb\u4f55\u82af\u7247\u5916\u5185\u5b58\u3002\u56e0\u6b64\uff0cGroq \u82af\u7247\u7684\u539f\u6750\u6599\u6210\u672c\u8981\u4f4e\u5f97\u591a\u3002Groq \u8fd8\u662f\u4e00\u5bb6\u521d\u521b\u516c\u53f8\uff0c\u56e0\u6b64\u4ed6\u4eec\u7684\u82af\u7247\u4ea7\u91cf\u8981\u4f4e\u5f97\u591a\uff0c\u76f8\u5bf9\u56fa\u5b9a\u6210\u672c\u4e5f\u8981\u9ad8\u5f97\u591a\uff0c\u5176\u4e2d\u5305\u62ec\u5fc5\u987b\u5411Marvell\u652f\u4ed8\u9ad8\u989d\u5229\u6da6\u4ee5\u83b7\u53d6\u5b9a\u5236ASIC\u670d\u52a1\u3002<\/p>\n\n\n\n<p>\u4e0b\u8868\u5217\u51fa\u4e86\u4e09\u79cd\u90e8\u7f72\uff0c\u4e00\u79cd\u662f\u9488\u5bf9 Groq \u7684\u90e8\u7f72\uff0c\u91c7\u7528\u4ed6\u4eec\u5f53\u524d\u7684\u7ba1\u9053\u5e76\u884c\u6027\uff0cbatch size \u8bbe\u7f6e\u4e3a 3\u3002\u636e\u6211\u4eec\u4e86\u89e3\uff0c\u4ed6\u4eec\u8ba1\u5212\u5728\u4e0b\u5468\u5f00\u59cb\u5728\u751f\u4ea7\u4e2d\u5b9e\u65bd\u8fd9\u79cd\u90e8\u7f72\u3002\u53e6\u5916\u4e24\u79cd\u5206\u522b\u662f\u9488\u5bf9\u5ef6\u8fdf\u4f18\u5316\u7684 H100 \u63a8\u7406\u90e8\u7f72\uff0c\u4ee5\u53ca\u9488\u5bf9\u541e\u5410\u91cf\u4f18\u5316\u7684 H100 \u63a8\u7406\u90e8\u7f72\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"1080\" height=\"504\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-44.webp\" alt=\"\" class=\"wp-image-2796\"\/><\/figure><\/div>\n\n\n<p>\u4e0a\u8868\u5927\u5927\u7b80\u5316\u4e86\u7ecf\u6d4e\u6548\u76ca\uff08\u5ffd\u7565\u4e86\u5927\u91cf\u7cfb\u7edf\u7ea7\u6210\u672c\uff0c\u6211\u4eec\u7a0d\u540e\u5c06\u6df1\u5165\u8ba8\u8bba\uff0c\u540c\u65f6\u4e5f\u5ffd\u7565\u4e86 Nvidia \u7684\u5de8\u5927\u5229\u6da6\uff09\u3002\u8fd9\u91cc\u7684\u91cd\u70b9\u662f\u8981\u6307\u51fa\uff0c\u4e0e\u7ecf\u8fc7\u5ef6\u8fdf\u4f18\u5316\u7684 Nvidia \u7cfb\u7edf\u76f8\u6bd4\uff0cGroq \u5728\u82af\u7247\u67b6\u6784\u4e0a\u7684\u4f18\u52bf\u5728\u4e8e\u6bcf token \u8f93\u51fa\u6240\u4f7f\u7528\u7684\u7845\u6750\u6599\u6210\u672c\u3002<\/p>\n\n\n\n<p>8xA100 GPU \u53ef\u4ee5\u4e3a Mixtral \u63d0\u4f9b\u670d\u52a1\uff0c\u5e76\u5b9e\u73b0\u6bcf\u7528\u6237\u6bcf\u79d2\u7ea6220 \u4e2a token \u7684\u541e\u5410\u91cf\u3002\u800c 8xH100 \u5728\u4e0d\u8fdb\u884c\u63a8\u6d4b\u89e3\u7801\uff08speculative decoding\uff09\u7684\u60c5\u51b5\u4e0b\uff0c\u6bcf\u7528\u6237\u6bcf\u79d2\u53ef\u8fbe\u5230\u7ea6 280 \u4e2a token\u3002\u5f53\u4f7f\u7528\u63a8\u6d4b\u89e3\u7801\u65f6\uff0c8xH100 \u63a8\u7406\u5355\u5143\u7684\u541e\u5410\u91cf\u53ef\u63a5\u8fd1\u6bcf\u7528\u6237\u6bcf\u79d2 420 \u4e2a token\u3002\u867d\u7136\u541e\u5410\u91cf\u53ef\u80fd\u4f1a\u8d85\u8fc7\u8fd9\u4e00\u6570\u5b57\uff0c\u4f46\u5728 MoE \u6a21\u578b\u4e0a\u5b9e\u73b0\u63a8\u6d4b\u89e3\u7801\u5177\u6709\u4e00\u5b9a\u6311\u6218\u6027\u3002<\/p>\n\n\n\n<p>\u76ee\u524d\u5c1a\u4e0d\u5b58\u5728\u9488\u5bf9\u5ef6\u8fdf\u4f18\u5316\u7684 API \u670d\u52a1\uff0c\u56e0\u4e3a\u5176\u7ecf\u6d4e\u6548\u76ca\u5f88\u5dee\u3002\u76ee\u524d\uff0cAPI \u670d\u52a1\u5546\u5c1a\u672a\u770b\u5230\u5e02\u573a\u4e0a\u613f\u610f\u652f\u4ed8 10 \u500d\u8d39\u7528\u4ee5\u964d\u4f4e\u5ef6\u8fdf\u7684\u9700\u6c42\u3002\u7136\u800c\uff0c\u968f\u7740 agent \u548c\u5176\u4ed6\u6781\u4f4e\u5ef6\u8fdf\u4efb\u52a1\u7684\u666e\u53ca\uff0c\u57fa\u4e8e GPU \u7684 API \u670d\u52a1\u5546\u5f88\u53ef\u80fd\u4f1a\u5728\u63a8\u51fa\u5ef6\u8fdf\u4f18\u5316 API\uff0c\u4ee5\u8865\u5145\u4ed6\u4eec\u5f53\u524d\u9488\u5bf9\u541e\u5410\u91cf\u4f18\u5316\u7684API\u3002<\/p>\n\n\n\n<p>Groq \u5c06\u5728\u4e0b\u5468\u5b9e\u65bd\u4ed6\u4eec\u7684 batch \u7cfb\u7edf\uff0c\u4e0e\u6ca1\u6709\u91c7\u7528\u63a8\u6d4b\u89e3\u7801\u7684 Groq \u76f8\u6bd4\uff0c\u867d\u7136 Nvidia \u91c7\u7528\u4e86\u5177\u6709\u63a8\u6d4b\u89e3\u7801\u7684\u4f4e\u5ef6\u8fdf\u4f18\u5316\u7cfb\u7edf\uff0c\u4f46\u5728\u541e\u5410\u91cf\u548c\u6210\u672c\u65b9\u9762\u4ecd\u7136\u8fdc\u8fdc\u843d\u540e\u4e8e Groq\u3002\u6b64\u5916\uff0cGroq \u91c7\u7528\u7684\u662f\u8f83\u8001\u7684 14 \u7eb3\u7c73\u5de5\u827a\u6280\u672f\uff0c\u5e76\u5411 Marvell \u652f\u4ed8\u4e86\u53ef\u89c2\u7684\u82af\u7247\u5229\u6da6\u3002\u5982\u679c Groq \u80fd\u591f\u83b7\u5f97\u66f4\u591a\u8d44\u91d1\uff0c\u5e76\u5728 2025 \u5e74\u4e0b\u534a\u5e74\u5de6\u53f3\u6295\u4ea7\u4e0b\u4e00\u4ee3 4 \u7eb3\u7c73\u82af\u7247\uff0c\u90a3\u4e48\u7ecf\u6d4e\u6548\u76ca\u53ef\u80fd\u4f1a\u53d1\u751f\u91cd\u5927\u53d8\u5316\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0cNvidia \u5e76\u4e0d\u4f1a\u5750\u4ee5\u5f85\u6bd9\uff0c\u636e\u6211\u4eec\u6240\u77e5\uff0c\u4ed6\u4eec\u5c06\u5728\u672a\u6765\u4e0d\u5230\u4e00\u4e2a\u6708\u7684\u65f6\u95f4\u5185\u53d1\u5e03\u4e0b\u4e00\u4ee3 B100 \u82af\u7247\u3002<\/p>\n\n\n\n<p>\u5728\u4e00\u4e2a\u4ee5\u541e\u5410\u91cf\u4e3a\u4f18\u5316\u76ee\u6807\u7684\u7cfb\u7edf\u4e2d\uff0c\u7ecf\u6d4e\u60c5\u51b5\u4f1a\u53d1\u751f\u663e\u8457\u53d8\u5316\u3002\u6839\u636e BOM \u8ba1\u7b97\uff0cNvidia \u7cfb\u7edf\u6bcf\u7f8e\u5143\u7684\u6027\u80fd\u63d0\u5347\u4e00\u4e2a\u6570\u91cf\u7ea7\uff0c\u4f46\u6bcf\u4e2a\u7528\u6237\u7684\u541e\u5410\u91cf\u8f83\u4f4e\u3002\u5bf9\u4e8e\u541e\u5410\u91cf\u4f18\u5316\u65b9\u6848\uff0cGroq \u5728\u67b6\u6784\u4e0a\u5b8c\u5168\u4e0d\u5177\u5907\u7ade\u4e89\u529b\u3002<\/p>\n\n\n\n<p>\u7136\u800c\uff0c\u5bf9\u4e8e\u8d2d\u4e70\u548c\u90e8\u7f72\u7cfb\u7edf\u7684\u7528\u6237\u6765\u8bf4\uff0c\u4e0a\u8ff0\u7b80\u5316\u5206\u6790\u5e76\u4e0d\u662f\u6b63\u786e\u7684\u5546\u4e1a\u6848\u4f8b\u5206\u6790\u65b9\u6cd5\uff0c\u56e0\u4e3a\u8fd9\u79cd\u5206\u6790\u5ffd\u7565\u4e86\u7cfb\u7edf\u6210\u672c\u3001\u5229\u6da6\u548c\u529f\u8017\u7b49\u56e0\u7d20\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u4ecb\u7ecd\u6027\u80fd \/ TCO \u7684\u5206\u6790\u6cd5\u3002<\/p>\n\n\n\n<p>\u4e00\u65e6\u6211\u4eec\u8003\u8651\u5230\u8fd9\u4e9b\u56e0\u7d20\uff0ctoken \u7ecf\u6d4e\u5b66\uff08\u8fd9\u4e2a\u65f6\u9ae6\u7684\u65b0\u8bcd\u662f\u7531 swyx \u521b\u9020\u7684\uff09\u7684\u60c5\u51b5\u5c31\u5b8c\u5168\u4e0d\u540c\u4e86\u3002\u5bf9\u4e8e Nvidia \u6765\u8bf4\uff0c\u6211\u4eec\u5c06\u4f7f\u7528\u6b64\u5904\u6240\u89e3\u91ca\u7684 GPU \u4e91\u7ecf\u6d4e\u5b66\uff0c\u5177\u4f53\u5185\u5bb9\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"1080\" height=\"770\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-45.webp\" alt=\"\" class=\"wp-image-2797\"\/><figcaption class=\"wp-element-caption\">\u8d44\u672c\u6210\u672c\u5305\u62ec\u6700\u4f4e\u8d44\u672c\u56de\u62a5\u7387\uff0c\u5373\u63d0\u51fa\u8be5\u5546\u4e1a\u6848\u4f8b\u7684\u4eba\u671f\u671b\u83b7\u5f97\u7684\u6295\u8d44\u56de\u62a5\uff0c\u4ee5\u8bc1\u660e\u9879\u76ee\u7684\u98ce\u9669\u662f\u5408\u7406\u7684\u3002<\/figcaption><\/figure><\/div>\n\n\n<p>Nvidia \u7684 GPU \u4e3b\u677f\u6bdb\u5229\u7387\u975e\u5e38\u9ad8\u3002\u5176\u670d\u52a1\u5668\u4ef7\u683c\u9ad8\u8fbe 350,000 \u7f8e\u5143\uff08\u8fdc\u9ad8\u4e8e H100 \u670d\u52a1\u5668\u7684\u8d85\u7ea7\u5206\u9891\u5668\u6210\u672c\uff09\uff0c\u5176\u4e2d\u8fd8\u5305\u62ec\u5185\u5b58\u30018 \u4e2a\u603b\u5e26\u5bbd\u4e3a 3.2Tbps \u7684 InfiniBand \u7f51\u5361\uff08\u5bf9\u4e8e\u8fd9\u79cd\u63a8\u7406\u5e94\u7528\u6765\u8bf4\u5e76\u4e0d\u9700\u8981\uff09\u7684\u5de8\u989d\u6210\u672c\uff0c\u4ee5\u53ca\u5728Nvidia\u6bdb\u5229\u7387\u4e4b\u4e0a\u53e0\u52a0\u4e86\u76f8\u5f53\u53ef\u89c2\u7684OEM\u5236\u9020\u5546\u7684\u5229\u6da6\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e Groq \u800c\u8a00\uff0c\u6211\u4eec\u4f30\u7b97\u7684\u662f\u7cfb\u7edf\u6210\u672c\uff0c\u5e76\u5c06\u82af\u7247\u3001\u5c01\u88c5\u3001\u7f51\u7edc\u3001CPU\u3001\u5185\u5b58\u7b49\u7ec6\u8282\u56e0\u7d20\u8003\u8651\u5728\u5185\uff0c\u540c\u65f6\u5047\u5b9a\u4e86\u8f83\u4f4e\u7684\u6574\u4f53 ODM \u5229\u6da6\u3002\u6211\u4eec\u4e5f\u4e0d\u5305\u62ec Groq \u56e0\u9500\u552e\u786c\u4ef6\u800c\u6536\u53d6\u7684\u5229\u6da6\uff0c\u56e0\u6b64\uff0c\u867d\u7136\u770b\u8d77\u6765\u662f\u82f9\u679c\u4e0e\u6a58\u5b50\u7684\u5bf9\u6bd4\uff0c\u4f46\u4e5f\u662f Groq \u6210\u672c\u4e0e\u63a8\u7406 API \u670d\u52a1\u5546\u6210\u672c\u7684\u516c\u5e73\u5bf9\u6bd4\uff0c\u56e0\u4e3a\u4e24\u8005\u63d0\u4f9b\u7684\u662f\u76f8\u540c\u7684\u4ea7\u54c1\/\u6a21\u578b\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"903\" height=\"626\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-46.webp\" alt=\"\" class=\"wp-image-2798\"\/><\/figure><\/div>\n\n\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c8 \u4e2a Nvidia GPU \u53ea\u9700\u8981 2 \u4e2a CPU\uff0c\u800c Groq \u7684 576 \u82af\u7247\u7cfb\u7edf\u76ee\u524d\u5219\u914d\u5907\u4e86 144 \u4e2a CPU \u548c 144TBs \u7684 RAM\u3002<\/p>\n\n\n\n<p>\u5c06\u8fd9\u4e9b\u7ec4\u4ef6\u6210\u672c\u76f8\u52a0\uff0c\u6211\u4eec\u5f97\u51fa\u6bcf\u4e2a Groq LPU \u670d\u52a1\u5668\u7684\u6210\u672c\u4e3a 35,000 \u7f8e\u5143\uff0c\u5176\u4e2d\u5305\u62ec 8 \u4e2a Groq LPU \u548c\u4e0a\u8ff0\u6240\u6709\u5176\u4ed6\u786c\u4ef6\u3002Mixtral Groq \u63a8\u7406\u90e8\u7f72\u4f7f\u7528\u4e86 8 \u4e2a\u673a\u67b6\uff0c\u6bcf\u4e2a\u673a\u67b6\u6709 9 \u53f0\u670d\u52a1\u5668\uff0c\u5373\u603b\u5171 576 \u4e2a LPU \u82af\u7247\u7684\u63a8\u7406\u90e8\u7f72\u6210\u672c\u4e3a 252 \u4e07\u7f8e\u5143\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5178\u578b\u7684 H100 HGX \u7cfb\u7edf\u7684\u524d\u671f\u8d44\u672c\u652f\u51fa\u4e3a 350,000 \u7f8e\u5143\uff0c\u5305\u62ec 8 \u4e2a H100\u3002\u5927\u591a\u6570\u57fa\u4e8e H100 \u7684 Mixtral \u63a8\u7406\u5b9e\u4f8b\u53ea\u4f7f\u7528 2 \u4e2a H100 \u82af\u7247\uff0c\u56e0\u6b64\u6bcf\u53f0 H100 \u670d\u52a1\u5668\u53ef\u90e8\u7f72 4 \u4e2a\u63a8\u7406\u5b9e\u4f8b\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"1080\" height=\"714\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-47.webp\" alt=\"\" class=\"wp-image-2799\"\/><\/figure><\/div>\n\n\n<p>\u5047\u8bbe\u95e8\u69db\u6536\u76ca\u7387\u4e3a18%\uff0c\u4f7f\u7528\u5bff\u547d\u4e3a 5 \u5e74\uff0c\u90a3\u4e48\u4e00\u4e2a H100 \u7cfb\u7edf\uff08\u5373\u5177\u6709 8 \u4e2a H100 GPU \u7684 H100 HGX\uff09\u7684\u603b\u644a\u9500\u6708\u8d44\u672c\u6210\u672c\u4e3a 8,888 \u7f8e\u5143\uff0c\u800c\u6bcf\u6708\u6258\u7ba1\u8d39\u7528\u4e3a 2,586 \u7f8e\u5143\uff0cTCO \u4e3a 11,474 \u7f8e\u5143\uff0c\u800c\u66f4\u5927\u89c4\u6a21\u7684 Groq \u7cfb\u7edf\u7684\u603b\u62e5\u6709\u6210\u672c\u4e3a\u6bcf\u6708 122,400 \u7f8e\u5143\u3002<\/p>\n\n\n\n<p>\u867d\u7136\u6574\u4e2a Groq \u7cfb\u7edf\uff089 \u53f0\u670d\u52a1\u5668\u4e58\u4ee5 8 \u4e2a\u673a\u67b6\uff09\u7684\u6708\u644a\u9500\u8d44\u672c\u6210\u672c\u662f H100 HGX \u7684 7.2 \u500d\uff08Groq \u4e3a 63,991 \u7f8e\u5143\uff0cH100 HGX \u4e3a 8,888 \u7f8e\u5143\uff09\uff0c\u4f46 Groq \u7cfb\u7edf\u5728 FP16 \u65f6\u7684 FLOPS \u6027\u80fd\u5374\u662f H100 HGX \u7684 13.7 \u500d\uff08Groq \u4e3a 108,000 TFLOPS\uff0cH100 HGX \u4e3a 7,912 TFLOPS\uff09\u3002\u7531\u4e8e\u5185\u5b58\u5899\u7684\u5b58\u5728\uff0c\u57fa\u4e8e H100 \u7684\u63a8\u7406\u7cfb\u7edf\u901a\u5e38\u5177\u6709\u8f83\u4f4e\u7684 FLOPS \u5229\u7528\u7387\uff0c\u800c Groq \u7684\u67b6\u6784\u901a\u8fc7\u82af\u7247\u4e0a\u7684 SRAM \u907f\u5f00\u4e86\u5185\u5b58\u5899\u3002\u5c3d\u7ba1\u5982\u6b64\uff0c\u7531\u4e8e\u67d0\u4e9b\u539f\u56e0\uff08\u65e0\u8bba\u662f\u7f3a\u4e4f\u7f13\u51b2\u533a\u8fd8\u662f VLIW \u67b6\u6784\uff09\uff0c\u5373\u4f7f\u4e0b\u5468\u63a8\u9001\u7684 batch size 3 \u5df2\u7ecf\u5b9e\u73b0\uff0cGroq \u7684 FLOPS \u5229\u7528\u7387\u90fd\u8fd8\u662f\u4f4e\u4e8eNvidia\u3002<\/p>\n\n\n\n<p>\u63a8\u65ad API \u670d\u52a1\u5546\uff08\u5982 Nvidia\uff09\u8d2d\u4e70\u7cfb\u7edf\u65f6\uff0c\u8003\u8651\u5230 OEM \u9644\u52a0\u5229\u6da6\uff0c\u5176\u6bdb\u5229\u7387\u901a\u5e38\u8d85\u8fc7 80%\uff0c\u800c Groq \u5b9e\u9645\u4e0a\u662f\u4ee5\u6210\u672c\u4ef7\u8d2d\u4e70\u7cfb\u7edf\u3002\u5c3d\u7ba1\u4ed6\u4eec\u5fc5\u987b\u5411 SuperMicro \u548c Marvell \u5206\u522b\u652f\u4ed8\u7cfb\u7edf\u548c\u82af\u7247\u7684\u6210\u672c\uff0c\u4f46\u8fdc\u672a\u8fbe\u5230 API \u670d\u52a1\u5546\u53ca\u5176 GPU \u4e91\u670d\u52a1\u5546\u7684\u6c34\u5e73\u3002\u56e0\u6b64\uff0cGroq \u7684\u603b\u62e5\u6709\u6210\u672c\u53d7\u8d44\u672c\u6295\u5165\u7684\u5f71\u54cd\u8981\u5c0f\u5f97\u591a\uff0cGroq \u7684\u8d44\u672c\u6295\u5165\u5360\u603b\u62e5\u6709\u6210\u672c\u7684 52%\uff0c\u800c H100 \u7cfb\u7edf\u7684\u8d44\u672c\u6295\u5165\u5360\u603b\u62e5\u6709\u6210\u672c\u7684\u8fd1 80%\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"1080\" height=\"597\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-48.webp\" alt=\"\" class=\"wp-image-2800\"\/><\/figure><\/div>\n\n\n<p>8xH100 \u7684\u5ef6\u8fdf\u4f18\u5316\u63a8\u7406\u90e8\u7f72\u6bcf\u767e\u4e07\u4e2a token \u7684\u6210\u672c\u4e3a 5.2 \u7f8e\u5143\u30022xH100 \u7684\u541e\u5410\u91cf\u4f18\u5316\u63a8\u7406\u90e8\u7f72\u6bcf\u767e\u4e07\u4e2a token \u7684\u6210\u672c\u4e3a 0.57 \u7f8e\u5143\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cGroq \u7684\u6210\u672c\u4e3a\u6bcf\u767e\u4e07\u4e2a token 1.94 \u7f8e\u5143\uff0c\u6bd4 8xH100 \u66f4\u5feb\u3001\u66f4\u4fbf\u5b9c\u3002<\/p>\n\n\n\n<p>\u4e0e\u8bb8\u591a\u63a8\u7406\u670d\u52a1\u670d\u52a1\u5546\u4e00\u6837\uff0cGroq \u76ee\u524d\u7684\u5546\u4e1a\u6a21\u5f0f\u5904\u4e8e\u8d1f\u6bdb\u5229\u72b6\u6001\uff0c\u5b83\u81f3\u5c11\u9700\u8981\u5c06\u541e\u5410\u91cf\u63d0\u9ad8 7 \u500d\u4ee5\u4e0a\u624d\u80fd\u5b9e\u73b0\u76c8\u4e8f\u5e73\u8861\u3002\u4e0e\u5ef6\u8fdf\u4f18\u5316\u63a8\u7406\u90e8\u7f72\u4e2d\u7684 8xH100 \u8bbe\u5907\u76f8\u6bd4\uff0c\u8fd9\u4e2a\u5dee\u8ddd\u975e\u5e38\u5927\u3002\u5982\u679c\u91c7\u7528\u76f8\u540c\u7684\u5b9a\u4ef7\uff0c8xH100 \u8bbe\u5907\u4e0e\u6536\u652f\u5e73\u8861\u76f8\u5dee\u8fd1 20 \u500d\u3002<\/p>\n\n\n\n<p>\u9664\u4e86\u9500\u552e\u63a8\u7406 API \u670d\u52a1\u5916\uff0cGroq \u7684\u5546\u4e1a\u6a21\u5f0f\u8fd8\u5305\u62ec\u76f4\u63a5\u9500\u552e\u7cfb\u7edf\u3002\u5982\u679c Groq \u5c06\u5176\u7cfb\u7edf\u4ee5 60% \u7684\u6bdb\u5229\u7387\u51fa\u552e\u7ed9\u7b2c\u4e09\u65b9\u8fd0\u8425\u5546\uff0c\u8fd9\u5c06\u4e0e H100 HGX \u7684\u603b\u62e5\u6709\u6210\u672c\u7684\u8d44\u672c\u5bc6\u5ea6\u5927\u81f4\u76f8\u5f53\uff0c\u7cfb\u7edf\u4ef7\u683c\u5c06\u5927\u7ea6\u4e3a 635 \u4e07\u7f8e\u5143\u3002<\/p>\n\n\n\n<p>Groq \u58f0\u79f0\u4ed6\u4eec\u5728\u529f\u8017\u65b9\u9762\u5177\u6709\u4f18\u52bf\uff0c\u4f46\u6211\u4eec\u5e76\u672a\u770b\u5230\u6709\u660e\u663e\u8bc1\u636e\u80fd\u591f\u652f\u6301\u8fd9\u4e00\u70b9\u3002\u5373\u4f7f\u5bf9\u4e8e H100 \u670d\u52a1\u5668\u91c7\u53d6\u6700\u4fdd\u5b88\u7684\u5047\u8bbe\uff0c\u5373 10 \u5343\u74e6\uff08\u5305\u62ec CPU \u548c\u6240\u6709 8 \u4e2a\u7f51\u5361\u5168\u901f\u8fd0\u884c\uff09\uff0c\u4e5f\u6bd4 576 \u82af\u7247\u7684 Groq \u670d\u52a1\u5668\u66f4\u9ad8\u6548\uff0c\u540e\u8005\u9700\u8981 230 \u5343\u74e6\uff0c\u5373\u6bcf\u53f0 8 \u82af\u7247\u670d\u52a1\u5668\u9700\u8981 3.2 \u5343\u74e6\u3002\u867d\u7136 Groq \u58f0\u79f0\u5177\u6709\u6bcf\u74e6\u6027\u80fd\u4f18\u52bf\uff0c\u4f46\u6211\u4eec\u5e76\u4e0d\u6e05\u695a\u5176\u8ba1\u7b97\u65b9\u5f0f\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"1080\" height=\"291\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-49.webp\" alt=\"\" class=\"wp-image-2801\"\/><\/figure><\/div>\n\n\n<p>\u6211\u4eec\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u867d\u7136 Groq \u76ee\u524d\u5728\u5176 API \u4e1a\u52a1\u4e0a\u5904\u4e8e\u4e8f\u635f\u72b6\u6001\uff0c\u9700\u8981\u8d85\u8fc7 7.2 \u500d\u7684\u541e\u5410\u91cf\u624d\u80fd\u5b9e\u73b0\u76c8\u5229\uff08\u5305\u62ec\u4e0b\u5468\u8ba1\u5212\u5b9e\u65bd\u7684 batch 3 \u6539\u8fdb\uff09\uff0c\u4f46\u4ed6\u4eec\u58f0\u79f0\u5df2\u7ecf\u5236\u5b9a\u4e86\u4e00\u4e2a\u6539\u8fdb\u8def\u7ebf\u56fe\uff0c\u53ef\u4ee5\u5728\u672a\u6765\u51e0\u4e2a\u5b63\u5ea6\u5185\u5b9e\u73b0\u76c8\u4e8f\u5e73\u8861\uff0c\u4e3b\u8981\u901a\u8fc7\u4ee5\u4e0b\u4e09\u4e2a\u6539\u8fdb\u9014\u5f84\u6765\u5b9e\u73b0\uff1a<\/p>\n\n\n\n<ol>\n<li>\u7ee7\u7eed\u7f16\u8bd1\u5de5\u4f5c\uff0c\u63d0\u9ad8\u541e\u5410\u91cf\uff1b<\/li>\n\n\n\n<li>\u91c7\u7528\u65b0\u7684\u670d\u52a1\u5668\u67b6\u6784\uff0c\u5927\u5e45\u964d\u4f4e\u5305\u62ec\u677f\u5361\u5728\u5185\u7684\u975e\u82af\u7247\u6210\u672c\uff0c\u4f7f\u7528\u66f4\u5c11\u7684 CPU \u548c\u66f4\u5c11\u7684 DRAM\uff1b<\/li>\n\n\n\n<li>\u90e8\u7f72\u66f4\u5927\u89c4\u6a21\u7684\u7cfb\u7edf\uff0c\u901a\u8fc7\u589e\u52a0\u66f4\u591a\u7684\u7ba1\u9053\u5b9e\u73b0\u8d85\u7ebf\u6027\u6027\u80fd\u6269\u5c55\uff0c\u4ece\u800c\u63d0\u9ad8 batching \u5904\u7406\u80fd\u529b\uff0c\u5e76\u6700\u7ec8\u652f\u6301\u66f4\u5927\u7684\u6a21\u578b\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u867d\u7136\u8fd9\u4e9b\u9879\u76ee\u4e2d\u7684\u6bcf\u4e00\u9879\u90fd\u662f\u5408\u4e4e\u903b\u8f91\u7684\uff0c\u6211\u4eec\u4e5f\u5e0c\u671b\u4ed6\u4eec\u80fd\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u4f46 7 \u500d\u541e\u5410\u91cf\u7684\u6539\u8fdb\u662f\u76f8\u5f53\u56f0\u96be\u7684\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u8ba4\u4e3a\u5b58\u5728\u51e0\u4e2a\u4e3b\u8981\u6311\u6218\uff1a<\/p>\n\n\n\n<p>\u76ee\u524d\uff0c\u6700\u5927\u7684 MoE \u6a21\u578b\u7684\u53c2\u6570\u8303\u56f4\u5728 1-2 \u4e07\u4ebf\u4e2a\uff0c<strong>\u6211\u4eec\u9884\u8ba1\u8c37\u6b4c\u548cOpenAI\u5c06\u5728\u660e\u5e74\u63a8\u51fa\u8d85\u8fc710\u4e07\u4ebf\u4e2a\u53c2\u6570\u7684\u6a21\u578b<\/strong>\uff0c\u8fd9\u5c06\u9700\u8981\u6570\u767e\u4e2a GPU \u548c\u6570\u5341TB\u5185\u5b58\u7684\u63a8\u7406\u7cfb\u7edf\u3002\u867d\u7136 Mixtral \u76ee\u524d\u662f\u5fae\u8c03\u3001API \u670d\u52a1\u548c\u672c\u5730\u90e8\u7f72\u4e2d\u6700\u91cd\u8981\u7684\u6a21\u578b\uff0c\u4f46 3 \u4e2a\u6708\u540e\u60c5\u51b5\u5c06\u4f1a\u6709\u6240\u4e0d\u540c\u3002LLAMA 3 \u548c\u66f4\u5927\u89c4\u6a21\u7684 Mistral \u6a21\u578b\u5373\u5c06\u63a8\u51fa\u3002<\/p>\n\n\n\n<p>Groq \u516c\u53f8\u5df2\u7ecf\u8bc1\u660e\u4ed6\u4eec\u80fd\u591f\u6784\u5efa\u9002\u7528\u4e8e\u5c11\u4e8e 1,000 \u4ebf\u53c2\u6570\u7684\u6a21\u578b\u7684\u7cfb\u7edf\uff0c\u5373\u4f7f\u662f\u50cf LLAMA 3 \u8fd9\u6837\u5927\u5c0f\u7684\u6a21\u578b\u4e5f\u5f88\u96be\u9002\u5e94\u672a\u6765\u7406\u8bba\u4e0a\u7684\u4e0a\u5343\u82af\u7247\u7cfb\u7edf\u3002Groq \u8ba1\u5212\u5728\u4e24\u5e74\u5185\u90e8\u7f72 100 \u4e07\u4e2a\u82af\u7247\uff0c\u6bcf\u4e2a\u63a8\u7406\u7cfb\u7edf\u7684\u89c4\u6a21\u90fd\u5c06\u8d85\u8fc7\u76ee\u524d\u90e8\u7f72\u7684 576 \u4e2a\u82af\u7247\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e5f\u5f15\u51fa\u4e86\u6211\u4eec\u770b\u5230\u7684\u53e6\u4e00\u4e2a\u540c\u6837\u8270\u5de8\u7684\u6311\u6218\uff0c\u5373\u6781\u957f\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002\u8c37\u6b4c\u5728 Gemini 1.5 Pro \u4e0a\u5c55\u793a\u4e86\u4e00\u4e2a\u4ee4\u4eba\u60ca\u53f9\u7684\u529f\u80fd\uff0c\u80fd\u591f\u5904\u7406 1,000 \u4e07\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684 token\u3002\u8fd9\u6837\u7684\u957f\u5ea6\u8db3\u4ee5\u5bb9\u7eb3 10 \u5c0f\u65f6\u7684\u89c6\u9891\u3001110 \u5c0f\u65f6\u7684\u97f3\u9891\u300130 \u4e07\u884c\u7684\u4ee3\u7801\u6216 700 \u4e07\u4e2a\u5b57\u3002\u6211\u4eec\u9884\u8ba1\u8bb8\u591a\u516c\u53f8\u548c\u670d\u52a1\u673a\u6784\u90fd\u80fd\u63d0\u4f9b\u5177\u6709\u8d85\u5927\u5e8f\u5217\u957f\u5ea6\u7684\u6a21\u578b\uff0c\u4ee5\u4fbf\u80fd\u591f\u5904\u7406\u4ee3\u7801\u5e93\u548c\u6587\u6863\u5e93\u7b49\u5927\u91cf\u63d0\u793a\u4fe1\u606f\uff0c\u56e0\u4e3a\u8fd9\u79cd\u6027\u80fd\u8fdc\u8fdc\u8d85\u8d8a\u4e86 RAG \u6a21\u578b\uff0c\u5728\u5b9e\u9645\u5e94\u7528\u4e2d RAG \u6a21\u578b\u901a\u5e38\u8868\u73b0\u4e0d\u4f73\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"741\" height=\"520\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/02\/image-50.webp\" alt=\"\" class=\"wp-image-2802\"\/><\/figure><\/div>\n\n\n<p>\u5f53\u7136\uff0c\u9884\u5904\u7406\u5728\u521d\u59cb\u9636\u6bb5\u53ef\u80fd\u4f1a\u82b1\u8d39\u76f8\u5f53\u957f\u7684\u65f6\u95f4\uff08\u8c37\u6b4c\u6570\u636e\u663e\u793a\uff0c\u9700\u8981\u9700\u8981\u6574\u6574\u4e00\u5206\u949f\u624d\u80fd\u8f93\u51fa\u7b2c\u4e00\u4e2a token\uff09\uff0c\u4f46\u5728\u8fd9\u4e4b\u540e\uff0c\u9884\u5904\u7406\u7684\u6210\u672c\u4f1a\u5728\u4f17\u591a\u8bf7\u6c42\u4e2d\u644a\u9500\uff0c\u5927\u89c4\u6a21\u7684\u7279\u5b9a\u4e8e\u5ba2\u6237\u7684\u63d0\u793a\u4fe1\u606f\u4e0d\u9700\u8981\u9891\u7e41\u91cd\u65b0\u8ba1\u7b97\u3002\u867d\u7136\u8c37\u6b4c\u6ca1\u6709\u4f7f\u7528\u6269\u5c55\u4e3aO(n^2)\u7684\u6807\u51c6\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4f46 Gemini 1.5 Pro \u4ecd\u7136\u9700\u8981\u6570\u767e GB \u751a\u81f3\u662f TB \u7ea7\u522b\u7684\u5185\u5b58\u6765\u5b58\u50a8 KVCache\u3002<\/p>\n\n\n\n<p>\u8003\u8651\u5230 KVCache \u7684\u5927\u5c0f\u8981\u6c42\uff0c\u6211\u4eec\u5f88\u96be\u770b\u51fa Groq \u5982\u4f55\u80fd\u591f\u5b9e\u73b0\u8d85\u5927\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002\u5b83\u9700\u8981\u6210\u5343\u4e0a\u4e07\u82af\u7247\u7684\u7cfb\u7edf\uff0c\u800c\u4e0d\u662f\u50cf\u8c37\u6b4c\u3001Nvidia \u548c AMD \u63a8\u7406\u89e3\u51b3\u65b9\u6848\u90a3\u6837\u53ea\u4f7f\u7528 10 \u6216 100 \u5757\u82af\u7247\u3002<\/p>\n\n\n\n<p>\u867d\u7136 Groq \u5728\u4f4e\u5ef6\u8fdf\u7f51\u7edc\u82af\u7247\u65b9\u9762\u7684\u80fd\u529b\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\uff0c\u4f46\u8981\u5c06\u5176\u6269\u5c55\u5230\u9700\u8981\u6570\u4e07\u5757\u82af\u7247\u7684\u8d85\u957f\u4e0a\u4e0b\u6587\uff0c\u4ee5\u53ca\u50cf Gemini Pro 1.5 \u8fd9\u6837\u7684\u4e2d\u7b49\u89c4\u6a21\u6a21\u578b\u6216\u50cf GPT-5 \u548c Gemini Ultra 2 \u8fd9\u6837\u7684\u8d85\u5927\u6a21\u578b\uff0c\u5c06\u4f1a\u975e\u5e38\u56f0\u96be\u3002<\/p>\n\n\n\n<p>\u8fd9\u4f7f\u5f97\u4eba\u4eec\u5f00\u59cb\u8d28\u7591\u8fd9\u4e9b\u5e9e\u5927\u7684\u4eba\u5de5\u667a\u80fd\u5efa\u8bbe\u9879\u76ee\u7684\u4f7f\u7528\u5bff\u547d\u3002\u5bf9\u4e8e GPU \u6765\u8bf4\uff0c\u7531\u4e8e\u5176\u7075\u6d3b\u6027\uff0c\u5f88\u5bb9\u6613\u770b\u5230\u5b83\u4eec\u5728\u672a\u6765 4 \u5e74\u5185\u5bf9\u65b0\u6a21\u578b\u4ecd\u7136\u6709\u7528\u3002\u4f46\u7531\u4e8e Groq \u7f3a\u4e4f DRAM\uff0c\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u4e0d\u65ad\u589e\u52a0\uff0c\u5176\u7075\u6d3b\u6027\u4e5f\u4f1a\u5927\u6253\u6298\u6263\u3002\u5982\u679c\u8fd9\u6210\u4e3a\u4e00\u4e2a\u95ee\u9898\uff0c\u90a3\u4e48\u5b83\u7684\u7cfb\u7edf\u6298\u65e7\u5e74\u9650\u5c06\u4e0d\u518d\u662f 5 \u5e74\uff0c\u800c\u662f\u66f4\u77ed\u3002\u8fd9\u5c06\u5927\u5927\u589e\u52a0\u6210\u672c\u3002<\/p>\n\n\n\n<p>Groq \u9762\u4e34\u7684\u53e6\u4e00\u4e2a\u6311\u6218\u662f\uff0c\u63a8\u6d4b\u89e3\u7801\u548c Medusa \u7b49\u6280\u672f\u7684\u5feb\u901f\u53d1\u5c55\u3002Tree\/branch \u63a8\u7406\u6cd5\u4f7f\u63a8\u6d4b\u89e3\u7801\u901f\u5ea6\u63d0\u9ad8\u4e86 3 \u500d\u4ee5\u4e0a\u3002\u5982\u679c\u8fd9\u4e9b\u6280\u672f\u80fd\u591f\u5728\u751f\u4ea7\u7ea7\u7cfb\u7edf\u4e0a\u8fdb\u884c\u6709\u6548\u90e8\u7f72\uff0c\u90a3\u4e48\u4e00\u4e2a 8x H100 \u7cfb\u7edf\u6bcf\u79d2\u5c31\u80fd\u5904\u7406\u8d85\u8fc7 600 \u4e2a token\u3002<\/p>\n\n\n\n<p>Groq \u8fd8\u8868\u793a\u4ed6\u4eec\u8ba1\u5212\u5728\u672a\u6765\u5b9e\u73b0\u63a8\u6d4b\u89e3\u7801\uff0c\u4f46\u6211\u4eec\u4e0d\u660e\u767d\u8fd9\u5c06\u5982\u4f55\u4e0e\u4ed6\u4eec\u7684\u786e\u5b9a\u6027\u67b6\u6784\u76f8\u914d\u5408\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u63a8\u6d4b\u89e3\u7801\u9700\u8981\u6d88\u8017\u66f4\u591a\u7684 FLOPS \u4ee5\u6362\u53d6\u901a\u8fc7\u63d0\u9ad8 batch size \u6765\u63d0\u9ad8\u5e26\u5bbd\u6548\u7387\u3002Groq \u4e3b\u8981\u53d7\u9650\u4e8e FLOPS \u548c\u7f51\u7edc\uff0c\u800c\u4e0d\u662f SRAM \u5e26\u5bbd\u3002Groq \u9700\u8981\u5c06 batch \u80fd\u529b\u5927\u5e45\u63d0\u9ad8\u5230 3 \u4ee5\u4e0a\uff0c\u624d\u80fd\u6709\u6548\u5b9e\u73b0\u63a8\u6d4b\u89e3\u7801\u3002<\/p>\n\n\n\n<p>\u6700\u540e\uff0cB100 \u5c06\u4e8e\u4e0b\u4e2a\u6708\u5ba3\u5e03\u63a8\u51fa\uff0c\u5e76\u5728\u4e0b\u534a\u5e74\u5f00\u59cb\u53d1\u8d27\uff0c\u636e\u4f20\u95fb\uff0c\u5176\u6027\u80fd \/ TCO \u6539\u8fdb\u9884\u8ba1\u5c06\u8d85\u8fc7 H100 \u7684\u4e24\u500d\u3002\u6b64\u5916\uff0cNvidia \u7684\u53d1\u5c55\u901f\u5ea6\u975e\u5e38\u8fc5\u901f\uff0c\u9884\u8ba1\u4e24\u4e2a\u5b63\u5ea6\u540e\u5c06\u63a8\u51fa B200\uff0c\u5e76\u5728\u4e4b\u540e\u4e24\u4e2a\u5b63\u5ea6\u63a8\u51fa X\/R100\u3002Nvidia \u7684\u53d1\u5c55\u76ee\u6807\u662f\u4e0d\u65ad\u53d8\u5316\u7684\u3002<\/p>\n\n\n\n<p>\u5c3d\u7ba1\u5982\u6b64\uff0c\u5982\u679c Groq \u80fd\u591f\u6709\u6548\u5730\u6269\u5c55\u5230\u6570\u5343\u82af\u7247\u7684\u7cfb\u7edf\uff0c\u7ba1\u9053\u6570\u91cf\u5c06\u4f1a\u5927\u5e45\u589e\u52a0\uff0c\u968f\u7740\u7ba1\u9053\u6570\u91cf\u7684\u589e\u52a0\uff0c\u5c06\u4ea7\u751f\u989d\u5916\u7684 SRAM\uff0c\u4ece\u800c\u4e3a\u66f4\u5927\u7684 batch size \u63d0\u4f9b\u66f4\u591a\u7684 KVCache\u3002\u8fd9\u53cd\u8fc7\u6765\u53c8\u80fd\u5b9e\u73b0\u5927\u4e8e 10 \u7684 batch size\uff0c\u5e76\u6709\u671b\u5927\u5e45\u964d\u4f4e\u6210\u672c\u3002\u6211\u4eec\u8ba4\u4e3a\u8fd9\u662f\u6709\u53ef\u80fd\u7684\uff0c\u4f46\u53ef\u80fd\u6027\u4e0d\u662f\u5f88\u9ad8\u3002\u73b0\u5728\u7684\u5173\u952e\u5728\u4e8e\u770b\u770b Groq \u80fd\u5426\u8bc1\u660e\u6211\u4eec\u7684\u9884\u6d4b\u662f\u9519\u8bef\u7684\uff0c\u5e76\u4e14\u80fd\u591f\u663e\u8457\u63d0\u9ad8\u541e\u5410\u91cf\u3002<\/p>\n\n\n\n<p>\u4e0d\u8fc7\uff0c\u771f\u6b63\u91cd\u8981\u7684\u95ee\u9898\u662f\uff0c\u4f4e\u5ef6\u8fdf\u5c0f\u578b\u6a21\u578b\u63a8\u7406\u672c\u8eab\u5e02\u573a\u662f\u5426\u8db3\u591f\u5927\uff0c\u5982\u679c\u662f\uff0c\u90a3\u4e48\u5728\u7075\u6d3b\u7684GPU \u57fa\u7840\u8bbe\u65bd\u6210\u672c\u63a5\u8fd1\u76f8\u540c\uff0c\u5e76\u4e14\u53ef\u4ee5\u76f8\u5f53\u5bb9\u6613\u5730\u91cd\u65b0\u90e8\u7f72\u7528\u4e8e\u541e\u5410\u91cf\u6216\u5927\u578b\u6a21\u578b\u5e94\u7528\u7684\u60c5\u51b5\u4e0b\uff0c\u662f\u5426\u503c\u5f97\u62e5\u6709\u4e13\u95e8\u7684\u57fa\u7840\u8bbe\u65bd\u3002<\/p>\n\n\n\n<p><em>*\u4ee5\u4e0a\u6587\u7ae0\u7ffb\u8bd1\u81ea\u300aGroq Inference Tokenomics: Speed, But At What Cost?\u300b\uff0c\u5982\u9700\u539f\u6587\uff0c\u8bf7\u4e0e\u6211\u4eec\u8054\u7cfb\u3002<\/em><\/p>\n\n\n\n<p>WF Research \u662f\u4ee5\u7b2c\u4e00\u6027\u539f\u7406\u4e3a\u57fa\u7840\u7684\u4e13\u4e1a\u987e\u95ee\u670d\u52a1\u673a\u6784\uff0c\u6b22\u8fce\u5173\u6ce8\u548c\u7559\u8a00\uff01<\/p>\n\n\n\n<p>\u5fae\u4fe1\uff1aAlexqjl<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8fd1\u671f\uff0cAI \u786c\u4ef6\u521d\u521b\u516c\u53f8 Groq \u5f15\u8d77\u4e86\u5e7f\u6cdb\u5173\u6ce8\uff0c\u4ed6\u4eec\u5728\u63a8\u7406 API \u4e0a\u6240\u5c55\u793a\u7684\u9886\u5148\u5f00\u6e90\u6a21\u578b Mistral [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":2803,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"_links":{"self":[{"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/posts\/2791"}],"collection":[{"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2791"}],"version-history":[{"count":1,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/posts\/2791\/revisions"}],"predecessor-version":[{"id":2804,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/posts\/2791\/revisions\/2804"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/media\/2803"}],"wp:attachment":[{"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2791"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2791"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2791"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}