{"id":34257,"date":"2024-12-03T05:57:50","date_gmt":"2024-12-02T21:57:50","guid":{"rendered":"https:\/\/17aitech.com\/?p=34257"},"modified":"2024-12-03T05:57:50","modified_gmt":"2024-12-02T21:57:50","slug":"%e5%8f%b2%e4%b8%8a%e6%9c%80%e5%bf%abai%e8%8a%af%e7%89%87%e3%80%8csohu%e3%80%8d%ef%bc%8c%e9%80%9f%e5%ba%a610%e5%80%8d%e4%ba%8eb200%ef%bc%8c%e5%93%88%e4%bd%9b%e8%be%8d%e5%ad%a6%e7%94%9f%e6%89%93","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34257","title":{"rendered":"\u53f2\u4e0a\u6700\u5febAI\u82af\u7247\u300cSohu\u300d\uff0c\u901f\u5ea610\u500d\u4e8eB200\uff0c\u54c8\u4f5b\u8f8d\u5b66\u751f\u6253\u9020"},"content":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u53f2\u4e0a\u6700\u5febAI\u82af\u7247\u300cSohu\u300d\uff0c\u901f\u5ea610\u500d\u4e8eB200\uff0c\u54c8\u4f5b\u8f8d\u5b66\u751f\u6253\u9020<\/a><\/p>\n

\n
\n
\n

\u751f\u6210\u5f0f AI \u63a8\u7406\u6027\u4ef7\u6bd4\u662f GPU \u7684 140 \u500d\u3002<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n

\u5927\u6a21\u578b\u65f6\u4ee3\uff0c\u5168\u7403\u90fd\u7f3a\u7b97\u529b\uff0c\u4e70\u94f2\u5b50\u7684\u82f1\u4f1f\u8fbe\u5e02\u503c\u88ab\u7092\u4e0a\u4e86\u5929\u3002<\/p>\n

\u73b0\u5728\uff0c\u7ec8\u4e8e\u6709\u4e00\u5bb6\u516c\u53f8\u5e26\u7740\u81ea\u5df1\u7684 AI \u82af\u7247\u6765\u53eb\u677f\u4e86\u3002<\/p>\n

\u4eca\u5929\u51cc\u6668\uff0c\u79d1\u6280\u5708\u8fce\u6765\u4e86\u4e00\u4e2a\u91cd\u8981\u65b0\u95fb\u3002\u7f8e\u56fd\u82af\u7247\u521b\u4e1a\u516c\u53f8 Etched \u63a8\u51fa\u4e86\u81ea\u5df1\u7684\u7b2c\u4e00\u5757 AI \u82af\u7247 Sohu\uff0c\u5b83\u8fd0\u884c\u5927\u6a21\u578b\u7684\u901f\u5ea6\u6bd4\u82f1\u4f1f\u8fbe H100 \u8981\u5feb 20 \u500d\uff0c\u6bd4\u4eca\u5e74 3 \u6708\u624d\u63a8\u51fa\u7684\u9876\u914d\u82af\u7247 B200 \u4e5f\u8981\u5feb\u4e0a\u8d85\u8fc7 10 \u500d\u3002<\/p>\n

<\/a><\/p>\n

\u4e00\u53f0 Sohu \u7684\u670d\u52a1\u5668\u8fd0\u884c Llama 70B \u6bcf\u79d2\u53ef\u8f93\u51fa\u8d85\u8fc7 50 \u4e07\u4e2a token\uff0c\u6bd4 H100 \u670d\u52a1\u5668\uff0823,000 \u4e2a token \/ \u79d2\uff09\u591a 20 \u500d\uff0c\u6bd4 B200 \u670d\u52a1\u5668\uff08\u7ea6 45,000 \u4e2a token \/ \u79d2\uff09\u591a 10 \u500d\u3002<\/sup><\/em><\/a><\/p>\n

Sohu \u662f\u4e16\u754c\u7b2c\u4e00\u6b3e\u4e13\u7528\u4e8e Transformer \u8ba1\u7b97\u7684\u82af\u7247\uff0c\u5386\u65f6\u4e24\u5e74\u6253\u9020\u3002<\/p>\n

\u4f5c\u4e3a\u4e00\u5757 ASIC\uff08\u4e13\u7528\u96c6\u6210\u7535\u8def<\/mark>\uff09\uff0cSohu \u628a\u5bf9\u4e8e transformer \u67b6\u6784\u7684\u4f18\u5316\u786c\u5316\u5728\u82af\u7247\u4e2d\uff0c\u65e0\u6cd5\u8fd0\u884c\u5927\u591a\u6570\u300c\u4f20\u7edf\u300d\u7684 AI \u6a21\u578b\uff1a\u5982\u4e3a Instagram \u5e7f\u544a\u63d0\u4f9b\u652f\u6301\u7684 DLRM\u3001AlphaFold<\/mark> 2 \u7b49\u86cb\u767d\u8d28\u6298\u53e0\u6a21\u578b\u6216 Stable Diffusion 2 \u7b49\u8f83\u65e7\u7684\u56fe\u50cf\u751f\u6210<\/mark>\u6a21\u578b\u3002\u6211\u4eec\u4e5f\u65e0\u6cd5\u8fd0\u884c CNN\u3001RNN \u6216 LSTM\u3002<\/p>\n

\u4f46\u53e6\u4e00\u65b9\u9762\uff0c\u5bf9\u4e8e transformer \u6765\u8bf4\uff0cSohu \u5c31\u662f\u6709\u53f2\u4ee5\u6765\u6700\u5feb\u7684\u82af\u7247\uff0c\u4e0e\u5176\u4ed6\u4ea7\u54c1\u4e4b\u95f4\u662f\u91cf\u7ea7\u7684\u533a\u522b\u3002\u5982\u4eca\u7684\u6bcf\u6b3e\u4e3b\u6d41 AI \u4ea7\u54c1\u5982 ChatGPT\u3001Claude\u3001Gemini \u548c Sora \u90fd\u662f\u7531 transformer \u9a71\u52a8\u7684\u3002<\/p>\n

\u6700\u8fd1\u4e00\u6bb5\u65f6\u95f4\uff0c\u7531\u4e8e\u6469\u5c14\u5b9a\u5f8b<\/mark>\u653e\u7f13\uff0cGPU \u6027\u80fd\u7684\u63d0\u5347\u5f88\u5927\u7a0b\u5ea6\u4e0a\u9700\u8981\u4f9d\u8d56\u4e8e\u589e\u52a0\u82af\u7247\u9762\u79ef\u548c\u529f\u8017\u3002\u4e0d\u8bba\u662f\u82f1\u4f1f\u8fbe B200\u3001AMD<\/mark> MI300X \u8fd8\u662f Intel Gaudi 3\uff0c\u90fd\u4e0d\u7ea6\u800c\u540c\u7684\u4f7f\u7528\u300c\u4e8c\u5408\u4e00\u300d\u7684\u65b9\u5f0f\u63d0\u5347\u6027\u80fd\uff0c\u529f\u8017\u4e5f\u7ffb\u500d\u4e86\u3002<\/p>\n

<\/a><\/section>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u4ece 2022 \u5e74\u5230 2025 \u5e74\uff0cAI \u82af\u7247\u7684\u6027\u80fd\u5e76\u6ca1\u6709\u771f\u6b63\u53d8\u5f97\u66f4\u597d\uff0c\u800c\u662f\u53d8\u5f97\u66f4\u5927\u4e86\u3002<\/sup><\/em><\/section>\n

\u4f46\u5982\u679c\u5927\u6a21\u578b\u5e7f\u6cdb\u4f7f\u7528 Transformer \u67b6\u6784\uff0c\u8ffd\u6c42\u4e13\u4e1a\u5316\u6216\u8bb8\u662f\u63d0\u9ad8\u6027\u80fd\u7684\u597d\u65b9\u5411\u3002\u200d<\/p>\n

\u4f5c\u4e3a\u4e00\u4e2a\u65b0\u5174\u9886\u57df\uff0cAI \u6a21\u578b\u7684\u67b6\u6784\u8fc7\u53bb\u53d8\u5316\u5f88\u5927\u3002\u4f46\u81ea GPT-2<\/mark> \u4ee5\u6765\uff0c\u6700\u5148\u8fdb\u7684\u6a21\u578b\u51e0\u4e4e\u90fd\u5728\u4f7f\u7528 Transformer\uff0c\u4ece OpenAI \u7684 GPT \u7cfb\u5217\u3001\u8c37\u6b4c\u7684 PaLM\u3001Facebook \u7684 LLaMa\uff0c\u518d\u5230\u7279\u65af\u62c9 FSD \u81ea\u52a8\u9a7e\u9a76\u6240\u9700\u7684\u6a21\u578b\u3002<\/p>\n

<\/a><\/p>\n

Etched \u7ed9\u6211\u4eec\u7b97\u4e86\u4e00\u7b14\u8d26\uff1a\u82af\u7247\u9879\u76ee\u7684\u6210\u672c\u4e3a 5000 \u4e07\u81f3 1 \u4ebf\u7f8e\u5143\uff0c\u9700\u8981\u6570\u5e74\u65f6\u95f4\u624d\u80fd\u6295\u5165\u751f\u4ea7\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5f53\u6a21\u578b\u8bad\u7ec3\u6210\u672c\u8d85\u8fc7 10 \u4ebf\u7f8e\u5143\u3001\u63a8\u7406\u6210\u672c\u8d85\u8fc7 100 \u4ebf\u7f8e\u5143\u65f6\uff0c\u4f7f\u7528\u4e13\u7528\u82af\u7247\u662f\u4e0d\u53ef\u907f\u514d\u7684\u3002\u5728\u8fd9\u79cd\u4ea7\u4e1a\u89c4\u6a21\u4e0b\uff0c1% \u7684\u6539\u8fdb\u5c31\u80fd\u64ac\u52a8\u786c\u4ef6\u67b6\u6784\u7684\u66f4\u65b0\u3002<\/p>\n

\u901f\u5ea6\u8d85 H100 20 \u500d\uff0cFLOPS \u5229\u7528\u7387\u8d85 90%<\/strong><\/p>\n

\u4f5c\u4e3a\u4e16\u754c\u4e0a\u9996\u6b3e transformer ASIC\uff08\u5e94\u7528\u578b\u4e13\u7528\u96c6\u6210\u7535\u8def<\/mark>\uff09\u82af\u7247\uff0c\u4e00\u53f0\u96c6\u6210\u4e86 8 \u5757 Sohu \u7684\u670d\u52a1\u5668\u53ef\u4ee5\u5339\u654c 160 \u5757 H100 GPU\u3002\u4e5f\u5373\uff0cSohu \u7684\u8fd0\u884c\u901f\u5ea6\u662f H100 \u7684 20 \u591a\u500d\u3002<\/p>\n

\u5177\u4f53\u6765\u8bb2\uff0c\u901a\u8fc7\u4e13\u95e8\u5316\uff0cSohu \u5177\u5907\u4e86\u524d\u6240\u672a\u6709\u7684\u6027\u80fd\u3002\u4e00\u53f0\u96c6\u6210 8 \u5757 Sohu \u82af\u7247\u7684\u670d\u52a1\u5668\u6bcf\u79d2\u53ef\u4ee5\u5904\u7406 50 \u4e07\u7684 Llama 7B tokens\u3002<\/p>\n

\u9488\u5bf9 Llama 3 70B \u7684 FP8 \u7cbe\u5ea6\u57fa\u51c6<\/mark>\u6d4b\u8bd5\u663e\u793a\uff1a\u65e0\u7a00\u758f\u6027\u30018 \u500d\u6a21\u578b\u5e76\u884c\u30012048 \u8f93\u5165\u6216 128 \u8f93\u51fa\u957f\u5ea6\u3002<\/p>\n

\u6b64\u5916\uff0c\u5bf9\u4e8e Llama\u3001Stable Diffusion 3\uff0cSohu \u4ec5\u652f\u6301 transformer \u63a8\u7406\u3002Sohu \u652f\u6301\u4e86\u5f53\u524d\u8c37\u6b4c\u3001Meta\u3001\u5fae\u8f6f\u3001OpenAI\u3001Anthropic \u7b49\u5404\u5bb6\u7684\u6a21\u578b\uff0c\u672a\u6765\u8fd8\u4f1a\u9002\u914d\u6a21\u578b\u8c03\u6574\u3002<\/p>\n

\u7531\u4e8e Sohu \u4ec5\u80fd\u8fd0\u884c\u4e00\u79cd\u7b97\u6cd5\uff0c\u56e0\u6b64\u53ef\u4ee5\u5220\u9664\u7edd\u5927\u591a\u6570\u63a7\u5236\u6d41\u903b\u8f91<\/mark>\uff0c\u4ece\u800c\u5141\u8bb8\u62e5\u6709\u66f4\u591a\u6570\u5b66\u5757\u3002\u4e5f\u56e0\u6b64\uff0cSohu \u5b9e\u73b0\u4e86 90% \u4ee5\u4e0a\u7684 FLOPS \u5229\u7528\u7387\uff0c\u800c\u4f7f\u7528 TRT-LLM \u7684 GPU \u7ea6\u4e3a 30%\u3002<\/p>\n

Sohu \u4e3a\u4f55\u80fd\u8f93\u51fa\u66f4\u591a FLOPS\uff1f<\/strong><\/p>\n

\u82f1\u4f1f\u8fbe H200 \u652f\u6301 989 TFLOPS \u7684 FP16\/BF16 \u8ba1\u7b97\u80fd\u529b\uff0c\u5e76\u4e14\u6ca1\u6709\u7a00\u758f\u6027\u3002\u8fd9\u662f\u5f53\u524d\u6700\u5148\u8fdb\u7684\u82af\u7247\uff0c\u800c 2025 \u5e74\u63a8\u51fa\u7684 GB200 \u5c06\u5728\u8ba1\u7b97\u80fd\u529b\u4e0a\u63d0\u5347 25%\uff0c\u652f\u6301 1250 TFLOPS\u3002<\/p>\n

\u7531\u4e8e GPU \u7684\u7edd\u5927\u90e8\u5206\u533a\u57df\u90fd\u662f\u53ef\u7f16\u7a0b\u7684\uff0c\u56e0\u6b64\u4e13\u6ce8\u4e8e transformer \u4f1a\u5bb9\u7eb3\u66f4\u591a\u7684\u8ba1\u7b97\u3002\u8fd9\u53ef\u4ee5\u4ece\u7b2c\u4e00\u6027\u539f\u7406\u4e2d\u8bc1\u660e\uff1a<\/p>\n

\u6784\u5efa\u5355\u4e2a FP16\/BF16\/FP8 \u4e58\u52a0\u7535\u8def\u9700\u8981 10000 \u4e2a\u6676\u4f53\u7ba1\uff0c\u8fd9\u662f\u6240\u6709\u77e9\u9635\u6570\u5b66\u7684\u57fa\u77f3\u3002H100 SXM \u62e5\u6709 528 \u4e2a\u5f20\u91cf<\/mark>\u6838\u5fc3\uff0c\u6bcf\u4e2a\u6838\u5fc3\u62e5\u6709 4\u00d7 8 \u00d7 16 FMA \u7535\u8def\u3002\u4e58\u6cd5\u544a\u8bc9\u6211\u4eec\uff1aH100 \u6709 27 \u4ebf\u4e2a\u6676\u4f53\u7ba1\u7528\u4e8e\u5f20\u91cf<\/mark>\u6838\u5fc3\u3002<\/p>\n

\u4f46\u662f\uff0cH100 \u5374\u6709 800 \u4ebf\u4e2a\u6676\u4f53\u7ba1\u3002\u8fd9\u610f\u5473\u7740 H100 GPU \u4e0a\u53ea\u6709 3.3% \u7684\u6676\u4f53\u7ba1\u7528\u4e8e\u77e9\u9635\u4e58\u6cd5\u3002\u8fd9\u662f\u82f1\u4f1f\u8fbe\u548c\u5176\u4ed6\u82af\u7247\u5382\u5546\u7ecf\u8fc7\u6df1\u601d\u719f\u8651\u7684\u8bbe\u8ba1\u51b3\u5b9a\u3002\u5982\u679c\u4f60\u60f3\u652f\u6301\u6240\u6709\u7c7b\u578b\u7684\u6a21\u578b\uff08CNN\u3001LSTM\u3001SSM \u7b49\uff09\uff0c\u90a3\u4e48\u6ca1\u6709\u6bd4\u8fd9\u66f4\u597d\u7684\u4e86\u3002<\/p>\n

\u800c\u901a\u8fc7\u4ec5\u8fd0\u884c transformer\uff0cEtched \u53ef\u4ee5\u8ba9 Sohu \u82af\u7247\u8f93\u51fa\u66f4\u591a\u7684 FLOPS\uff0c\u4e14\u9700\u8981\u964d\u4f4e\u7cbe\u5ea6\u6216\u7a00\u758f\u6027\u3002<\/p>\n

\u5185\u5b58\u5e26\u5bbd\u4e5f\u4e0d\u662f\u74f6\u9888<\/strong><\/p>\n

\u5b9e\u9645\u4e0a\uff0c\u5bf9\u4e8e\u50cf Llama 3 \u8fd9\u6837\u7684\u6a21\u578b\uff0c\u60c5\u51b5\u5e76\u975e\u5982\u6b64\u3002<\/p>\n

\u6211\u4eec\u4ee5\u82f1\u4f1f\u8fbe\u548c AMD<\/mark> \u7684\u6807\u51c6\u57fa\u51c6<\/mark>\u4e3a\u4f8b\uff1a2048 \u4e2a\u8f93\u5165 token \u548c 128 \u4e2a\u8f93\u51fa token\u3002\u5927\u591a\u6570 AI \u4ea7\u54c1\u7684 prompt \u66f4\u957f\uff0c\u6bd4\u5982\u6700\u65b0\u7684 Claude \u804a\u5929\u673a\u5668\u4eba<\/mark>\u5728\u7cfb\u7edf prompt \u4e2d\u62e5\u6709 1000+tokens\u3002<\/p>\n

\u5728 Sohu \u4e0a\uff0c\u63a8\u7406\u662f\u5206 batch \u8fd0\u884c\u7684\u3002\u6bcf\u4e2a batch \u90fd\u9700\u8981\u52a0\u8f7d\u6240\u6709\u6a21\u578b\u6743\u91cd<\/mark>\u4e00\u6b21\uff0c\u5e76\u5728 batch \u7684\u6bcf\u4e2a token \u4e2d\u91cd\u590d\u4f7f\u7528\u3002\u901a\u5e38\u6765\u8bf4\uff0cLLM \u8f93\u5165\u662f\u8ba1\u7b97\u5bc6\u96c6\u578b\u7684\uff0c\u800c LLM \u8f93\u51fa\u662f\u5185\u5b58\u5bc6\u96c6\u578b\u7684\u3002\u5f53\u6211\u4eec\u5c06\u8f93\u5165\u548c\u8f93\u51fa token \u4e0e\u8fde\u7eed batch \u7ed3\u5408\u65f6\uff0c\u5de5\u4f5c\u8d1f\u8f7d\u53d8\u6210\u4e86\u9ad8\u5ea6\u8ba1\u7b97\u5bc6\u96c6\u578b\u3002<\/p>\n

\u4ee5\u4e0b\u4e3a LLM \u8fde\u7eed batching \u5904\u7406\u7684\u793a\u4f8b\uff0c\u8fd9\u91cc\u8fd0\u884c\u5177\u6709\u56db\u4e2a\u8f93\u5165 token \u548c\u56db\u4e2a\u8f93\u51fa token \u7684\u5e8f\u5217\u3002\u6bcf\u79cd\u989c\u8272\u4ee3\u8868\u4e0d\u540c\u7684\u5e8f\u5217\u3002<\/p>\n

<\/a><\/p>\n

\u6211\u4eec\u53ef\u4ee5\u6269\u5c55\u76f8\u540c\u7684\u6280\u5de7\uff0c\u4ece\u800c\u8fd0\u884c\u5177\u6709 2048 \u4e2a\u8f93\u5165 token \u548c 128 \u4e2a\u8f93\u51fa token \u7684 Llama 3 70B\u3002\u6bcf\u4e2a batch \u4e2d\u5305\u542b\u7528\u4e8e\u4e00\u4e2a\u5e8f\u5217\u7684 2048 \u4e2a\u8f93\u5165 token\uff0c\u4ee5\u53ca\u7528\u4e8e 127 \u4e2a\u4e0d\u540c\u5e8f\u5217\u7684 127 \u4e2a\u8f93\u51fa token\u3002<\/p>\n

\u5982\u679c\u8fd9\u6837\u505a\u4e86\uff0c\u5219\u6bcf\u4e2a batch \u9700\u8981\u5927\u7ea6 (2048 + 127)\u00d770B \u53c2\u6570<\/mark> \u00d7 \u6bcf\u4e2a\u53c2\u6570<\/mark> 2 \u5b57\u8282 = 304 TFLOP\uff0c\u800c\u4ec5\u9700\u8981\u52a0\u8f7d 70B \u53c2\u6570<\/mark> \u00d7 \u6bcf\u4e2a\u53c2\u6570<\/mark> 2 \u5b57\u8282 = 140 GB \u7684\u6a21\u578b\u6743\u91cd<\/mark>\u548c\u5927\u7ea6 127\u00d7 64 \u00d7 8 \u00d7 128 \u00d7 (2048 + 127) \u00d7 2 \u00d7 2 = 72GB \u7684 KV \u7f13\u5b58\u6743\u91cd<\/mark>\u3002\u8fd9\u6bd4\u5185\u5b58\u5e26\u5bbd\u9700\u8981\u7684\u8ba1\u7b97\u91cf\u591a\u5f97\u591a\uff1aH200 \u9700\u8981 6.8 PFLOPS \u7684\u8ba1\u7b97\u624d\u80fd\u6700\u5927\u5316\u5176\u5185\u5b58\u5e26\u5bbd\u3002\u8fd9\u8fd8\u662f\u5229\u7528\u7387\u4e3a 100% \u7684\u60c5\u51b5\uff0c\u5982\u679c\u5229\u7528\u7387\u4ec5\u4e3a 30%\uff0c\u5219\u9700\u8981 3 \u500d\u4ee5\u4e0a\u7684\u5185\u5b58\u3002<\/p>\n

Sohu \u62e5\u6709\u4e86\u66f4\u591a\u7684\u8ba1\u7b97\u80fd\u529b\u4e14\u5229\u7528\u7387\u975e\u5e38\u9ad8\uff0c \u56e0\u6b64\u53ef\u4ee5\u8fd0\u884c\u5de8\u5927\u7684\u541e\u5410\u91cf\uff0c\u800c\u4e0d\u4f1a\u51fa\u73b0\u5185\u5b58\u5e26\u5bbd\u74f6\u9888\u3002<\/p>\n

\u8f6f\u4ef6\u5982\u4f55\u5de5\u4f5c<\/strong><\/p>\n

\u5728 GPU \u548c TPU \u4e0a\uff0c\u8f6f\u4ef6\u662f\u4e00\u573a\u5669\u68a6\u3002\u5904\u7406\u4efb\u610f CUDA \u548c PyTorch \u4ee3\u7801\u9700\u8981\u6781\u5176\u590d\u6742\u7684\u7f16\u8bd1\u5668\u3002\u7b2c\u4e09\u65b9 AI \u82af\u7247\uff08\u5982 AMD<\/mark>\u3001Intel\u3001AWS \u7b49\uff09\u5728\u8f6f\u4ef6\u4e0a\u603b\u5171\u82b1\u8d39\u4e86\u6570\u5341\u4ebf\u7f8e\u5143\uff0c\u4f46\u6536\u6548\u751a\u5fae\u3002<\/p>\n

\u800c Sohu \u53ea\u8fd0\u884c transformer\uff0c\u56e0\u6b64\u53ea\u9700\u8981\u4e3a transformer \u7f16\u5199\u8f6f\u4ef6\u3002<\/p>\n

\u5927\u591a\u6570\u8fd0\u884c\u5f00\u6e90\u6216\u5185\u90e8\u6a21\u578b\u7684\u516c\u53f8\u90fd\u4f7f\u7528\u7279\u5b9a\u4e8e transformer \u7684\u63a8\u7406\u5e93\uff0c\u6bd4\u5982 TensorRT-LLM\u3001vLLM \u6216 HuggingFace \u7684 TGI\u3002<\/p>\n

\u8fd9\u4e9b\u6846\u67b6\u975e\u5e38\u50f5\u5316\uff0c\u867d\u7136\u4f60\u53ef\u4ee5\u8fdb\u884c\u6a21\u578b\u8d85\u53c2\u6570<\/mark>\u8c03\u4f18\uff0c\u4f46\u5b9e\u9645\u4e0a\u4e0d\u652f\u6301\u66f4\u6539\u5e95\u5c42\u6a21\u578b\u4ee3\u7801\u3002\u4f46\u8fd9\u6ca1\u5173\u7cfb\uff0c\u56e0\u4e3a\u6240\u6709 transformer \u6a21\u578b\u90fd\u975e\u5e38\u76f8\u4f3c\uff08\u751a\u81f3\u662f\u6587\u672c \/ \u56fe\u50cf \/ \u89c6\u9891\u6a21\u578b\uff09\uff0c\u6240\u4ee5\u8d85\u53c2\u6570<\/mark>\u8c03\u4f18\u5c31\u662f\u4f60\u771f\u6b63\u9700\u8981\u7684\u3002<\/p>\n

\u867d\u7136 95% \u7684 AI \u516c\u53f8\u662f\u8fd9\u6837\uff0c\u4f46\u4e00\u4e9b\u6700\u5927\u7684 AI \u5b9e\u9a8c\u5ba4\u91c7\u7528\u5b9a\u5236\u65b9\u5f0f\u3002\u4ed6\u4eec\u6709\u5de5\u7a0b\u5e08\u56e2\u961f\u6765\u624b\u52a8\u8c03\u6574 GPU \u6838\u5fc3\u4ee5\u5b9e\u73b0\u66f4\u9ad8\u7684\u5229\u7528\u7387\uff0c\u5e76\u8fdb\u884c\u9006\u5411\u5de5\u7a0b<\/mark>\u4ee5\u5c06\u5bc4\u5b58\u5668\u5bf9\u6bcf\u4e2a\u5f20\u91cf<\/mark>\u6838\u5fc3\u7684\u5ef6\u8fdf\u5c06\u81f3\u6700\u4f4e\u3002<\/p>\n

Etched \u8ba9\u6211\u4eec\u4e0d\u9700\u8981\u518d\u8fdb\u884c\u9006\u5411\u5de5\u7a0b<\/mark>\uff0c\u4ed6\u4eec\u7684\u8f6f\u4ef6\uff08\u4ece\u9a71\u52a8\u7a0b\u5e8f\u3001\u5185\u6838\u5230\u670d\u52a1\u5806\u6808\uff09\u90fd\u5c06\u662f\u5f00\u6e90\u7684\u3002\u5982\u679c\u4f60\u60f3\u5b9e\u73b0\u81ea\u5b9a\u4e49 transformer \u5c42\uff0c\u5219\u5185\u6838\u5411\u5bfc\u53ef\u4ee5\u81ea\u7531\u5730\u8fd9\u6837\u505a\u3002<\/p>\n

\u521b\u4e1a\u56e2\u961f\uff1a\u54c8\u4f5b\u8f8d\u5b66\u751f\u9886\u8854<\/strong><\/p>\n

Etched \u7684 CEO Gavin Uberti \u544a\u8bc9\u8bb0\u8005\uff1a\u300c\u5982\u679c\u672a\u6765 Transformer \u4e0d\u518d\u662f\u4e3b\u6d41\uff0c\u90a3\u6211\u4eec\u5c31\u4f1a\u706d\u4ea1\u3002\u4f46\u5982\u679c\u5b83\u7ee7\u7eed\u5b58\u5728\uff0c\u6211\u4eec\u5c31\u4f1a\u6210\u4e3a\u6709\u53f2\u4ee5\u6765\u6700\u5927\u7684\u516c\u53f8\u3002\u300d<\/p>\n

\u6253\u9020 Sohu \u82af\u7247\u7684 Etched \u4f4d\u4e8e\u52a0\u5dde\u5e93\u6bd4\u8482\u8bfa\uff0c\u516c\u53f8\u6210\u7acb\u4ec5\u4e24\u5e74\uff0c\u76ee\u524d\u56e2\u961f\u53ea\u6709 35 \u4eba\uff0c\u521b\u59cb\u4eba\u662f\u4e00\u5bf9\u54c8\u4f5b\u8f8d\u5b66\u751f Gavin Uberti\uff08\u524d OctoML \u548c\u524d Xnor.ai \u5458\u5de5\uff09\u548c Chris Zhu\uff0c\u4ed6\u4eec\u4e0e Robert Wachen \u548c\u524d\u8d5b\u666e\u62c9\u65af\u534a\u5bfc\u4f53\u516c\u53f8\u9996\u5e2d\u6280\u672f\u5b98 Mark Ross \u4e00\u8d77\uff0c\u4e00\u76f4\u81f4\u529b\u4e8e\u6253\u9020\u4e13\u7528\u4e8e AI \u5927\u6a21\u578b\u7684\u82af\u7247\u3002<\/p>\n

<\/a><\/p>\n

\u5728 Sohu \u82af\u7247\u53d1\u5e03\u7684\u540c\u65f6\uff0cEtched \u4e5f\u5ba3\u5e03\u5df2\u5b8c\u6210\u4e86 1.2 \u4ebf\u7f8e\u5143\u7684 A \u8f6e\u878d\u8d44\uff0c\u7531 Primary Venture Partners \u548c Positive Sum Ventures \u5171\u540c\u9886\u6295\u3002Etched \u7684\u603b\u878d\u8d44\u989d\u5df2\u8fbe\u5230 1.2536 \u4ebf\u7f8e\u5143\uff0c\u672c\u8f6e\u878d\u8d44\u7684\u91cd\u8981\u6295\u8d44\u8005\u5305\u62ec Peter Thiel\u3001GitHub \u9996\u5e2d\u6267\u884c\u5b98 Thomas Dohmke\u3001Cruise \u8054\u5408\u521b\u59cb\u4eba Kyle Vogt \u548c Quora \u8054\u5408\u521b\u59cb\u4eba Charlie Cheever\u3002<\/p>\n

\u4e0d\u8fc7\u5bf9\u4e8e\u5360\u636e\u8d85\u8fc7 80% AI \u82af\u7247\u5e02\u573a\u4efd\u989d\u7684\u82f1\u4f1f\u8fbe\u6765\u8bf4\uff0c1.2 \u4ebf\u7f8e\u5143\u53ea\u76f8\u5f53\u4e8e\u5b83\u534a\u5929\u7684\u6536\u5165\u3002<\/p>\n

\u300c\u6211\u4eec\u5982\u6b64\u5174\u594b\u7684\u539f\u56e0\uff0c\u9009\u62e9\u8f8d\u5b66\u7684\u539f\u56e0\uff0c\u4ee5\u53ca\u6211\u4eec\u53ec\u96c6\u56e2\u961f\uff0c\u6295\u8eab\u82af\u7247\u9879\u76ee\u7684\u539f\u56e0\u5728\u4e8e \u2014\u2014 \u8fd9\u662f\u6700\u91cd\u8981\u7684\u5de5\u4f5c\uff0c\u300dEtched \u8fd0\u8425\u4e3b\u7ba1 Robert Wachen \u8bf4\u9053\u3002\u300c\u6574\u4e2a\u6280\u672f\u7684\u672a\u6765\u5c06\u53d6\u51b3\u4e8e\u7b97\u529b\u57fa\u7840\u8bbe\u65bd\u80fd\u5426\u5b9e\u73b0\u5927\u89c4\u6a21\u3002\u300d<\/p>\n

Uberti \u58f0\u79f0\u5230\u76ee\u524d\u4e3a\u6b62\uff0c\u5df2\u6709\u533f\u540d\u5ba2\u6237\u9884\u8ba2\u4e86\u300c\u6570\u5343\u4e07\u7f8e\u5143\u300d\u7684\u786c\u4ef6\uff0c\u9884\u8ba1\u5728\u4eca\u5e74\u4e09\u5b63\u5ea6\uff0cSohu \u5c06\u63a8\u5411\u5e02\u573a\u3002<\/p>\n

\u672a\u6765\u771f\u7684\u5982 Uberti \u6240\u8bf4\uff0c\u53ea\u6709\u5728 Sohu \u8fd9\u6837\u7684\u82af\u7247\u4e0a\uff0c\u89c6\u9891\u751f\u6210<\/mark>\u3001\u97f3\u9891\u751f\u6210\u3001\u5177\u8eab\u667a\u80fd\u7b49\u6280\u672f\u624d\u80fd\u771f\u6b63\u843d\u5730\u5417\uff1f<\/p>\n

\u53c2\u8003\u5185\u5bb9\uff1a<\/sup><\/em><\/p>\n

https:\/\/www.etched.com\/announcing-etched<\/em><\/sup><\/p>\n

https:\/\/twitter.com\/Etched\/status\/1805625693113663834<\/em><\/sup><\/p>\n

https:\/\/www.cnbc.com\/2024\/06\/25\/etched-raises-120-million-to-build-chip-to-take-on-nvidia-in-ai.html<\/em><\/sup><\/p>\n

https:\/\/techcrunch.com\/2024\/06\/25\/etched-is-building-an-ai-chip-that-only-runs-transformer-models\/<\/sup><\/em><\/p>\n

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u53f2\u4e0a\u6700\u5febAI\u82af\u7247\u300cSohu\u300d\uff0c\u901f\u5ea610\u500d\u4e8eB200\uff0c\u54c8\u4f5b\u8f8d\u5b66\u751f\u6253\u9020<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u53f2\u4e0a\u6700\u5febAI\u82af\u7247\u300cSo […]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68],"class_list":["post-34257","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"\n\u53f2\u4e0a\u6700\u5febAI\u82af\u7247\u300cSohu\u300d\uff0c\u901f\u5ea610\u500d\u4e8eB200\uff0c\u54c8\u4f5b\u8f8d\u5b66\u751f\u6253\u9020 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=34257\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=34257\",\"url\":\"https:\/\/17aitech.com\/?p=34257\",\"name\":\"\u53f2\u4e0a\u6700\u5febAI\u82af\u7247\u300cSohu\u300d\uff0c\u901f\u5ea610\u500d\u4e8eB200\uff0c\u54c8\u4f5b\u8f8d\u5b66\u751f\u6253\u9020 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=34257#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=34257#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-5d780cf0805eda687e76e0b0130290e5.png\",\"datePublished\":\"2024-12-02T21:57:50+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=34257#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=34257\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=34257#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-5d780cf0805eda687e76e0b0130290e5.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-5d780cf0805eda687e76e0b0130290e5.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=34257#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u53f2\u4e0a\u6700\u5febAI\u82af\u7247\u300cSohu\u300d\uff0c\u901f\u5ea610\u500d\u4e8eB200\uff0c\u54c8\u4f5b\u8f8d\u5b66\u751f\u6253\u9020\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u53f2\u4e0a\u6700\u5febAI\u82af\u7247\u300cSohu\u300d\uff0c\u901f\u5ea610\u500d\u4e8eB200\uff0c\u54c8\u4f5b\u8f8d\u5b66\u751f\u6253\u9020 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=34257","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34257","url":"https:\/\/17aitech.com\/?p=34257","name":"\u53f2\u4e0a\u6700\u5febAI\u82af\u7247\u300cSohu\u300d\uff0c\u901f\u5ea610\u500d\u4e8eB200\uff0c\u54c8\u4f5b\u8f8d\u5b66\u751f\u6253\u9020 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34257#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34257#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-5d780cf0805eda687e76e0b0130290e5.png","datePublished":"2024-12-02T21:57:50+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34257#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34257"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34257#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-5d780cf0805eda687e76e0b0130290e5.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/06\/frc-5d780cf0805eda687e76e0b0130290e5.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34257#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u53f2\u4e0a\u6700\u5febAI\u82af\u7247\u300cSohu\u300d\uff0c\u901f\u5ea610\u500d\u4e8eB200\uff0c\u54c8\u4f5b\u8f8d\u5b66\u751f\u6253\u9020"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34257","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=34257"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34257\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34257"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34257"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34257"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}