<script type="application/ld+json">{"@context":"http://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://www.simcentric.com/tc/"},{"@type":"ListItem","position":2,"name":"度量大模型推理效能——解析 2026 年的 Token 生成效率","item":"https://www.simcentric.com/tc/japan-dedicated-server-tc/measuring-llm-performance-analyzing-token-efficiency-in-2026/"}]}</script> {"id":34108,"date":"2026-08-17T12:18:17","date_gmt":"2026-08-17T04:18:17","guid":{"rendered":"https:\/\/www.simcentric.com\/uncategorized-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/"},"modified":"2026-08-17T12:20:49","modified_gmt":"2026-08-17T04:20:49","slug":"measuring-llm-performance-analyzing-token-efficiency-in-2026","status":"publish","type":"post","link":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/","title":{"rendered":"\u5ea6\u91cf\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd\u2014\u2014\u89e3\u6790 2026 \u5e74\u7684 Token \u751f\u6210\u6548\u7387"},"content":{"rendered":"<p>\u5728\u751f\u6210\u5f0f AI \u90e8\u7f72\u4e2d\uff0c\u4f60\u900f\u904e\u4e00\u7cfb\u5217\u95dc\u9375\u6307\u6a19\u4f86\u8861\u91cf<a href=\"https:\/\/www.simcentric.com\/tc\/america-dedicated-server-tc\/rtx-4090-ai-inference-hosting-concurrency\/\" target=\"_self\">\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd<\/a>\u3002Time to first token\uff08TTFT\uff0c\u9996 Token \u5ef6\u9072\uff09\u7528\u65bc\u8ffd\u8e64\u63d0\u793a\u8655\u7406\u6642\u5ef6\uff1btime per output token\uff08TPOT\uff0c\u55ae\u4e00 Token \u751f\u6210\u6642\u9593\uff09\u548c inter-token latency\uff08Token \u9593\u5ef6\u9072\uff09\u7528\u65bc\u8861\u91cf\u8f38\u51fa\u751f\u6210\u901f\u5ea6\uff1btokens per second\uff08\u6bcf\u79d2 Token \u6578\uff09\u53cd\u6620\u7cfb\u7d71\u541e\u5410\u91cf\uff0c\u800c Goodput \u5247\u7d71\u8a08\u5728\u56b4\u683c\u670d\u52d9\u76ee\u6a19\u4e0b\u6210\u529f\u5b8c\u6210\u7684\u8acb\u6c42\u6578\u3002<\/p>\n<p>\u5728 2026 \u5e74\u7684\u751f\u7522\u74b0\u5883\u4e2d\uff0c\u4f8b\u5982\u90e8\u7f72\u5728<a href=\"https:\/\/www.simcentric.com\/tc\/products\/dedicated-server-jp\/\" target=\"_blank\" rel=\"noopener\">\u65e5\u672c\u5c08\u7528\u4f3a\u670d\u5668<\/a>\u53e2\u96c6\u4e0a\u7684\u5927\u898f\u6a21\u7cfb\u7d71\uff0c\u4f60\u900f\u904e\u786c\u9ad4\u4f7f\u7528\u7387\u5c0d\u71df\u904b\u6210\u672c\u7684\u58d3\u7e2e\u7a0b\u5ea6\u4f86\u5b9a\u7fa9 Token \u751f\u6210\u6548\u7387\u3002\u4e00\u689d\u6e05\u6670\u7684\u8a08\u7b97\u908a\u754c\u5c07\u5927\u8a9e\u8a00\u6a21\u578b\u7684\u57f7\u884c\u5283\u5206\u70ba\u5169\u500b\u968e\u6bb5\uff1a\u9810\u586b\u5145\uff08prefill\uff09\u968e\u6bb5\u5e73\u884c\u8655\u7406\u8f38\u5165 Token\uff0c\u5f62\u6210\u4ee5\u904b\u7b97\u70ba\u4e3b\u7684\u74f6\u9838\uff1b\u89e3\u78bc\uff08decode\uff09\u968e\u6bb5\u4f9d\u5e8f\u7522\u751f Token\uff0c\u5e36\u4f86\u4ee5\u8a18\u61b6\u9ad4\u983b\u5bec\u70ba\u4e3b\u7684\u9650\u5236\u3002\u5de5\u7a0b\u5e2b\u904b\u7528\u9019\u4e9b\u6548\u80fd\u6307\u6a19\uff0c\u5728\u5927\u6a21\u578b\u63a8\u7406\u512a\u5316\u904e\u7a0b\u4e2d\u5b9a\u4f4d\u7cfb\u7d71\u6975\u9650\u3002<\/p>\n<h2><strong>\u95dc\u9375\u91cd\u9ede\u7e3d\u89bd<\/strong><\/h2>\n<ul>\n<li>\u5927\u6a21\u578b\u8655\u7406\u5305\u542b\u5169\u500b\u968e\u6bb5\uff1a\u9810\u586b\u5145\u968e\u6bb5\u5feb\u901f\u8655\u7406\u63d0\u793a\u8a5e\uff0c\u89e3\u78bc\u968e\u6bb5\u5247\u76f8\u5c0d\u7de9\u6162\u5730\u751f\u6210\u6587\u5b57\u3002<\/li>\n<li>Goodput \u900f\u904e\u53ea\u7d71\u8a08\u7b26\u5408\u5ef6\u9072\u76ee\u6a19\u7684\u8f38\u51fa Token\uff0c\u4f86\u8861\u91cf\u7cfb\u7d71\u771f\u6b63\u7684\u6210\u529f\u7387\u3002<\/li>\n<li>\u9023\u7e8c\u6279\u6b21\u8655\u7406\u53ef\u4ee5\u6d88\u9664\u8a18\u61b6\u9ad4\u6d6a\u8cbb\uff0c\u5c07\u7cfb\u7d71\u901f\u5ea6\u63d0\u5347\u81f3\u539f\u4f86\u7684\u5927\u7d04 3 \u500d\u3002<\/li>\n<li>\u63a8\u6e2c\u89e3\u78bc\u904b\u7528\u5c0f\u6a21\u578b\u52a0\u901f\u8a5e\u5143\u751f\u6210\uff0c\u540c\u6642\u4e0d\u72a7\u7272\u8f38\u51fa\u54c1\u8cea\u3002<\/li>\n<li>\u73fe\u4ee3\u986f\u793a\u6676\u7247\u53ef\u4ee5\u964d\u4f4e\u71df\u904b\u6210\u672c\uff0c\u4e26\u9ad8\u6548\u57f7\u884c\u5927\u8a9e\u8a00\u6a21\u578b\u3002<\/li>\n<\/ul>\n<h2><strong>\u5927\u6a21\u578b Token \u751f\u6210\u6a5f\u5236\uff1a\u9810\u586b\u5145 vs \u89e3\u78bc<\/strong><\/h2>\n<p>\u4f60\u9700\u8981\u7406\u89e3\u5927\u8a9e\u8a00\u6a21\u578b\u5728\u8655\u7406\u4f7f\u7528\u8005\u8acb\u6c42\u6642\u7684\u5169\u500b\u4e0d\u540c\u8a08\u7b97\u6b65\u9a5f\uff1a\u9810\u586b\u5145\u968e\u6bb5\u8ca0\u8cac\u8655\u7406\u521d\u59cb\u63d0\u793a\u8a5e\uff0c\u89e3\u78bc\u968e\u6bb5\u8ca0\u8cac\u751f\u6210\u56de\u61c9\u6587\u5b57\u3002\u6bcf\u500b\u968e\u6bb5\u5728\u5927\u6a21\u578b\u57f7\u884c\u671f\u9593\u90fd\u6703\u5e36\u4f86\u7368\u7279\u7684\u904b\u884c\u9700\u6c42\u3002<\/p>\n<h3>\u9810\u586b\u5145\u968e\u6bb5\u8207\u904b\u7b97\u74f6\u9838<\/h3>\n<p>\u5728\u9810\u586b\u5145\u968e\u6bb5\uff0c\u6a21\u578b\u6703\u540c\u6642\u651d\u53d6\u5168\u90e8\u8f38\u5165 Token\u3002\u4f60\u900f\u904e\u91cf\u6e2c\u9019\u4e00\u968e\u6bb5\u4f86\u78ba\u5b9a\u90e8\u7f72\u4e2d\u7684\u9996 Token \u5ef6\u9072\uff08TTFT\uff09\u3002\u7576\u5927\u6a21\u578b\u6536\u5230\u4e00\u500b\u63d0\u793a\u8a5e\u6642\uff0c\u52a0\u901f\u5668\u6703\u5e73\u884c\u8a08\u7b97\u6240\u6709\u63d0\u793a Token \u914d\u5c0d\u4e4b\u9593\u7684\u6ce8\u610f\u529b\u5206\u6578\u3002\u4f8b\u5982\uff0c\u5c0d\u65bc 4096 \u500b Token \u7684\u8f38\u5165\u63d0\u793a\uff0c\u5927\u7d04\u6703\u7522\u751f 1680 \u842c\u6b21\u6ce8\u610f\u529b\u5206\u6578\u8a08\u7b97\u3002<\/p>\n<div class=\"qc-default-table-wrapper \">\n<table style=\"min-width: 75px;\">\n<colgroup>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/> <\/colgroup>\n<tbody>\n<tr>\n<th colspan=\"1\" rowspan=\"1\">\u904b\u884c\u6307\u6a19<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u9810\u586b\u5145\u968e\u6bb5<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u89e3\u78bc\u968e\u6bb5<\/th>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u4e3b\u8981\u57f7\u884c\u6a21\u5f0f<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u5e73\u884c\u6279\u6b21\u8655\u7406<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u5e8f\u5217 Token \u751f\u6210<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u7b97\u8853\u5f37\u5ea6<\/td>\n<td colspan=\"1\" rowspan=\"1\">200-400 ops\/byte<\/td>\n<td colspan=\"1\" rowspan=\"1\">60-80 ops\/byte<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">GPU \u4f7f\u7528\u7387<\/td>\n<td colspan=\"1\" rowspan=\"1\">90-95%<\/td>\n<td colspan=\"1\" rowspan=\"1\">20-40%<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u786c\u9ad4\u9650\u5236<\/td>\n<td colspan=\"1\" rowspan=\"1\">Tensor Core\uff08\u904b\u7b97\uff09<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u8a18\u61b6\u9ad4\u532f\u6d41\u6392\uff08\u983b\u5bec\uff09<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>\u9019\u7a2e\u5e73\u884c\u57f7\u884c\u5728\u786c\u9ad4\u4e0a\u5be6\u73fe\u4e86\u9ad8\u5ea6\u7684\u8cc7\u6599\u91cd\u8907\u4f7f\u7528\u3002\u4f8b\u5982\uff0c\u5728 70B FP16 \u6a21\u578b\u4e0a\u8655\u7406 1024 Token \u7684\u9810\u586b\u5145\u6642\uff0c\u5176\u7b97\u8853\u5f37\u5ea6\u7d04\u70ba 1020 FLOP\/byte\uff0c\u9060\u9ad8\u65bc NVIDIA H100 SXM GPU \u7684 295 FLOP\/byte\u300c\u5c71\u810a\u9ede\u300d\uff08ridge point\uff09\u3002\u9ad8\u7b97\u8853\u5f37\u5ea6\u6703\u5c07 Tensor Core \u4f7f\u7528\u7387\u63a8\u9ad8\u81f3 90\u201395%\uff0c\u4f7f\u5f97\u904b\u7b97\u541e\u5410\u6210\u70ba\u4e3b\u8981\u74f6\u9838\u3002<\/p>\n<h3>\u89e3\u78bc\u968e\u6bb5\u8207\u8a18\u61b6\u9ad4\u983b\u5bec\u9650\u5236<\/h3>\n<p>\u89e3\u78bc\u968e\u6bb5\u4f9d\u5e8f\u9010\u4e00\u751f\u6210\u8f38\u51fa Token\u3002\u6bcf\u500b\u751f\u6210\u7684 Token \u90fd\u9700\u8981\u5c0d\u6574\u500b\u7db2\u8def\u57f7\u884c\u4e00\u6b21\u7368\u7acb\u7684\u524d\u5411\u904b\u7b97\uff0c\u5c0e\u81f4\u5728\u591a\u6b65\u8f38\u51fa\u4e2d\u5f62\u6210\u8f03\u9577\u7684 Token \u751f\u6210\u6642\u9593\u3002\u4f8b\u5982\uff0c\u5c0d\u65bc 1024 \u500b\u8f38\u51fa Token\uff0c\u4f60\u7684\u7cfb\u7d71\u5fc5\u9808\u57f7\u884c 1024 \u500b\u9023\u7e8c\u6b65\u9a5f\u3002<\/p>\n<p>\u5728\u6bcf\u4e00\u6b21\u89e3\u78bc\u6b65\u9a5f\u4e2d\uff0c\u786c\u9ad4\u90fd\u9700\u8981\u5f9e\u986f\u793a\u8a18\u61b6\u9ad4\u4e2d\u8b80\u53d6\u5b8c\u6574\u7684\u6a21\u578b\u6b0a\u91cd\u4ee5\u53ca\u5df2\u5feb\u53d6\u7684 KV\uff08\u9375\u503c\uff09\u5411\u91cf\u3002\u4ee5\u6279\u6b21\u5927\u5c0f\u70ba 1 \u7684 70B FP16 \u6a21\u578b\u70ba\u4f8b\uff0c\u6bcf\u751f\u6210\u4e00\u500b Token \u5927\u7d04\u9700\u8981\u8b80\u53d6 140 GB \u7684\u6b0a\u91cd\u8cc7\u6599\u3002\u9019\u7a2e\u6301\u7e8c\u7684\u8cc7\u6599\u4e32\u6d41\u6703\u5c07\u7b97\u8853\u5f37\u5ea6\u62c9\u4f4e\u81f3 60\u201380 ops\/byte\uff0c\u540c\u6642\u8b93 GPU \u4f7f\u7528\u7387\u964d\u81f3 20\u201340%\u3002<\/p>\n<p>\u8a18\u61b6\u9ad4\u983b\u5bec\u6703\u6210\u70ba\u7cfb\u7d71\u9650\u5236\u56e0\u7d20\uff0c\u56e0\u70ba GPU \u5927\u90e8\u5206\u6642\u8108\u9031\u671f\u90fd\u5728\u7b49\u5f85\u6b0a\u91cd\u50b3\u8f38\u3002\u900f\u904e\u653e\u5927\u6279\u6b21\u5927\u5c0f\uff0c\u53ef\u4ee5\u5c07\u9019\u4e9b\u6b0a\u91cd\u8b80\u53d6\u958b\u92b7\u6524\u63d0\u5230\u591a\u500b\u8acb\u6c42\u4e0a\uff0c\u5f9e\u800c\u63d0\u5347\u7e3d\u541e\u5410\u91cf\u3002\u59a5\u5584\u512a\u5316\u8a18\u61b6\u9ad4\u983b\u5bec\u53ef\u4ee5\u7a69\u5b9a\u6548\u80fd\uff0c\u4e26\u5728\u751f\u7522\u5de5\u4f5c\u8ca0\u8f09\u4e0b\u6574\u9ad4\u63d0\u5347 Token \u751f\u6210\u6548\u7387\u3002\u6709\u6548\u7ba1\u7406\u9019\u4e9b\u786c\u9ad4\u9650\u5236\uff0c\u53ef\u4ee5\u5728\u9ad8\u4f75\u767c\u63a8\u7406\u4efb\u52d9\u4e2d\u52a0\u901f\u5927\u6a21\u578b\u63a8\u7406\u6d41\u7a0b\uff0c\u540c\u6642\u964d\u4f4e\u6574\u9ad4\u786c\u9ad4\u71df\u904b\u652f\u51fa\u3002<\/p>\n<h2><strong>\u5ef6\u9072\u8207\u56de\u61c9\u6027\u7684\u95dc\u9375\u6307\u6a19<\/strong><\/h2>\n<p>\u5728\u751f\u7522\u74b0\u5883\u4e2d\uff0c\u4f60\u900f\u904e\u8ffd\u8e64\u5404\u57f7\u884c\u968e\u6bb5\u7684\u95dc\u9375\u6307\u6a19\u4f86\u8a55\u4f30\u4f7f\u7528\u8005\u9ad4\u9a57\u3002\u670d\u52d9\u7cfb\u7d71\u63a1\u7528\u7d71\u4e00\u516c\u5f0f\u8a08\u7b97\u7aef\u5230\u7aef\u56de\u61c9\u6642\u9593\uff1aLatency = TTFT + (TPOT * \u8f38\u51fa Token \u6578)\u3002\u6b64\u516c\u5f0f\u5c07\u521d\u59cb\u63d0\u793a\u8655\u7406\u6642\u9593\u8207\u7d2f\u7a4d Token \u751f\u6210\u6642\u9593\u76f8\u52a0\uff0c\u7cbe\u6e96\u8861\u91cf\u6574\u9ad4\u7cfb\u7d71\u56de\u61c9\u6027\u3002<\/p>\n<h3>TTFT\u3001TPOT \u8207 Token \u9593\u5ef6\u9072<\/h3>\n<p>\u6a21\u578b\u53c3\u6578\u898f\u6a21\u8207\u90e8\u7f72\u786c\u9ad4\u6703\u76f4\u63a5\u6c7a\u5b9a\u670d\u52d9\u901f\u5ea6\u3002\u4ee5 LLaMA-2 \u6a21\u578b\u70ba\u4f8b\uff0c\u5728\u5f15\u64ce\u7248\u672c 0.6.1\u30014 \u5f35 NVIDIA A100 GPU\u300125 \u500b\u4f75\u767c\u4f7f\u7528\u8005\u7684\u689d\u4ef6\u4e0b\uff0c\u57fa\u6e96\u6e2c\u8a66\u51f8\u986f\u4e86\u9019\u4e9b\u95dc\u9375\u904b\u884c\u6b0a\u8861\uff1a<\/p>\n<div class=\"qc-default-table-wrapper \">\n<table style=\"min-width: 175px;\">\n<colgroup>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/> <\/colgroup>\n<tbody>\n<tr>\n<th colspan=\"1\" rowspan=\"1\">\u6a21\u578b\u898f\u6a21<\/th>\n<th colspan=\"1\" rowspan=\"1\">TTFT p50\uff08\u79d2\uff09<\/th>\n<th colspan=\"1\" rowspan=\"1\">TTFT p95\uff08\u79d2\uff09<\/th>\n<th colspan=\"1\" rowspan=\"1\">TTFT p99\uff08\u79d2\uff09<\/th>\n<th colspan=\"1\" rowspan=\"1\">TPOT p50\uff08\u79d2\uff09<\/th>\n<th colspan=\"1\" rowspan=\"1\">TPOT p95\uff08\u79d2\uff09<\/th>\n<th colspan=\"1\" rowspan=\"1\">TPOT p99\uff08\u79d2\uff09<\/th>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">7B \u6a21\u578b<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.24<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.89<\/td>\n<td colspan=\"1\" rowspan=\"1\">1.42<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.019<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.037<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.056<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">13B \u6a21\u578b<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.31<\/td>\n<td colspan=\"1\" rowspan=\"1\">1.12<\/td>\n<td colspan=\"1\" rowspan=\"1\">1.78<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.033<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.067<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.096<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">70B \u6a21\u578b<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.87<\/td>\n<td colspan=\"1\" rowspan=\"1\">2.34<\/td>\n<td colspan=\"1\" rowspan=\"1\">3.67<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.122<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.231<\/td>\n<td colspan=\"1\" rowspan=\"1\">0.343<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>\u5728\u8f03\u4f4e\u4f75\u767c\u3001\u4ee5 NVIDIA H100 GPU \u70ba\u57fa\u790e\u7684\u60c5\u5883\u4e0b\uff0c\u8f03\u5c0f\u7684\u6a21\u578b\u8b8a\u9ad4\uff08\u5982 Llama 3.1 8B\uff09\u53ef\u5be6\u73fe\u5c0f\u65bc 80 ms \u7684\u9996 Token \u5ef6\u9072\uff0c\u4ee5\u53ca 11\u201321 ms \u7684 Token \u9593\u5ef6\u9072\u3002\u6a21\u578b\u67b6\u69cb\u540c\u6a23\u6703\u5f71\u97ff\u63a8\u7406\u57f7\u884c\u6642\u7684 Token \u5ef6\u9072\u3002Mixture-of-Experts\uff08\u5c08\u5bb6\u6df7\u5408\uff09\u67b6\u69cb\u6bcf\u5c64\u53ea\u555f\u7528\u90e8\u5206\u53c3\u6578\uff0c\u76f8\u8f03\u7a20\u5bc6\u7db2\u8def\u53ef\u5c07\u524d\u994b\u7db2\u8def\u901a\u8a0a\u958b\u92b7\u964d\u4f4e\u7d04 50%\uff1b\u800c\u8f03\u6dfa\u7684\u7db2\u8def\u6df1\u5ea6\u5247\u80fd\u6e1b\u5c11\u5e8f\u5217\u751f\u6210\u904e\u7a0b\u4e2d\u7684 Token \u9593\u5ef6\u9072\uff0c\u9032\u800c\u63d0\u5347\u6574\u9ad4\u5927\u6a21\u578b\u6548\u80fd\u3002<\/p>\n<h3>\u5c3e\u90e8\u5ef6\u9072\u8207\u5206\u4f4d\u6578\u5206\u4f48<\/h3>\n<p>\u4f01\u696d\u74b0\u5883\u6703\u7cbe\u7d30\u76e3\u63a7 Token \u4f7f\u7528\u60c5\u6cc1\uff0c\u4ee5\u5728\u6301\u7e8c\u64f4\u5927\u7684\u4e0a\u4e0b\u6587\u8996\u7a97\u4e2d\u7dad\u6301\u6548\u7387\u3002\u7cfb\u7d71\u63d0\u793a\u5728\u6578\u767e\u842c\u6b21\u8acb\u6c42\u4e2d\u91cd\u8907\uff0c\u6703\u7522\u751f\u9f90\u5927\u7684\u6307\u4ee4\u958b\u92b7\uff1b\u52d5\u614b\u8acb\u6c42\u6279\u6b21\u8655\u7406\u53ef\u4ee5\u5c07\u9019\u985e\u6307\u4ee4\u958b\u92b7\u6700\u591a\u964d\u4f4e 96.5%\uff0c\u907f\u514d\u56b4\u91cd\u7684\u904b\u7b97\u8207\u8a18\u61b6\u9ad4\u6d6a\u8cbb\u3002<\/p>\n<p>\u73fe\u4ee3\u9023\u7e8c\u6279\u6b21\u5f15\u64ce\u6703\u512a\u5148\u6392\u7a0b\u63d0\u793a\u9810\u586b\u5145\uff0c\u800c\u975e\u6b63\u5728\u9032\u884c\u7684\u89e3\u78bc\u6b65\u9a5f\u3002\u9019\u7a2e\u6392\u7a0b\u7b56\u7565\u6703\u5e36\u4f86\u7279\u5b9a\u7684\u904b\u884c\u74f6\u9838\uff1a<\/p>\n<ul>\n<li>\u63d2\u5165\u9810\u586b\u5145\u4efb\u52d9\u6642\uff0c\u6703\u66ab\u6642\u4e2d\u65b7\u73fe\u884c\u7684\u89e3\u78bc\u8fed\u4ee3\u3002<\/li>\n<li>\u5728\u5c16\u5cf0\u8ca0\u8f09\u4e0b\uff0c\u9019\u985e\u57f7\u884c\u505c\u9813\u6703\u8b93\u8f38\u51fa Token \u751f\u6210\u5ef6\u9072\u6578\u79d2\u3002<\/li>\n<li>\u6392\u7a0b\u5ef6\u9072\u6703\u5c0e\u81f4 P99 Token \u9593\u5ef6\u9072\u6307\u6a19\u51fa\u73fe\u56b4\u91cd\u5c3e\u90e8\u5ef6\u9072\u5c16\u5cf0\u3002<\/li>\n<\/ul>\n<p>\u5de5\u7a0b\u5e2b\u5fc5\u9808\u4ee5\u5b8c\u6574\u7684\u7d71\u8a08\u5206\u4f48\u4f86\u5206\u6790\u6548\u80fd\u6307\u6a19\uff0c\u800c\u975e\u53ea\u4f9d\u8cf4\u5e73\u5747\u503c\u3002\u5e73\u5747\u503c\u6703\u63a9\u84cb\u8cc7\u6e90\u722d\u7528\u4e0b\u7684\u5ef6\u9072\u5c16\u5cf0\u3002\u900f\u904e\u8ffd\u8e64\u9ad8\u5206\u4f4d\u6578\uff0c\u53ef\u4ee5\u770b\u6e05\u4f47\u5217\u74f6\u9838\u5728\u9ad8\u8ca0\u8f09\u6642\u5982\u4f55\u5f71\u97ff\u771f\u5be6\u7d42\u7aef\u4f7f\u7528\u8005\u3002\u5c07\u6574\u9ad4 tokens per second \u541e\u5410\u91cf\u8207\u56b4\u683c\u7684\u7e3d\u5ef6\u9072\u4e0a\u9650\u52a0\u4ee5\u5e73\u8861\uff0c\u6709\u52a9\u65bc\u7dad\u6301\u670d\u52d9\u54c1\u8cea\u3002\u4f60\u9700\u8981\u900f\u904e\u8abf\u6574\u5f15\u64ce\u53c3\u6578\u4f86\u7a69\u5b9a\u5927\u6a21\u578b\u63a8\u7406\u6d41\u7a0b\u4e2d\u7684\u6392\u968a\u5ef6\u9072\uff1b\u5c0d\u6548\u80fd\u6307\u6a19\u9032\u884c\u56b4\u5bc6\u76e3\u63a7\uff0c\u53ef\u5728\u8acb\u6c42\u91cf\u6210\u9577\u6642\u78ba\u4fdd\u9ad8\u56de\u61c9\u6027\u3002<\/p>\n<h2><strong>\u5ea6\u91cf\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd\u8207\u541e\u5410\u91cf<\/strong><\/h2>\n<p>\u4f60\u900f\u904e\u6bd4\u8f03\u6a19\u6e96\u7cfb\u7d71\u541e\u5410\u91cf\u8207\u5be6\u969b\u751f\u6210\u901f\u5ea6\u4f86\u8a55\u4f30\u670d\u52d9\u80fd\u529b\u3002\u50b3\u7d71 API \u5de5\u5177\u5728\u5ba2\u6236\u7aef\u5074\u4ee5\u6bcf\u79d2\u8acb\u6c42\u6578\uff08RPS\uff09\u548c\u6bcf\u79d2 Token \u6578\uff08TPS\uff09\u56de\u5831\u541e\u5410\u91cf\u3002\u5ba2\u6236\u7aef\u5de5\u5177 GenAI-Perf \u900f\u904e\u5411\u76f8\u5bb9 OpenAI \u7684\u670d\u52d9\u7aef\u9ede\u767c\u9001\u5de5\u4f5c\u8ca0\u8f09\u6d41\u91cf\uff0c\u5728\u7dda\u4e0a\u58d3\u529b\u6e2c\u8a66\u4e2d\u8ffd\u8e64\u9019\u4e9b\u6548\u80fd\u6307\u6a19\u3002<\/p>\n<h3>\u6bcf\u79d2\u8acb\u6c42\u6578\u8207\u4f75\u767c\u64f4\u5c55<\/h3>\n<p>\u4f60\u5fc5\u9808\u540c\u6642\u5206\u6790\u6bcf\u79d2\u8acb\u6c42\u6578\u8207\u6bcf\u79d2 Token \u6578\uff0c\u624d\u80fd\u5145\u5206\u7406\u89e3\u6574\u9ad4\u6279\u6b21\u8655\u7406\u6548\u7387\u3002\u6a19\u6e96 API \u7aef\u9ede\u900f\u904e\u5df2\u5b8c\u6210\u7684\u5ba2\u6236\u7aef\u8acb\u6c42\u6578\u4f86\u8ffd\u8e64\u4e00\u822c\u670d\u52d9\u80fd\u529b\uff0c\u800c\u804a\u5929\u670d\u52d9\u5247\u900f\u904e\u8f38\u51fa Token \u7684\u6bcf\u79d2\u751f\u6210\u901f\u5ea6\u4f86\u76e3\u63a7\u5373\u6642\u6548\u80fd\u3002<\/p>\n<ul>\n<li>\u5ba2\u6236\u7aef\u5074\u7684 Token \u751f\u6210\u541e\u5410\u53cd\u6620\u4f7f\u7528\u8005\u5728\u4e32\u6d41\u6703\u8a71\u4e2d\u7684\u4e3b\u89c0\u50b3\u8f38\u901f\u5ea6\u3002<\/li>\n<li>\u4f3a\u670d\u7aef\u60c5\u5883\u6703\u5728\u56fa\u5b9a\u5ef6\u9072\u76ee\u6a19\u4e0b\u8a55\u4f30 TTFT \u8207 TPOT\uff0c\u4f8b\u5982 450 ms \u7684\u9996 Token \u5ef6\u9072\u8207 40 ms \u7684\u55ae Token \u751f\u6210\u6642\u9593\u3002<\/li>\n<li>\u96e2\u7dda\u60c5\u5883\u5247\u5728\u5b8c\u6574\u8cc7\u6599\u96c6\u904b\u884c\u671f\u9593\uff0c\u805a\u5408\u6bcf\u79d2\u751f\u6210\u7684\u7e3d\u8f38\u51fa Token \u6578\u3002<\/li>\n<\/ul>\n<p>\u91dd\u5c0d\u6279\u6b21\u6700\u4f73\u5316\u7684\u57fa\u6e96\u6e2c\u8a66\u5f80\u5f80\u900f\u904e\u56fa\u5b9a\u9577\u5ea6\u63d0\u793a\u3001\u5728\u9810\u71b1\u5b8c\u6210\u7684 GPU \u4e0a\u57f7\u884c\uff0c\u4ee5\u63a9\u84cb\u5be6\u969b\u904b\u884c\u4e2d\u7684\u505c\u9813\u3002\u6df7\u5408\u63d0\u793a\u9577\u5ea6\u5728\u771f\u5be6\u74b0\u5883\u4e2d\u6703\u5c0e\u81f4 GPU \u4f7f\u7528\u7387\u4e0b\u964d\u8207\u52d5\u614b\u6392\u968a\u5ef6\u9072\u3002\u51b7\u5feb\u53d6\u672a\u547d\u4e2d\u4e5f\u6703\u589e\u52a0\u4e92\u52d5\u4f7f\u7528\u8005\u7684\u7e3d\u5ef6\u9072\u3002\u4f60\u61c9\u7d50\u5408\u8acb\u6c42\u91cf\u8207\u5be6\u969b\u8f38\u51fa Token \u5ea6\u91cf\uff0c\u4f86\u6700\u4f73\u5316\u90e8\u7f72\u786c\u9ad4\u7684\u8cc7\u6e90\u914d\u7f6e\u3002<\/p>\n<h3>Goodput \u8207 SLO \u9054\u6210\u7387<\/h3>\n<p>\u539f\u59cb\u541e\u5410\u6578\u64da\u4e26\u4e0d\u80fd\u4fdd\u8b49\u4f7f\u7528\u8005\u5728\u4f01\u696d\u4f7f\u7528\u60c5\u5883\u4e2d\u771f\u6b63\u6eff\u610f\u3002\u4f60\u900f\u904e\u8ffd\u8e64\u751f\u7522\u53e2\u96c6\u4e2d\u7684 Goodput \u4f86\u5ea6\u91cf\u6709\u6548\u670d\u52d9\u6210\u529f\u7387\u3002Goodput \u7d71\u8a08\u7684\u662f\u5728\u55ae\u4f4d\u6642\u9593\u5167\uff0c\u5b8c\u5168\u5728\u670d\u52d9\u7b49\u7d1a\u76ee\u6a19\uff08SLO\uff09\u5167\u5b8c\u6210\u7684\u6210\u529f\u8acb\u6c42 Token \u6578\u3002<\/p>\n<div class=\"qc-default-table-wrapper \">\n<table style=\"min-width: 75px;\">\n<colgroup>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/> <\/colgroup>\n<tbody>\n<tr>\n<th colspan=\"1\" rowspan=\"1\">\u6307\u6a19<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u5b9a\u7fa9<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u91cf\u6e2c\u65b9\u6cd5<\/th>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u670d\u52d9\u7b49\u7d1a\u76ee\u6a19\u9054\u6210\u7387<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u5728\u6240\u6709 Token \u4ea4\u4ed8\u671f\u9650\u5167\u5b8c\u6210\u7684\u8acb\u6c42\u5360\u6bd4<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u53ea\u7d71\u8a08\u7b26\u5408\u76ee\u6a19\u6642\u9593\u689d\u4ef6\u7684\u5b8c\u6574\u8acb\u6c42<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">Goodput<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u7b26\u5408\u670d\u52d9\u7b49\u7d1a\u76ee\u6a19\u7684 Token \u541e\u5410<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u4ee5\u6210\u529f\u8f38\u51fa\u7684 Token \u6578\u9664\u4ee5\u7e3d\u670d\u52d9\u6642\u9593<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u5e73\u6ed1 Goodput<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u9023\u7e8c\u5316\u7684\u670d\u52d9\u6548\u76ca\u8a08\u7b97<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u900f\u904e\u61f2\u7f70\u5ef6\u9072 Token\uff0c\u800c\u975e\u76f4\u63a5\u6368\u68c4\u903e\u6642\u8acb\u6c42\u4f86\u8a08\u7b97<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>\u91cf\u6e2c Goodput \u6709\u52a9\u65bc\u5c07\u771f\u6b63\u7684\u71df\u904b\u4ea4\u4ed8\u901f\u5ea6\u8207\u88ab\u6d6a\u8cbb\u7684\u904b\u7b97\u6b65\u9a5f\u5340\u5206\u958b\u4f86\u3002\u9ad8\u8ca0\u8f09\u6703\u589e\u52a0\u6392\u968a\u5ef6\u9072\uff0c\u9032\u800c\u524a\u5f31\u4e92\u52d5\u6548\u80fd\u3002\u73fe\u4ee3\u63a8\u7406\u5de5\u5177\u900f\u904e\u6263\u9664\u57fa\u65bc\u4f7f\u7528\u8005\u9ad4\u9a57\u6821\u6e96\u7684\u5ef6\u9072\u61f2\u7f70\u4f86\u8a08\u7b97\u5e73\u6ed1 Goodput\uff0c\u6b64\u4e00\u9023\u7e8c\u6307\u6a19\u6709\u52a9\u65bc\u5728\u771f\u5be6\u6d41\u91cf\u9ad8\u5cf0\u4e0b\u8a55\u4f30\u5927\u8a9e\u8a00\u6a21\u578b\u8868\u73fe\u3002\u4f60\u53ef\u4ee5\u900f\u904e\u8abf\u6574\u57f7\u884c\u6279\u6b21\u7b56\u7565\uff0c\u5728\u78ba\u4fdd\u6240\u6709\u7bc0\u9ede\u7b26\u5408\u56b4\u683c\u5ef6\u9072\u9580\u6abb\u7684\u524d\u63d0\u4e0b\uff0c\u7dad\u6301\u9ad8\u6c34\u6e96\u7684\u5927\u6a21\u578b\u6548\u80fd\u3002\u5c0d\u9019\u4e9b\u95dc\u9375\u6307\u6a19\u7684\u6a19\u6e96\u5316\u5ea6\u91cf\uff0c\u6709\u52a9\u65bc\u5efa\u69cb\u53ef\u64f4\u5c55\u3001\u5177\u6210\u672c\u6548\u76ca\u7684\u5927\u6a21\u578b\u63a8\u7406\u6d41\u7a0b\uff1b\u8ffd\u8e64 Goodput \u80fd\u78ba\u4fdd\u63a8\u7406\u5e73\u53f0\u5728\u4e0d\u6d6a\u8cbb\u786c\u9ad4\u904b\u7b97\u8cc7\u6e90\u7684\u60c5\u6cc1\u4e0b\uff0c\u6301\u7e8c\u63d0\u4f9b\u7a69\u5b9a\u56de\u61c9\u3002<\/p>\n<h2><strong>\u6700\u4f73\u5316\u5927\u6a21\u578b\u63a8\u7406\u4e2d\u7684\u7cfb\u7d71\u6b0a\u8861<\/strong><\/h2>\n<p>\u6700\u4f73\u5316\u5927\u6a21\u578b\u63a8\u7406\u9700\u8981\u5728\u8655\u7406\u901f\u5ea6\u8207\u786c\u9ad4\u8a18\u61b6\u9ad4\u4f7f\u7528\u4e4b\u9593\u53d6\u5f97\u5e73\u8861\u3002\u50b3\u7d71\u975c\u614b\u6279\u6b21\u8655\u7406\u6703\u5c0e\u81f4\u56b4\u91cd\u7684\u4f3a\u670d\u5668\u5ef6\u9072\uff0c\u56e0\u70ba\u77ed\u8f38\u51fa\u5fc5\u9808\u7b49\u5f85\u9577\u8f38\u51fa\u5b8c\u6210\u3002\u73fe\u4ee3\u670d\u52d9\u6846\u67b6\u900f\u904e\u9023\u7e8c\u6279\u6b21\u8655\u7406\u8207\u9032\u968e\u8a18\u61b6\u9ad4\u7ba1\u7406\u4f86\u89e3\u6c7a\u6b64\u4e00\u554f\u984c\u3002<\/p>\n<h3>\u9023\u7e8c\u6279\u6b21\u8655\u7406\u8207\u8a18\u61b6\u9ad4\u7ba1\u7406<\/h3>\n<p>\u9023\u7e8c\u6279\u6b21\u8655\u7406\u5728\u8fed\u4ee3\u5c64\u7d1a\u904b\u4f5c\uff0c\u4ee5\u6709\u6548\u8655\u7406\u52d5\u614b\u5de5\u4f5c\u8ca0\u8f09\u3002\u7cfb\u7d71\u6392\u7a0b\u5668\u6703\u5728\u6bcf\u500b\u8fed\u4ee3\u6b65\u9a5f\u5f8c\u6aa2\u67e5\u76ee\u524d\u6d3b\u52d5\u8acb\u6c42\u4f47\u5217\u3002<\/p>\n<ol>\n<li>\u8a18\u61b6\u9ad4\u5f15\u64ce\u5c07\u5e8f\u5217\u5feb\u53d6\u8cc7\u6599\u5207\u5206\u70ba\u908f\u8f2f\u8a18\u61b6\u9ad4\u5340\u584a\u3002<\/li>\n<li>\u57f7\u884c\u968e\u6bb5\u5c07\u9019\u4e9b\u908f\u8f2f\u5340\u584a\u5c0d\u61c9\u5230\u5be6\u9ad4\u986f\u793a\u8a18\u61b6\u9ad4\u5340\u584a\u3002<\/li>\n<li>\u5340\u584a\u5c0d\u7167\u8868\u8ffd\u8e64\u5206\u6563\u65bc\u5be6\u9ad4\u8a18\u61b6\u9ad4\u4e2d\u7684\u6240\u6709\u908f\u8f2f\u6620\u5c04\u3002<\/li>\n<li>\u670d\u52d9\u5f15\u64ce\u6703\u96a8\u8457\u5e8f\u5217\u7522\u751f Token \u52d5\u614b\u914d\u7f6e\u65b0\u7684\u8a18\u61b6\u9ad4\u5340\u584a\u3002<\/li>\n<li>\u8acb\u6c42\u7d50\u675f\u5f8c\uff0c\u7cfb\u7d71\u6703\u7acb\u5373\u91cb\u653e\u5c0d\u61c9\u7684\u5e8f\u5217\u8a18\u61b6\u9ad4\u5340\u584a\u3002<\/li>\n<li>\u5f15\u64ce\u6703\u5c07\u7b49\u5f85\u4e2d\u7684\u8acb\u6c42\u63d2\u5165\u7a7a\u9592\u7684\u6279\u6b21\u69fd\u4f4d\uff0c\u800c\u4e0d\u62d6\u6162\u6b63\u5728\u57f7\u884c\u7684\u5e8f\u5217\u3002<\/li>\n<\/ol>\n<p>\u9019\u7a2e\u52d5\u614b\u65b9\u5f0f\u6d88\u9664\u4e86\u9810\u5148\u914d\u7f6e\u8a18\u61b6\u9ad4\u7684\u6d6a\u8cbb\u3002PagedAttention \u5728\u751f\u7522\u6d41\u91cf\u4e0b\u53ef\u5c07\u8a18\u61b6\u9ad4\u6d6a\u8cbb\u58d3\u4f4e\u81f3\u7d04 4%\u3002<\/p>\n<div class=\"qc-default-table-wrapper \">\n<table style=\"min-width: 100px;\">\n<colgroup>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/> <\/colgroup>\n<tbody>\n<tr>\n<th colspan=\"1\" rowspan=\"1\">\u5de5\u4f5c\u8ca0\u8f09\u985e\u578b<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u56fa\u5b9a\u6279\u6b21\uff08\u8acb\u6c42\/\u5206\u9418\uff09<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u9023\u7e8c\u6279\u6b21\uff08\u8acb\u6c42\/\u5206\u9418\uff09<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u541e\u5410\u63d0\u5347<\/th>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u6df7\u5408\u56de\u61c9\u9577\u5ea6\uff0850\u2013800 Token\uff09<\/td>\n<td colspan=\"1\" rowspan=\"1\">25<\/td>\n<td colspan=\"1\" rowspan=\"1\">85<\/td>\n<td colspan=\"1\" rowspan=\"1\">3.4x<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u4e92\u52d5\u5f0f\u804a\u5929\uff08\u8f2a\u6b21\u53ef\u8b8a\uff09<\/td>\n<td colspan=\"1\" rowspan=\"1\">35<\/td>\n<td colspan=\"1\" rowspan=\"1\">110<\/td>\n<td colspan=\"1\" rowspan=\"1\">3.1x<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u7a69\u5b9a\u77ed\u56de\u61c9\uff08&lt;100 Token\uff09<\/td>\n<td colspan=\"1\" rowspan=\"1\">60<\/td>\n<td colspan=\"1\" rowspan=\"1\">95<\/td>\n<td colspan=\"1\" rowspan=\"1\">1.6x<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u9577\u6587\u672c\u751f\u6210\uff08\u5e73\u5747 &gt;500 Token\uff09<\/td>\n<td colspan=\"1\" rowspan=\"1\">12<\/td>\n<td colspan=\"1\" rowspan=\"1\">28<\/td>\n<td colspan=\"1\" rowspan=\"1\">2.3x<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>\u5728\u591a\u7a2e\u56de\u61c9\u9577\u5ea6\u4e0b\uff0c\u4f60\u90fd\u80fd\u7372\u5f97\u986f\u8457\u7684\u541e\u5410\u63d0\u5347\u3002<\/p>\n<ul>\n<li>\u9023\u7e8c\u6279\u6b21\u8655\u7406\u6703\u5728\u65e2\u6709\u8acb\u6c42\u91cb\u653e\u986f\u793a\u8a18\u61b6\u9ad4\u5f8c\uff0c\u7acb\u5373\u63a5\u7d0d\u6392\u968a\u4e2d\u7684\u8acb\u6c42\u3002<\/li>\n<li>\u5728\u5e8f\u5217\u751f\u6210\u6b65\u9a5f\u4e2d\uff0cGPU \u6279\u6b21\u59cb\u7d42\u7dad\u6301\u9ad8\u586b\u5145\u5ea6\u3002<\/li>\n<li>\u77ed\u5e8f\u5217\u6703\u63d0\u524d\u7d50\u675f\uff0c\u800c\u5176\u9918\u5e8f\u5217\u5247\u5728\u4e0d\u589e\u52a0\u984d\u5916\u5ef6\u9072\u7684\u60c5\u6cc1\u4e0b\u6301\u7e8c\u751f\u6210\u3002<\/li>\n<\/ul>\n<p>\u66f4\u9ad8\u7684\u4f75\u767c\u6703\u8b93 Token \u751f\u6210\u6b65\u9a5f\u8b8a\u6162\uff0c\u56e0\u6b64\u4f60\u5fc5\u9808\u5c07\u9023\u7e8c\u6279\u6b21\u8655\u7406\u8207\u6a21\u578b\u6700\u4f73\u5316\u7d50\u5408\uff0c\u624d\u80fd\u5728\u9ad8\u4f7f\u7528\u8005\u91cf\u4e0b\u7dad\u6301\u7cfb\u7d71\u6548\u7387\u3002<\/p>\n<h3>\u63a8\u6e2c\u89e3\u78bc\u8207\u89e3\u8026\u5f0f\u670d\u52d9<\/h3>\n<p>\u63a8\u6e2c\u89e3\u78bc\u900f\u904e\u5728\u4e3b\u6a21\u578b\u65c1\u908a\u57f7\u884c\u4e00\u500b\u8f03\u5c0f\u7684\u8349\u7a3f\u6a21\u578b\u4f86\u52a0\u901f\u63a8\u7406\u3002\u8349\u7a3f\u6a21\u578b\u53ef\u4ee5\u5feb\u901f\u751f\u6210\u5019\u9078 Token\uff0c\u800c\u5927\u578b\u76ee\u6a19\u6a21\u578b\u5247\u4ee5\u4e00\u6b21\u5e73\u884c\u524d\u5411\u904b\u7b97\u9a57\u8b49\u9019\u4e9b\u5019\u9078 Token\u3002<\/p>\n<div class=\"qc-default-table-wrapper \">\n<table style=\"min-width: 50px;\">\n<colgroup>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/> <\/colgroup>\n<tbody>\n<tr>\n<th colspan=\"1\" rowspan=\"1\">\u52a0\u901f\u6548\u679c<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u689d\u4ef6<\/th>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u6700\u9ad8 111% \u541e\u5410\u63d0\u5347<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u70ba LLaMa-65B \u8a2d\u8a08\u7684\u65b0\u8349\u7a3f\u6a21\u578b\uff0c\u76f8\u8f03\u65e2\u6709\u8349\u7a3f\u6a21\u578b\uff0c\u5728\u7dad\u6301\u6e96\u78ba\u7387\u7684\u524d\u63d0\u4e0b\u63d0\u5347\u541e\u5410\u3002<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u8d85\u904e 60% \u541e\u5410\u63d0\u5347<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u5728\u53c3\u6578\u7e3d\u91cf\u4e0d\u8b8a\u7684\u524d\u63d0\u4e0b\uff0c\u4ee5\u66f4\u6df1\u4e14\u66f4\u7a84\u7684\u65b0\u8349\u7a3f\u6a21\u578b\u53d6\u4ee3\u65e2\u6709\u6a21\u578b\u3002<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">97%\u2013111% \u541e\u5410\u63d0\u5347<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u5728\u6eab\u5ea6\u53d6\u6a23\u5c0d\u7167\u5be6\u9a57\u4e2d\uff0cNoFT-Wide-796M \u8349\u7a3f\u6a21\u578b\u76f8\u8f03\u67d0\u5df2\u5fae\u8abf\u6a21\u578b\u7684\u8868\u73fe\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>\u9019\u7a2e\u57f7\u884c\u65b9\u5f0f\u5728\u5b8c\u5168\u7dad\u6301\u76ee\u6a19\u8f38\u51fa\u54c1\u8cea\u7684\u524d\u63d0\u4e0b\uff0c\u5927\u5e45\u964d\u4f4e\u751f\u6210\u5ef6\u9072\u3002\u4f60\u53ef\u4ee5\u5728\u7121\u9700\u91cd\u65b0\u8a13\u7df4\u5927\u6a21\u578b\u7684\u60c5\u6cc1\u4e0b\uff0c\u7372\u5f97\u66f4\u5feb\u7684\u55ae\u4e00 Token \u751f\u6210\u901f\u5ea6\u3002<\/p>\n<p>\u591a\u7bc0\u9ede\u90e8\u7f72\u6703\u5c07\u904b\u7b97\u5bc6\u96c6\u7684\u9810\u586b\u5145\u4f5c\u696d\u8207\u8a18\u61b6\u9ad4\u5bc6\u96c6\u7684\u89e3\u78bc\u6b65\u9a5f\u62c6\u5206\u5230\u4e0d\u540c\u7684 GPU \u53e2\u96c6\u4e0a\u3002<\/p>\n<div class=\"qc-default-table-wrapper \">\n<table style=\"min-width: 50px;\">\n<colgroup>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/> <\/colgroup>\n<tbody>\n<tr>\n<th colspan=\"1\" rowspan=\"1\">\u7d50\u8ad6<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u767c\u73fe<\/th>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u5ef6\u9072\u6539\u5584<\/td>\n<td colspan=\"1\" rowspan=\"1\">PPD\uff08\u9810\u586b\u5145\/\u89e3\u78bc\u89e3\u8026\u670d\u52d9\u8b8a\u9ad4\uff09\u5728\u7dad\u6301 TPOT \u7af6\u722d\u529b\u7684\u540c\u6642\uff0c\u5c07\u7b2c 2 \u8f2a\u53ca\u4e4b\u5f8c\u7684 TTFT \u964d\u4f4e\u7d04 68%\u3002<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u541e\u5410\u8207\u8ca0\u8f09\u884c\u70ba<\/td>\n<td colspan=\"1\" rowspan=\"1\">PPD \u5728\u9ad8\u8ca0\u8f09\u60c5\u5883\u4e0b\u53ef\u7de9\u89e3 KV \u50b3\u8f38\u58c5\u585e\u3002<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">\u67b6\u69cb\u539f\u7406<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u5c07\u9810\u586b\u5145\u8207\u89e3\u78bc\u5be6\u9ad4\u62c6\u5206\u5230\u4e0d\u540c GPU \u8cc7\u6e90\u6c60\uff0c\u53ef\u4ee5\u6d88\u9664\u4e92\u76f8\u5e72\u64fe\uff0c\u4f7f P\/D \u8cc7\u6e90\u80fd\u7368\u7acb\u64f4\u7e2e\u4e26\u652f\u63f4\u786c\u9ad4\u591a\u6a23\u6027\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>\u5728\u7570\u8cea\u90e8\u7f72\u4e2d\u4f7f\u7528 A100 \u7cfb\u7d71\uff0c\u53ef\u4ee5\u81f3\u5c11\u5c07\u55ae\u4e00 Token \u5ef6\u9072\u964d\u4f4e 2 ms\uff1b\u800c\u5728 25 Gbps \u4e59\u592a\u7db2\u8def\u4e0b\uff0c\u8de8\u7bc0\u9ede\u8cc7\u6599\u50b3\u8f38\u6bcf\u500b Token \u7684\u984d\u5916\u958b\u92b7\u6700\u591a\u7d04\u70ba 0.21 ms\u3002\u900f\u904e\u89e3\u8026\u7bc0\u9ede\uff0c\u53ef\u4ee5\u5728\u5206\u6563\u5f0f\u53e2\u96c6\u4e2d\u6700\u4f73\u5316\u5927\u6a21\u578b\u63a8\u7406\u6d41\u7a0b\uff1b\u91dd\u5c0d\u6027\u7684\u786c\u9ad4\u6700\u4f73\u5316\u80fd\u5728\u751f\u7522\u898f\u6a21\u4e0b\u7a69\u5b9a\u56de\u61c9\u5ef6\u9072\u3002<\/p>\n<h2><strong>2026 \u5e74\u751f\u7522\u57fa\u6e96\u8207\u6c7a\u7b56\u67b6\u69cb<\/strong><\/h2>\n<p>\u4f60\u5fc5\u9808\u8a2d\u5b9a\u6e05\u6670\u7684\u57fa\u6e96\u76ee\u6a19\uff0c\u624d\u80fd\u5728\u771f\u5be6\u751f\u7522\u74b0\u5883\u4e2d\u8a55\u4f30\u7cfb\u7d71\u6548\u7387\u3002\u751f\u6210\u5f0f AI \u57fa\u790e\u8a2d\u65bd\u9700\u8981\u5c0d\u5ef6\u9072\u76ee\u6a19\u8207\u904b\u7b97\u4e0a\u9650\u9032\u884c\u56b4\u683c\u76e3\u63a7\uff0c\u4e0d\u540c\u4f7f\u7528\u60c5\u5883\u5728\u5be6\u969b\u90e8\u7f72\u4e2d\u5c0d\u61c9\u4e0d\u540c\u6548\u80fd\u76ee\u6a19\u3002<\/p>\n<h3>\u4f9d\u5de5\u4f5c\u8ca0\u8f09\u8a55\u4f30 Token \u751f\u6210\u6548\u7387<\/h3>\n<p>\u4e92\u52d5\u5f0f\u61c9\u7528\u9700\u8981\u5feb\u901f\u7684\u9996\u56de\u61c9\u4f86\u7dad\u6301\u4f7f\u7528\u8005\u9ecf\u8457\u5ea6\u3002GuideLLM \u70ba\u4f01\u696d\u7d1a\u804a\u5929\u90e8\u7f72\u63d0\u51fa\u4e86\u5177\u9ad4\u670d\u52d9\u76ee\u6a19\uff1a<\/p>\n<ul>\n<li>\u9996 Token \u5ef6\u9072\u76ee\u6a19\u9808\u63a7\u5236\u5728 200 ms \u4ee5\u5167\uff1b<\/li>\n<li>\u6b64\u56de\u61c9\u76ee\u6a19\u9700\u5728 99% \u7684\u7e3d\u8acb\u6c42\u4e2d\u9054\u6210\u3002<\/li>\n<\/ul>\n<p>\u4f60\u900f\u904e\u5728\u5ef6\u9072\u76ee\u6a19\u8207\u7cfb\u7d71\u786c\u9ad4\u6210\u672c\u4e4b\u9593\u6b0a\u8861\uff0c\u4f86\u8861\u91cf Token \u751f\u6210\u6548\u7387\u3002\u5373\u6642\u804a\u5929\u61c9\u7528\u66f4\u91cd\u8996\u4f4e\u9996\u5ef6\u9072\uff0c\u800c\u975e\u6700\u5927\u6279\u6b21\u5927\u5c0f\uff1b\u9577\u4e0a\u4e0b\u6587\u6458\u8981\u4efb\u52d9\u5247\u66f4\u91cd\u8996\u8f38\u51fa\u541e\u5410\uff0c\u800c\u5c0d\u9996\u56de\u61c9\u901f\u5ea6\u8981\u6c42\u8f03\u4f4e\u3002<\/p>\n<p>\u5de5\u7a0b\u5e2b\u6703\u5206\u6790\u6548\u80fd\u6307\u6a19\uff0c\u4ee5\u8a55\u4f30\u9ad8\u8ca0\u8f09\u4e0b\u6574\u9ad4\u4f47\u5217\u5065\u5eb7\u5ea6\u3002\u9ad8\u4f7f\u7528\u8005\u6d41\u91cf\u6703\u5728\u670d\u52d9\u7bc0\u9ede\u9593\u589e\u52a0\u6392\u968a\u5ef6\u9072\u3002\u4f60\u53ef\u4ee5\u900f\u904e\u8b93\u57f7\u884c\u968e\u6bb5\u7d44\u614b\u8cbc\u5408\u5be6\u969b\u5de5\u4f5c\u8ca0\u8f09\uff0c\u4f86\u4fdd\u8b77\u5927\u6a21\u578b\u63a8\u7406\u6d41\u7a0b\uff1b\u6709\u6548\u7684\u6a21\u578b\u6700\u4f73\u5316\u53ef\u4ee5\u5728\u4e0d\u89f8\u767c\u8a18\u61b6\u9ad4\u6ea2\u4f4d\u7684\u524d\u63d0\u4e0b\u7a69\u5b9a\u8655\u7406\u901f\u5ea6\u3002<\/p>\n<h3>\u6210\u672c\u6548\u7387\u8207\u786c\u9ad4\u9078\u578b\u7b56\u7565<\/h3>\n<p>\u786c\u9ad4\u9078\u578b\u76f4\u63a5\u6c7a\u5b9a\u6574\u9ad4\u8cc7\u672c\u652f\u51fa\u8207\u71df\u904b\u6210\u672c\u3002\u5c0d\u65bc\u5c0f\u578b\u5718\u968a\u800c\u8a00\uff0c\u5728\u672c\u5730\u786c\u9ad4\u4e0a\u90e8\u7f72\u5927\u8a9e\u8a00\u6a21\u578b\uff0c\u53ef\u4ee5\u66f4\u597d\u5730\u638c\u63e1\u8a18\u61b6\u9ad4\u8cc7\u6e90\u3002\u4f8b\u5982\uff0c\u4e00\u53f0\u642d\u8f09 M4 Ultra \u7684 Mac Studio \u5728 Q8 \u91cf\u5316\u689d\u4ef6\u4e0b\u57f7\u884c Llama 3 70B \u6642\uff0c\u53ef\u9054\u5230\u7d04 15\u201325 Token\/\u79d2\u7684\u901f\u5ea6\uff0c\u800c\u56e0\u5b58\u5728\u5378\u8f09\u958b\u92b7\u7684 RTX 5090 \u5247\u53ea\u6709\u7d04 3\u20135 Token\/\u79d2\u3002\u8cc7\u6599\u4e2d\u5fc3\u90e8\u7f72\u5247\u9700\u8981\u5c08\u7528\u52a0\u901f\u786c\u9ad4\u4f86\u627f\u8f09\u4f01\u696d\u7d1a\u9700\u6c42\u3002<\/p>\n<div class=\"qc-default-table-wrapper \">\n<table style=\"min-width: 100px;\">\n<colgroup>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/>\n<col style=\"min-width: 25px;\" \/> <\/colgroup>\n<tbody>\n<tr>\n<th colspan=\"1\" rowspan=\"1\">\u786c\u9ad4\u5e73\u53f0<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u5de5\u4f5c\u8ca0\u8f09\u6216\u57fa\u6e96\u60c5\u5883<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u71df\u904b CPM<\/th>\n<th colspan=\"1\" rowspan=\"1\">\u95dc\u9375\u6548\u80fd\u6307\u6a19<\/th>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">NVIDIA B200<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u4f01\u696d\u7d1a\u670d\u52d9<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u7d04 $0.02 \/ \u767e\u842c Token<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u7d04 60,000 Token\/\u79d2 \/ GPU<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">H100 SXM<\/td>\n<td colspan=\"1\" rowspan=\"1\">Llama 4 Scout 17B \u65bc vLLM \u4e0a<\/td>\n<td colspan=\"1\" rowspan=\"1\">$0.19 \/ \u767e\u842c Token<\/td>\n<td colspan=\"1\" rowspan=\"1\">4,200 Token\/\u79d2 \u541e\u5410\u91cf<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">H100 SXM<\/td>\n<td colspan=\"1\" rowspan=\"1\">70B FP8\uff0c\u7d50\u5408\u6700\u4f73\u6279\u6b21\u8a2d\u5b9a<\/td>\n<td colspan=\"1\" rowspan=\"1\">$0.32\u2013$0.54 \/ \u767e\u842c Token<\/td>\n<td colspan=\"1\" rowspan=\"1\">1,500\u20132,500 Token\/\u79d2 \u541e\u5410\u91cf<\/td>\n<\/tr>\n<tr>\n<td colspan=\"1\" rowspan=\"1\">MI300X<\/td>\n<td colspan=\"1\" rowspan=\"1\">Mixtral 8x7B\uff0c\u6279\u6b21\u5927\u5c0f\u70ba 1<\/td>\n<td colspan=\"1\" rowspan=\"1\">$22.22 \/ \u767e\u842c Token<\/td>\n<td colspan=\"1\" rowspan=\"1\">\u8cb7\u65b7\u8207\u79df\u7528\u7684\u640d\u76ca\u5169\u5e73\u9ede\u7d04\u70ba 9,045 \u5c0f\u6642<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>\u9078\u64c7\u73fe\u4ee3\u8cc7\u6599\u4e2d\u5fc3\u6676\u7247\u53ef\u4ee5\u5927\u5e45\u964d\u4f4e Token \u6210\u672c\u3002Blackwell \u67b6\u69cb\u76f8\u8f03\u65bc Hopper \u6676\u7247\uff0cToken \u6210\u672c\u6700\u9ad8\u53ef\u964d\u4f4e\u9054 10 \u500d\u3002<\/p>\n<p>\u4f60\u53ef\u4ee5\u900f\u904e\u5c07\u786c\u9ad4\u5e73\u53f0\u8207\u5be6\u969b\u6279\u6b21\u5927\u5c0f\u5206\u4f48\u5c0d\u9f4a\u4f86\u6700\u4f73\u5316\u71df\u904b\u6210\u672c\u3002MI300X GPU \u5728\u6279\u6b21\u5927\u5c0f\u5927\u65bc 256 \u7684\u60c5\u5883\u4e0b\u63d0\u4f9b\u66f4\u4f4e\u7684 Token \u6210\u672c\uff1bH100 GPU \u5247\u5728\u6279\u6b21\u5927\u5c0f 2\u2013128 \u7684\u4e2d\u7b49\u6279\u6b21\u5340\u9593\u5177\u5099\u66f4\u4f73\u7684\u55ae\u4f4d Token \u6210\u672c\u3002\u6700\u4f73\u5316\u5927\u6a21\u578b\u63a8\u7406\uff0c\u9700\u8981\u5c07\u786c\u9ad4\u8a18\u61b6\u9ad4\u983b\u5bec\u8207\u9810\u671f\u6279\u6b21\u5927\u5c0f\u7cbe\u6e96\u5339\u914d\uff1b\u826f\u597d\u7684\u63a8\u7406\u67b6\u69cb\u8a2d\u8a08\u4e0d\u50c5\u80fd\u63d0\u4f9b\u5f37\u52c1\u7684\u5927\u6a21\u578b\u6548\u80fd\uff0c\u4e5f\u80fd\u63a7\u5236\u786c\u9ad4\u652f\u51fa\u3002\u900f\u904e\u7cbe\u6e96\u7684\u6700\u4f73\u5316\uff0c\u4f60\u53ef\u4ee5\u5728\u9ad8\u751f\u7522\u6d41\u91cf\u4e0b\u7dad\u6301\u4f01\u696d\u90e8\u7f72\u7684\u7a69\u5b9a\u6027\uff0c\u540c\u6642\u63d0\u5347\u6574\u9ad4\u5927\u6a21\u578b\u5bb9\u91cf\u800c\u4e0d\u8d85\u51fa\u9810\u7b97\u3002<\/p>\n<div class=\"qc-divider-wrapper\">\n<div class=\"qc-divider\" style=\"border-top-style: solid; width: 100%; border-top-color: #d1d1d1;\"><\/div>\n<\/div>\n<p>\u5728\u63a8\u7406\u904e\u7a0b\u4e2d\uff0c\u4f60\u9700\u8981\u900f\u904e\u7ba1\u7406\u5169\u500b\u8655\u7406\u968e\u6bb5\u4f86\u5e73\u8861\u5927\u6a21\u578b\u6548\u80fd\uff1a\u4ee5\u904b\u7b97\u70ba\u4e3b\u7684\u9810\u586b\u5145\u968e\u6bb5\u9700\u8981\u5f37\u5927\u7684\u5e73\u884c\u904b\u7b97\u80fd\u529b\uff0c\u800c\u4ee5\u8a18\u61b6\u9ad4\u70ba\u4e3b\u7684\u89e3\u78bc\u968e\u6bb5\u5247\u4ef0\u8cf4\u9ad8\u986f\u793a\u8a18\u61b6\u9ad4\u983b\u5bec\u3002\u5c07\u9019\u5169\u500b\u8a08\u7b97\u968e\u6bb5\u9069\u5ea6\u9694\u96e2\uff0c\u53ef\u4ee5\u63d0\u5347\u6574\u500b\u53e2\u96c6\u7684 Token \u751f\u6210\u6548\u7387\u3002<\/p>\n<p>\u5728\u5927\u6a21\u578b\u63a8\u7406\u4e2d\uff0c\u552f\u6709\u5c07 Goodput \u8207\u670d\u52d9\u7b49\u7d1a\u76ee\u6a19\uff08SLO\uff09\u7f6e\u65bc\u539f\u59cb\u541e\u5410\u91cf\u4e4b\u4e0a\uff0c\u4f60\u624d\u80fd\u7372\u5f97\u771f\u6b63\u7684\u7cfb\u7d71\u50f9\u503c\u3002\u55ae\u7d14\u6700\u5927\u5316\u8f38\u51fa Token \u6578\uff0c\u4e26\u4e0d\u4ee3\u8868\u4f7f\u7528\u8005\u6eff\u610f\u5ea6\u96a8\u4e4b\u63d0\u5347\u3002<\/p>\n<p>\u4f60\u53ef\u4ee5\u4f9d\u64da\u76ee\u6a19\u61c9\u7528\u9700\u6c42\u9078\u64c7\u6700\u4f73\u5316\u8def\u5f91\uff1a<\/p>\n<ul>\n<li>\u5c0d\u65bc\u4e92\u52d5\u5f0f\u804a\u5929\u5de5\u4f5c\u8ca0\u8f09\uff0c\u53ef\u9078\u64c7\u63a8\u6e2c\u89e3\u78bc\u4ee5\u964d\u4f4e\u500b\u5225\u56de\u61c9\u5ef6\u9072\u3002<\/li>\n<li>\u5c0d\u65bc\u9ad8\u4f75\u767c\u4f01\u696d\u7d1a\u5927\u6a21\u578b\u63a8\u7406\u61c9\u7528\uff0c\u53ef\u90e8\u7f72\u89e3\u8026\u67b6\u69cb\uff0c\u5c07\u904b\u7b97\u5bc6\u96c6\u7684\u9810\u586b\u5145\u8207\u53d7\u8a18\u61b6\u9ad4\u9650\u5236\u7684\u89e3\u78bc\u8ca0\u8f09\u62c6\u5206\u81f3\u4e0d\u540c GPU \u53e2\u96c6\u3002<\/li>\n<\/ul>\n<h2><strong>\u5e38\u898b\u554f\u7b54\uff08FAQ\uff09<\/strong><\/h2>\n<h3>\u5982\u4f55\u8a08\u7b97\u63a8\u7406\u4e2d\u7684\u7aef\u5230\u7aef\u7e3d\u5ef6\u9072\uff1f<\/h3>\n<p>\u4f60\u53ef\u4ee5\u5c07\u9996 Token \u5ef6\u9072\uff08TTFT\uff09\u8207\u55ae\u4e00 Token \u6642\u9593\uff08TPOT\uff09\u4e58\u4e0a\u8f38\u51fa Token \u6578\u5f8c\u76f8\u52a0\uff0c\u4ee5\u8a08\u7b97\u7e3d\u5ef6\u9072\u3002\u6b64\u516c\u5f0f\u6709\u52a9\u65bc\u8861\u91cf\u4e92\u52d5\u5f0f\u61c9\u7528\u4e2d\u7684\u6574\u9ad4\u5927\u6a21\u578b\u56de\u61c9\u6027\u3002<\/p>\n<h3>\u9810\u586b\u5145\u8207\u89e3\u78bc\u968e\u6bb5\u5728\u8a08\u7b97\u4e0a\u7684\u6700\u5927\u5dee\u7570\u662f\u4ec0\u9ebc\uff1f<\/h3>\n<p>\u9810\u586b\u5145\u968e\u6bb5\u6703\u540c\u6642\u8655\u7406\u6240\u6709\u8f38\u5165 Token\uff0c\u5f37\u70c8\u4ef0\u8cf4 GPU \u904b\u7b97\u80fd\u529b\uff1b\u89e3\u78bc\u968e\u6bb5\u5247\u4f9d\u5e8f\u751f\u6210\u8f38\u51fa Token\uff0c\u4e3b\u8981\u53d7\u9650\u65bc\u8a18\u61b6\u9ad4\u983b\u5bec\u3002<\/p>\n<h3>\u70ba\u4ec0\u9ebc\u4f01\u696d\u5718\u968a\u61c9\u512a\u5148\u95dc\u6ce8 Goodput\uff0c\u800c\u4e0d\u662f\u539f\u59cb\u541e\u5410\u91cf\uff1f<\/h3>\n<p>\u539f\u59cb\u541e\u5410\u91cf\u53ea\u7d71\u8a08\u751f\u6210\u7684 Token \u6578\uff0c\u4e26\u4e0d\u5340\u5206\u662f\u5426\u903e\u6642\uff1bGoodput \u5247\u53ea\u7d71\u8a08\u7b26\u5408\u670d\u52d9\u7b49\u7d1a\u76ee\u6a19\u7684\u6210\u529f Token\u3002\u56e0\u6b64\uff0c\u76e3\u63a7 Goodput \u80fd\u5728\u9ad8\u6d41\u91cf\u4e0b\u5e6b\u52a9\u4f60\u7dad\u6301\u670d\u52d9\u54c1\u8cea\u3002<\/p>\n<h3>\u9023\u7e8c\u6279\u6b21\u8655\u7406\u5982\u4f55\u63d0\u5347\u751f\u6210\u6548\u7387\uff1f<\/h3>\n<p>\u9023\u7e8c\u6279\u6b21\u8655\u7406\u5728\u6bcf\u500b\u8fed\u4ee3\u6b65\u9a5f\u4e0a\u7ba1\u7406\u8acb\u6c42\u4f47\u5217\u3002\u6b64\u6280\u8853\u85c9\u7531 PagedAttention \u5c07\u8a18\u61b6\u9ad4\u6d6a\u8cbb\u964d\u4f4e\u5230\u7d04 4%\uff0c\u56e0\u6b64\u4f60\u53ef\u4ee5\u7528\u5b83\u4f86\u6700\u4f73\u5316\u5927\u8a9e\u8a00\u6a21\u578b\u7684\u6574\u9ad4\u541e\u5410\u91cf\u3002<\/p>\n<h3>\u63a8\u6e2c\u89e3\u78bc\u8207\u89e3\u8026\u5f0f\u670d\u52d9\u5982\u4f55\u964d\u4f4e\u5ef6\u9072\uff1f<\/h3>\n<p>\u63a8\u6e2c\u89e3\u78bc\u900f\u904e\u5c0f\u578b\u8349\u7a3f\u6a21\u578b\u52a0\u901f\u8f38\u51fa\u751f\u6210\uff1b\u89e3\u8026\u5f0f\u670d\u52d9\u5247\u5c07\u9810\u586b\u5145\u8207\u89e3\u78bc\u62c6\u5206\u81f3\u4e0d\u540c GPU \u53e2\u96c6\u3002\u4f60\u53ef\u4ee5\u4f9d\u7167\u76ee\u6a19\u5927\u6a21\u578b\u6548\u80fd\u6307\u6a19\uff0c\u5728\u5c16\u5cf0\u9700\u6c42\u60c5\u5883\u4e0b\u9078\u64c7\u9019\u4e9b\u6700\u4f73\u5316\u8def\u5f91\uff0c\u4ee5\u7a69\u5b9a\u5ef6\u9072\u8868\u73fe\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u751f\u6210\u5f0f AI \u90e8\u7f72\u4e2d\uff0c\u4f60\u900f\u904e\u4e00\u7cfb\u5217\u95dc\u9375\u6307\u6a19\u4f86\u8861\u91cf\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd\u3002Time to first token\uff08TT [&#8230;]<\/p>\n<p><a class=\"btn btn-secondary understrap-read-more-link\" href=\"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/\">Read More&#8230;<\/a><\/p>\n","protected":false},"author":2,"featured_media":34104,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[8116],"tags":[12833,12834,12835,12836,12698],"class_list":["post-34108","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-japan-dedicated-server-tc","tag-token-generation-efficiency","tag-ttft-tpot-goodput","tag-continuous-batching","tag-speculative-decoding","tag-llm-inference"],"acf":[],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v28.2 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>\u5ea6\u91cf\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd\u2014\u2014\u89e3\u6790 2026 \u5e74\u7684 Token \u751f\u6210\u6548\u7387<\/title>\n<meta name=\"description\" content=\"\u5728\u5927\u6a21\u578b\u63a8\u7406\u4e2d\u6700\u5927\u5316 Token \u751f\u6210\u6548\u7387\u3002\u5206\u6790 TTFT\u3001TPOT \u548c Goodput \u7b49\u95dc\u9375\u6548\u80fd\u6307\u6a19\uff0c\u512a\u5316\u9810\u586b\u5145\u8207\u89e3\u78bc\u968e\u6bb5\u7684\u57f7\u884c\u3002\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/posts\/34108\" \/>\n<meta property=\"og:locale\" content=\"zh_TW\" \/>\n<meta property=\"og:type\" content=\"company\" \/>\n<meta property=\"og:title\" content=\"\u5ea6\u91cf\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd\u2014\u2014\u89e3\u6790 2026 \u5e74\u7684 Token \u751f\u6210\u6548\u7387\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/posts\/34108\" \/>\n<meta property=\"og:site_name\" content=\"\u65b0\u5929\u57df\u4e92\u806f\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-17T04:18:17+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-08-17T04:20:49+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.simcentric.com\/wp-content\/uploads\/2026\/08\/sim-blog-2026-08-17-blogA.png\" \/>\n\t<meta property=\"og:image:width\" content=\"614\" \/>\n\t<meta property=\"og:image:height\" content=\"331\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u5ea6\u91cf\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd\u2014\u2014\u89e3\u6790 2026 \u5e74\u7684 Token \u751f\u6210\u6548\u7387","description":"\u5728\u5927\u6a21\u578b\u63a8\u7406\u4e2d\u6700\u5927\u5316 Token \u751f\u6210\u6548\u7387\u3002\u5206\u6790 TTFT\u3001TPOT \u548c Goodput \u7b49\u95dc\u9375\u6548\u80fd\u6307\u6a19\uff0c\u512a\u5316\u9810\u586b\u5145\u8207\u89e3\u78bc\u968e\u6bb5\u7684\u57f7\u884c\u3002","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:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/posts\/34108","og_locale":"zh_TW","og_type":"company","og_title":"\u5ea6\u91cf\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd\u2014\u2014\u89e3\u6790 2026 \u5e74\u7684 Token \u751f\u6210\u6548\u7387","og_url":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/posts\/34108","og_site_name":"\u65b0\u5929\u57df\u4e92\u806f","article_published_time":"2026-08-17T04:18:17+00:00","article_modified_time":"2026-08-17T04:20:49+00:00","og_image":[{"width":614,"height":331,"url":"https:\/\/www.simcentric.com\/wp-content\/uploads\/2026\/08\/sim-blog-2026-08-17-blogA.png","type":"image\/png"}],"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/#article","isPartOf":{"@id":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/"},"author":{"name":"Debbie Ng","@id":"https:\/\/simcentric.com\/tc\/#\/schema\/person\/0bfc8768eb7caacbfab4a0855d805626"},"headline":"\u5ea6\u91cf\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd\u2014\u2014\u89e3\u6790 2026 \u5e74\u7684 Token \u751f\u6210\u6548\u7387","datePublished":"2026-08-17T04:18:17+00:00","dateModified":"2026-08-17T04:20:49+00:00","mainEntityOfPage":{"@id":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/"},"wordCount":267,"publisher":{"@id":"https:\/\/simcentric.com\/tc\/#organization"},"image":{"@id":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/#primaryimage"},"thumbnailUrl":"https:\/\/www.simcentric.com\/wp-content\/uploads\/2026\/08\/sim-blog-2026-08-17-blogA.png","keywords":["Token \u751f\u6210\u6548\u7387","\u9996 Token \u5ef6\u9072 TTFT","\u9023\u7e8c\u6279\u6b21\u8655\u7406","\u63a8\u6e2c\u89e3\u78bc","LLM\u63a8\u8ad6"],"articleSection":["\u65e5\u672c\u4f3a\u670d\u5668"],"inLanguage":"zh-HK"},{"@type":"WebPage","@id":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/","url":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/","name":"\u5ea6\u91cf\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd\u2014\u2014\u89e3\u6790 2026 \u5e74\u7684 Token \u751f\u6210\u6548\u7387","isPartOf":{"@id":"https:\/\/simcentric.com\/tc\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/#primaryimage"},"image":{"@id":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/#primaryimage"},"thumbnailUrl":"https:\/\/www.simcentric.com\/wp-content\/uploads\/2026\/08\/sim-blog-2026-08-17-blogA.png","datePublished":"2026-08-17T04:18:17+00:00","dateModified":"2026-08-17T04:20:49+00:00","description":"\u5728\u5927\u6a21\u578b\u63a8\u7406\u4e2d\u6700\u5927\u5316 Token \u751f\u6210\u6548\u7387\u3002\u5206\u6790 TTFT\u3001TPOT \u548c Goodput \u7b49\u95dc\u9375\u6548\u80fd\u6307\u6a19\uff0c\u512a\u5316\u9810\u586b\u5145\u8207\u89e3\u78bc\u968e\u6bb5\u7684\u57f7\u884c\u3002","breadcrumb":{"@id":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/#breadcrumb"},"inLanguage":"zh-HK","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/"]}]},{"@type":"ImageObject","inLanguage":"zh-HK","@id":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/#primaryimage","url":"https:\/\/www.simcentric.com\/wp-content\/uploads\/2026\/08\/sim-blog-2026-08-17-blogA.png","contentUrl":"https:\/\/www.simcentric.com\/wp-content\/uploads\/2026\/08\/sim-blog-2026-08-17-blogA.png","width":614,"height":331,"caption":"2026 \u5927\u6a21\u578b\u63a8\u7406\u8207 Token \u751f\u6210\u6548\u7387\u793a\u610f\u5716"},{"@type":"BreadcrumbList","@id":"https:\/\/www.simcentric.com\/tc\/japan-dedicated-server-tc\/measuring-llm-performance-analyzing-token-efficiency-in-2026\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.simcentric.com\/tc\/"},{"@type":"ListItem","position":2,"name":"\u5ea6\u91cf\u5927\u6a21\u578b\u63a8\u7406\u6548\u80fd\u2014\u2014\u89e3\u6790 2026 \u5e74\u7684 Token \u751f\u6210\u6548\u7387"}]},{"@type":"WebSite","@id":"https:\/\/simcentric.com\/tc\/#website","url":"https:\/\/simcentric.com\/tc\/","name":"Simcentric Solutions","description":"","publisher":{"@id":"https:\/\/simcentric.com\/tc\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/simcentric.com\/tc\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-HK"},{"@type":"Organization","@id":"https:\/\/simcentric.com\/tc\/#organization","name":"Simcentric Solutions","url":"https:\/\/simcentric.com\/tc\/","logo":{"@type":"ImageObject","inLanguage":"zh-HK","@id":"https:\/\/simcentric.com\/tc\/#\/schema\/logo\/image\/","url":"https:\/\/www.simcentric.com\/wp-content\/uploads\/2023\/06\/sim-logo-2023.png","contentUrl":"https:\/\/www.simcentric.com\/wp-content\/uploads\/2023\/06\/sim-logo-2023.png","width":800,"height":222,"caption":"Simcentric Solutions"},"image":{"@id":"https:\/\/simcentric.com\/tc\/#\/schema\/logo\/image\/"}},{"@type":"Person","@id":"https:\/\/simcentric.com\/tc\/#\/schema\/person\/0bfc8768eb7caacbfab4a0855d805626","name":"Debbie Ng","image":{"@type":"ImageObject","inLanguage":"zh-HK","@id":"https:\/\/secure.gravatar.com\/avatar\/9f3420f60d27329ed6e921ddaae5206353d0a7414d1abb6f1f0b32b4fd849965?s=96&d=mm&r=g","url":"https:\/\/secure.gravatar.com\/avatar\/9f3420f60d27329ed6e921ddaae5206353d0a7414d1abb6f1f0b32b4fd849965?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/9f3420f60d27329ed6e921ddaae5206353d0a7414d1abb6f1f0b32b4fd849965?s=96&d=mm&r=g","caption":"Debbie Ng"}}]}},"_links":{"self":[{"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/posts\/34108","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/comments?post=34108"}],"version-history":[{"count":1,"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/posts\/34108\/revisions"}],"predecessor-version":[{"id":34113,"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/posts\/34108\/revisions\/34113"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/media\/34104"}],"wp:attachment":[{"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/media?parent=34108"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/categories?post=34108"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.simcentric.com\/tc\/wp-json\/wp\/v2\/tags?post=34108"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}