😱炸了！智谱Zai创始人唐杰在清华课堂硬核算账：大模型尽头是算账！揭秘GLM系列模型跑在纯国产芯片是怎么“玩得起”的？！
This is a dev post classified by AGK Intelligence as Other (a tutorial), kept by the Dev Radar because it carries real work, not commentary.
😱炸了！智谱Zai创始人唐杰在清华课堂硬核算账：大模型尽头是算账！揭秘GLM系列模型跑在纯国产芯片是怎么“玩得起”的？！ 智谱唐杰在清华课上把“Scaling to be used”摊开讲：算法再巧妙，也得跑起来、玩得起，整节课就是一笔经济账，结合GLM真实经历，把推理成本与国产卡实战讲透。 关键细节： 1️⃣GPU按小时收费，利用率10% = 十倍价格 卡再贵，闲着就是浪费，实际成本直接翻倍。 2️⃣GLM 5.0发布当天量超指数、直接停售 过年把训练卡全拿去推理，年后做研究发现一张卡都没有，半年持续缺卡。 3️⃣1000 TPS超快版：一台机器只能服务一个人 有人说多贵都要，唐杰直说那得百万级一年，最后降到两三百TPS换并发。 4️⃣国产卡实战：底线40 TPS/用户，硬优化到70 现在的Flash就跑在国产卡上，从能跑到能打全是工程硬刚。 5️⃣推理贵的本质：Decode卡的是带宽，不是算力 Prefill能打满GPU，Decode一个token一个token蹦，大部分时间在等数据，H100每读1 byte要做约295次计算才不亏，新卡算力涨得比带宽快。 6️⃣降本两大招 ◦压KV Cache：DeepSeek V3每token七万多byte，V4.1 Flash压到890 byte。 ◦投机解码：小模型先猜一串，大模型一次性验证，用闲着的算力白捡。 7️⃣A
Posted by SuSu_酥酥👅 (25k followers) 3 h ago · 17 likes · 3.3k views · view the original post on X. Kept by the Dev Radar as Other.
More dev work like this
- took @rehan_shei's open-source pixel fx into a real 3D scene and made the wizard… — @majidmanzarpour
- 彭博终端贵得让个人用户直摇头，这感受你应该不陌生。 — @bkdgiffug
- I went quiet for a few weeks because what I’m building now is better than every library… — @dwhitedesign
- Celestia’s Voyager handled 21M transactions per second in its first big test — @BSCNews
- Our annual kiroween hackathon is here, and submissions are officially open. 🎃 — @kirodotdev
- Em paralelo ao projeto do Launcher de Colin McRae (que tá funcionando bonitinho,… — @AkitaOnRails
- Ano passado brinquei de 4K upscaling com Real ERSGAN. Bom ver o claudinho automatizando… — @AkitaOnRails
- Anchorpoint's HKDAP jumped from 16.1M to about 103M on Ethereum today. 3 mints, 2 of… — @envio_indexer
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.2k posts from 5k X accounts over the last 21 days, 8.8k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-10-30 20:27 UTC. Full method.