[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1w9yihhgl600t":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":15,"bodyZh":19,"bodyEn":20,"sourceName":21,"sourceUrl":22,"updatedAt":23},"dual-3090-qengine132-power-capped-bench-2026-11-01","HARDWARE","2026-10-31T23:05:53.272Z","双 RTX 3090 新自报：220W 限功耗下的速度，不能当默认配置基线","Dual RTX 3090 report: 220 W caps are part of the result","社区公开 QEngine 1.3.2、Qwen3.8-27B INT4 的双卡报告，列出限功耗、NVLink、上下文及分阶段指标。本站按社区L0对待，不冒充独立复现，不用短探针代替稳态速度。","A community report documents dual RTX 3090 inference with QEngine 1.3.2, explicit power caps and phase-specific measurements. It remains L0 self-reported evidence, not independent AraoAI replication.",[12],{"id":13,"name":14},5,"Qwen3.8-27B",[16],{"id":17,"name":18},8,"2x NVIDIA RTX 3090","同样写“双3090”，测试条件可能完全不同。10月31日新发布的 club-3090 issue #1897 给出一份 Qwen3.8-27B INT4、QEngine 1.3.2 报告：两张 RTX 3090 都限制在220W，启用 NVLink，系统为 Ubuntu 24.04.5 LTS。报告生成时间为当日17:51 UTC，位于本次24小时窗口内。以下数字都是该作者自报，本站未独立复现。\n\n一、先读条件，再读速度。原文列出驱动615.71.10、FP8 KV、262144上下文上限和采样设置，两卡 PCIe 实际协商均为 Gen4 x8。220W 是这次运行的功率上限，不是实测平均耗电，更不是显卡型号的额定功耗。卡间有 NVLink 也不代表所有计算阶段都走同一条通信路径，不能据此推导其他机器的速度。\n\n二、不同阶段的数字不能拼。报告中的10K输入阶段，三个样本的 prefill 均值为2363.43 tok\u002Fs，同阶段 decode 为106.44 tok\u002Fs；90K阶段只有一个样本，分别为1920.08和117.68 tok\u002Fs。两阶段输出预算均只有16 token，后者不能当成长回答稳态速度。叙事和代码阶段又是不同提示，不应抽一列最快的 decode，再搭另一阶段的 prefill 做成万能成绩。\n\n三、保留会限制结论的信息。自报包含 MTP 草稿相关配置，不能直接套“显存带宽除权重大小”的普通稠密单步上界判假。报告还注明部分主板与内存信息因非root运行未采集。这些空缺应当留着，不从相邻帖子、显卡容量或引擎配置目标反推实测值。\n\n这份报告的用途，是提供可追溯的复现条件，而不是证明双3090优于其他显卡。本站仅按L0证据整理对应阶段记录，正式入库仍进草稿待审。另需注意，QEngine 的许可证为非商业许可，不能把权重可下载或引擎可运行理解为商业使用已获授权。\n\n来源：https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F1897\n许可：https:\u002F\u002Fgithub.com\u002Fbpx03\u002FQEngine\u002Fblob\u002Fmain\u002FLICENSE","Two reports labeled “dual 3090” need not describe comparable configurations. A new club-3090 issue #1897 documents Qwen3.8-27B INT4 on QEngine 1.3.2 with two RTX 3090 cards capped at 220 W each, NVLink, and Ubuntu 24.04.5 LTS. The report was generated on October 31 at 17:51 UTC, inside this window. Every measurement below is the author's self-report, not an independent AraoAI reproduction.\n\n1. Read conditions first. The report records driver 615.71.10, FP8 KV, a 262144-token serving limit, sampling settings and negotiated PCIe Gen4 x8 links. A 220 W cap is neither measured average power nor the card's rated TDP. NVLink presence does not establish which transport every execution phase uses.\n\n2. Keep phases separate. At the 10K-input phase, three samples give mean prefill of 2363.43 tok\u002Fs and same-phase decode of 106.44 tok\u002Fs. At the 90K phase, one sample gives 1920.08 and 117.68 tok\u002Fs respectively. Both phases allow only 16 output tokens; these are short depth probes, not long-response steady-state rates. Narrative and code prompts are different phases and must not be spliced into a universal result.\n\n3. Preserve limitations. MTP-related draft configuration invalidates the ordinary one-dense-pass bandwidth bound. Some motherboard and memory details were not collected without root. Missing measurements must not be inferred from another post, nominal card capacity or engine allocation targets.\n\nThis report supplies traceable reproduction conditions, not a GPU ranking. AraoAI treats it as L0 evidence; record imports remain drafts pending review. QEngine carries a noncommercial license, so availability is not commercial authorization.\n\nSource: https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F1897\nLicense: https:\u002F\u002Fgithub.com\u002Fbpx03\u002FQEngine\u002Fblob\u002Fmain\u002FLICENSE","club-3090 community self-report (L0)","https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F1897","2026-10-31T23:05:53.273Z"]