[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fwz5wanta3su2":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},"dual3090-qengine136-depth-paired-measurements-2026-11-20","HARDWARE","2026-11-19T23:07:55.837Z","双3090的QEngine新报告：短提示词、深上下文和整段输出必须分开读","New dual-3090 QEngine report separates short prompts, deep context and whole replies","11月19日生成的社区报告记录QEngine 1.3.6-vision、双3090与260W限功耗。10K与90K阶段提供成对速度，但短输出探针不是长输出稳态；整段返回的TTFT还包含生成时间。","A November 19 report covers QEngine 1.3.6-vision on two 3090s capped at 260 W each. Paired 10K\u002F90K measurements are short-output probes, while whole-reply TTFT includes generation. This remains a community self-report.",[12],{"id":13,"name":14},11,"Qwen3.8-Flash-Next",[16],{"id":17,"name":18},8,"2x NVIDIA RTX 3090","11月19日，club-3090新增#2103社区报告，原文生成时间为当天08:21:24 UTC。配置包括Ubuntu 24.04.5 LTS、QEngine镜像1.3.6-vision，以及运行Qwen3.8-Flash-Next-QE-INT4的两张RTX 3090。两卡均限功耗260W，使用NVLink及打过P2P补丁的驱动；这不是原厂功耗或原样驱动的通用基线。本站只核对公开原文，尚未独立复现。\n\n1. 同一阶段的数值才能配对。报告10K阶段的prefill均值为4366.41 tok\u002Fs，对应深度decode均值139.31 tok\u002Fs，三次测量；90K阶段为5443.05和170.80 tok\u002Fs，仅一次测量，实际输入89998 tokens。这些阶段的测量输出预算是128 tokens，属于深上下文短探针，不是长输出稳态吞吐。两组速度可以分别保留，不能混成一条平均配置。\n\n2. 漂亮的短提示词数字不能填补另一阶段。原文另列叙事和代码任务，各测五次，但其PP tok\u002Fs明确为n\u002Fa。维护者评论也说明短提示词不足以给出可靠prefill。把短任务decode与10K或90K的prefill拼成一条完整跑分，看似字段齐全，实际改变了实验条件。\n\n3. 整段输出会改变TTFT含义。持续测试的25次回复均为单个输出块；报告使用引擎计数器，而不是客户端流式窗口计算decode。其TTFT包含整段回复生成时间，不能直接当成纯首token延迟。报告的PASS仅表示该样本未出现失败信号，不等于对驱动补丁、任意长上下文任务或模型能力作出全面验证。\n\n还有两个边界必须随数字保留：专家权重有主机内存承载，不能套用全权重常驻显存的稠密模型速度公式；机器还装有第三张5070 Ti，但跑分指纹明确只测GPU 0、1，整机显存读数不能自动算成双3090的阶段峰值。这里有价值的不是把一个速度写进显卡排名，而是能追到同阶段、同条件的原始记录。\n\n原文与维护者说明：https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F2103","On November 19, club-3090 received community report #2103, generated at 08:21:24 UTC that day. It documents Ubuntu 24.04.5 LTS, QEngine image 1.3.6-vision and Qwen3.8-Flash-Next-QE-INT4 served on two RTX 3090s, each capped at 260 W, with NVLink and a P2P-patched driver. This is not a stock-power, unmodified-driver baseline. AraoAI checked the public source but has not independently reproduced it.\n\n1. Pair measurements only within a test stage. The 10K stage reports mean prefill of 4366.41 tok\u002Fs and decode at depth of 139.31 tok\u002Fs across three measurements. At 90K, the corresponding values are 5443.05 and 170.80 tok\u002Fs from one measurement, with 89998 actual input tokens. Both measured stages budget 128 output tokens: they are short-output deep-context probes, not sustained long-generation throughput. Keep them as distinct configurations.\n\n2. Short-prompt results cannot fill another stage. Narrative and code tasks each have five measurements, but their PP tok\u002Fs is explicitly n\u002Fa. The maintainer explains that these short prompts cannot establish meaningful prefill rates. Combining their decode with the 10K or 90K prefill would silently change experimental conditions.\n\n3. Whole replies change the meaning of TTFT. All 25 soak replies arrived in one output chunk. Decode uses engine counters rather than a client streaming window, and TTFT includes generation of the entire reply. PASS means no failure signal on this sample, not comprehensive validation of the driver patch or model capability.\n\nExperts are partly host-backed, so a dense, fully GPU-resident bandwidth ceiling is inappropriate. A third 5070 Ti is installed, but the benchmark fingerprint explicitly measures GPUs 0 and 1; whole-machine VRAM cannot be assigned to the dual-3090 stage. The useful artifact is a traceable, condition-matched record, not a GPU ranking.\n\nReport and maintainer clarification: https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F2103","club-3090 community report \u002F maintainer clarification","https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F2103","2026-11-19T23:07:55.838Z"]