[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f2ljqr3i2k2x6p":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":18,"bodyZh":22,"bodyEn":23,"sourceName":24,"sourceUrl":25,"updatedAt":26},"rtx5080-power-cap-sweeps-corrections-2026-11-16","HARDWARE","2026-11-15T23:09:00.079Z","四张5080的功耗扫描：先分清功耗上限和实际耗电，再谈省电","Four RTX 5080 power-cap sweeps: distinguish the configured limit from actual draw","QEngine 1.3.5.9社区报告给出逐卡扫描及公开更正。该机的额外功耗额度未带来稳定吞吐收益，但不能据此推算所有5080或整机电费。","A QEngine 1.3.5.9 community report includes per-card sweeps and explicit corrections. Higher caps did not produce a consistent throughput gain on this rig; it does not establish universal RTX 5080 savings.",[12,15],{"id":13,"name":14},5,"Qwen3.8-27B",{"id":16,"name":17},11,"Qwen3.8-Flash-Next",[19],{"id":20,"name":21},14,"NVIDIA RTX 5080","把功耗上限调低，不代表电费就按相同比例下降。11月15日UTC，club-3090 issue #2041发布四张RTX 5080的逐卡扫描，同日补充MoE结果并公开纠正口径。它是一台机器的一份社区报告，不是本站独立复现，但恰好说明为什么不能只抄一行“推荐瓦数”。\n\n报告锁定QEngine 1.3.5.9、构建ee98bd1，四张16GB卡经PCIe Gen5 x8交换芯片互联。稠密模型为Qwen3.8-27B-INT4、TP4、上下文262144、FP8 KV驻显存；补测Flash-Next使用另一套并发与主机内存KV布局。两套结果只能分别解释，不能合成同条件对照。扫描按卡执行，10W步进，工具每档重复测量，不能读成四卡同时降到同一档的整机实验。\n\n更值得看的是维护者追问后的更正：GPU0合法范围252—360W，GPU1与GPU2为250—396W，GPU3是250—360W，原叙述把GPU3下限写成252W是错的。GPU0实际扫描从260W起，GPU3从250W起。同一个5080型号，合法范围也不能一概而论；作者同时撤回了“该范围内发现拐点”的说法。\n\n现有扫描支持的窄结论是：这台机器、这个引擎和相应负载下，提高上限没有稳定换来更多吞吐。特别是MoE补测，实际功耗低于设定上限，不能把“允许用更多电”当成“真的多用了电”，也不能反过来把降上限幅度写成节电幅度。报告没有墙插电表数据，本文不计算整机节能比例。\n\n可复制的方法是先查每张卡的合法范围，再同时记录实际功耗、吞吐和限功耗状态，最后测自己的任务。保留更正比保留漂亮结论更重要。原始表格及更正链：https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F2041","Lowering a configured power limit does not reduce electricity use by the same percentage. On 15 November UTC, club-3090 issue #2041 published per-card sweeps on a four-RTX-5080 machine, followed by MoE results and explicit corrections. This is one community report from one rig, not an independent AraoAI reproduction.\n\nThe setup pins QEngine 1.3.5.9, build ee98bd1, with four 16GB cards connected through a PCIe Gen5 x8 switch. The dense workload uses Qwen3.8-27B INT4, TP4, a 262144 context limit and FP8 KV in VRAM. The Flash-Next follow-up uses different concurrency and host-RAM KV placement. Those workloads must be interpreted separately. The tool sweeps one card at a time in 10W steps with repeated runs; this is not a simultaneous whole-rig cap experiment.\n\nThe correction chain matters. GPU0 permits 252–360W, GPUs 1 and 2 permit 250–396W, and GPU3 permits 250–360W. The original prose incorrectly grouped GPU3 with GPU0 at a 252W minimum. GPU0's sweep actually starts at 260W, while GPU3 starts at 250W. The author also withdrew the claim that a knee had been observed within the reachable range. A shared product name is not enough to assume identical per-card limits.\n\nThe narrow finding is that increasing the cap did not consistently buy throughput for these workloads on this machine and engine. In the MoE follow-up, actual draw remained below the configured limits. Permission to consume more power is not measured consumption; conversely, a cap reduction is not a measured saving. No wall-socket energy measurement was supplied, so no whole-system saving percentage is calculated here.\n\nA useful local procedure is to read each card's legal limits, then record actual draw, throughput and power-throttling status together on your own workload. Preserve the correction history rather than copying an attractive recommended wattage. Original tables and corrections: https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F2041","club-3090社区功耗扫描及更正评论","https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F2041","2026-11-15T23:09:00.080Z"]