[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fd4pbgggn1fka":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":15,"bodyZh":16,"bodyEn":17,"sourceName":18,"sourceUrl":19,"updatedAt":6},"four5080-qengine1359-report-measurement-boundaries-2026-11-15","ARENA","2026-11-14T23:07:06.853Z","四张5080的Flash-Next新报告：448.60 tok\u002Fs不能脱离功耗与输出方式","A four-5080 Flash-Next report: 448.60 tok\u002Fs needs power and output conditions","社区公开QEngine 1.3.5.9四卡报告，明确限功耗、主机卸载与整块回复的计时限制。权重量化缺明确原文，不为凑字段导入跑分。","A QEngine 1.3.5.9 community report states power caps, host offload and whole-reply timing limits. Weight quantization is not explicit in the retrieved report, so no benchmark row is imported.",[12],{"id":13,"name":14},11,"Qwen3.8-Flash-Next",[],"2026年11月14日12:20 UTC，club-3090 issue #2033公开四张RTX 5080运行Qwen3.8-Flash-Next的完整报告，原始报告生成于11:22:11 UTC，均落在本次观察窗。作者使用QEngine 1.3.5.9、构建ee98bd1，报告code decode均值448.60、narrative均值366.93 tok\u002Fs，每种形状5次测量。这是一份社区自报，不是本站复现。\n\n第一，四张卡不是唯一条件。每卡通过PCIe Gen5 x8交换网络连接，没有NVLink；GPU0限252W，其余限250W。报告说明部分专家由主机内存承载，并启用MTP与n-gram查找，KV格式为fp8_e4m3。这些条件不能删掉，再拿数字与默认功耗、纯显存常驻或关闭投机的记录排高低。主机承载规模也不是实测显存占用。\n\n第二，能寻址大上下文，不等于记得住。报告的ctx为1048576，但上下文边界探测、压力、持续运行检查和能力题集回答的是不同问题。作者另列思考开关两组能力测试，并明确采样标记仍为sampling=server。维护者确认新修复不会追溯改写旧运行的标记，因此不能把这些分数写成受控的思考模式因果比较。\n\n第三，TTFT也可能混入整段生成。该报告的12轮工具回复全部一次性到达；客户端第一块出现时，整段回复已生成。维护者随后合并PR #2035，取消对此类曲线作提示处理增长评级，但保留原始数字。不能把曲线上升直接解释成读取长提示变慢，更不能从块间传输时间反算decode。\n\n本站保留原文与缺项，不将这份报告强凑成标准跑分记录：已取回正文和完整评论，但未见明确的权重量化档。KV精度与附件文件名都不能代替权重量化。一个条件完整、可复查的数字，才有资格进入后续比较。\n\n来源：https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F2033\n计时修正：https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F2035","At 12:20 UTC on November 14, 2026, club-3090 issue #2033 published a four-RTX-5080 Qwen3.8-Flash-Next report generated at 11:22:11 UTC, within this observation window. QEngine 1.3.5.9, build ee98bd1, reportedly achieved mean decode rates of 448.60 tok\u002Fs for code and 366.93 for narrative over five measured runs per shape. This is one contributor's report, not an AraoAI reproduction.\n\nFour cards alone do not identify the configuration. They use PCIe Gen5 x8 through a switch without NVLink. GPU0 is capped at 252W and the others at 250W. Experts are partly backed by host memory; MTP and n-gram lookup are enabled, and KV uses fp8_e4m3. These conditions prevent direct ranking against stock-power, all-VRAM or non-speculative configurations. A host-backing amount is not a measured VRAM footprint.\n\nAddressable context is not recall quality. The configured context is 1048576, but boundary probes, stress checks, soak runs and ability tasks answer different questions. Separate thinking-on and thinking-off quality runs retain sampling=server stamps. A maintainer confirms that a later sampling-stamp fix cannot retroactively verify the old runs. Their scores do not establish a controlled causal benefit from thinking mode.\n\nAll twelve tool replies in the agentic run arrived in one output chunk. The reported first-chunk time therefore includes generation of the entire reply. Subsequent PR #2035 removes prompt-read growth grades for these cases while preserving measurements. Do not interpret the curve as prompt processing alone or infer decode from transmission gaps.\n\nAraoAI retains the source and its missing fields rather than forcing a benchmark import. The retrieved issue and full comments do not explicitly state weight quantization. KV precision and attachment filenames cannot fill that field. A comparison starts with complete, auditable conditions.\n\nSource: https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F2033\nTiming correction: https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F2035","club-3090 社区报告及维护者修正","https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fissues\u002F2033"]