[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fnvyixblcg8l5":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},"qengine1342-dual3090-host-ram-unit-correction-2026-11-03","HARDWARE","2026-11-02T23:04:48.807Z","双 3090 社区新报告：主机内存门槛改为 94GB，别把 GiB 和 GB 混用","New dual-3090 report corrects host RAM guidance to 94 GB: GiB and GB are not interchangeable","club-3090 报告 QEngine 1.3.4.2 与新权重的同机测试，并修正内存单位。90K 阶段仅测一次，版本、权重、槽位和视觉设置同时变化，不能单因归功。","A club-3090 report covers QEngine 1.3.4.2 and new weights, correcting RAM units. The 90K phase has one sample and several settings changed, so causal speedup claims are not justified.",[12],{"id":13,"name":14},11,"Qwen3.8-Flash-Next",[16],{"id":17,"name":18},8,"2x NVIDIA RTX 3090","11 月 2 日 21:40 UTC，club-3090 新增并随后合并一份双 RTX 3090 报告：Qwen3.8-Flash-Next 使用 QEngine 1.3.4.2、新默认权重 c0022018、8 个槽位、视觉开启、FLASH_COMM=auto，两卡各限功耗 230W。以下均为该仓维护者自报，不是 AraoAI 独立复现；报告只写测量日期，不能据发帖时间替代精确测量起止。\n\n1. 先看最实用的修正。配置的主机内存门槛由 88 改为 94GB。报告解释，旧注释曾把 GiB 读数和十进制 GB 门槛直接比较；这次启动就绪时 MemAvailable 减少 93.5GB，整轮测试后减少 104.7GB。这是系统可用内存下降量，不等于某个进程的私有占用，也不等于显存。买机器时不能只看两张卡能否装下权重。\n\n2. 速度必须连同阶段看。叙事与代码生成分别自报 165.52、158.36 token\u002Fs；10K 与 90K 预填充分别为 1879、3615 token\u002Fs。它们属于不同测试阶段，不能把叙事生成和 90K 预填充拼成一条同条件记录。90K 预填充及该深度生成均只有一次测量，暂不足以说明稳定收益。\n\n3. 对比不等于归因。旧轮使用另一引擎版本、另一份权重、16 槽位且未开视觉。作者明确不把差异归给单一原因；我们也不把长上下文数字包装成版本全面提速。仓库配置同时警告 QEngine 为 PolyForm Noncommercial 1.0.0，商用前须核对许可，能运行不等于能商用。\n\n选型落点不是追一个最高速度，而是同时核实主机内存、显存余量、并发槽位和许可。本条可作为新的社区观察，但尚未成为站内已验证的独立复现记录；缺操作系统与精确测量时间等信息时，先保留候选，不补猜测值。\n\n来源：[报告 PR #1914](https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F1914)，含配置变更、测试阶段及单位修正。","At 21:40 UTC on November 2, club-3090 opened a report subsequently merged for Qwen3.8-Flash-Next on two RTX 3090 cards: QEngine 1.3.4.2, new default weights c0022018, eight slots, vision enabled, FLASH_COMM=auto and a 230 W power limit per card. This is the repository maintainer's self-report, not an independent AraoAI reproduction. Only a measurement date is given, so publication time cannot establish exact measurement boundaries.\n\n1. The practical correction is RAM guidance. The configuration changes its host-RAM threshold from 88 to 94 decimal GB. Earlier notes compared GiB readings directly with a GB threshold. The reported MemAvailable decrease is 93.5 GB at ready and 104.7 GB after the benchmark. This is a system-wide change in available RAM, not private process memory or GPU memory. Fitting the weights on two cards is not the whole hardware requirement.\n\n2. Keep phases separate. Narrative and code decode are reported as 165.52 and 158.36 tokens\u002Fs. Prefill at 10K and 90K is 1879 and 3615 tokens\u002Fs. These are different phases; pairing narrative decode with 90K prefill would invent a single configuration result. The 90K measurements have one sample each and are not evidence of stable gains.\n\n3. Comparison is not attribution. The earlier run used another engine version and weights, sixteen slots and no vision. The author explicitly avoids a single-cause explanation. We likewise do not call this a universal version speedup. The repository configuration also flags PolyForm Noncommercial 1.0.0 for QEngine: verify licensing before commercial deployment.\n\nFor selection, check host RAM, GPU headroom, slot count and licensing together. This is a new community observation, not a verified independent reproduction in our database. Missing OS and precise measurement-time evidence remain gaps rather than guessed fields.\n\nSource: [report PR #1914](https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F1914), including configuration diffs and the RAM-unit correction.","club-3090 maintainer self-report","https:\u002F\u002Fgithub.com\u002Fnoonghunna\u002Fclub-3090\u002Fpull\u002F1914","2026-11-02T23:04:48.808Z"]