← News
Hardware

New dual-3090 QEngine report separates short prompts, deep context and whole replies

A November 19 report covers QEngine 1.3.6-vision on two 3090s capped at 260 W each. Paired 10K/90K measurements are short-output probes, while whole-reply TTFT includes generation. This remains a community self-report.

Source: club-3090 community report / maintainer clarification

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.

  1. Pair measurements only within a test stage. The 10K stage reports mean prefill of 4366.41 tok/s and decode at depth of 139.31 tok/s across three measurements. At 90K, the corresponding values are 5443.05 and 170.80 tok/s 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.

  2. Short-prompt results cannot fill another stage. Narrative and code tasks each have five measurements, but their PP tok/s is explicitly n/a. 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.

  3. 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.

Experts 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.

Report and maintainer clarification: https://github.com/noonghunna/club-3090/issues/2103