New four-RTX-5080 report: disabling P2P may fall back to host staging, not the copy engine
club-3090 #1952 reports an on/off P2P comparison on one machine. The author identifies host-staged transfers and explicitly did not test forced copy-engine transport. This is one community report, not a universal GPU ranking.
Source: club-3090 社区自报 / QEngine LICENSE
club-3090 #1952, published at 14:29 UTC on November 5, presents a P2P on/off comparison using QEngine 1.3.4.3 and four RTX 5080 GPUs. Its useful contribution is the logged fallback transport, not merely another throughput result. This is one community self-report, not independently reproduced by AraoAI. Exact measurement start and end times are absent, so we report the newly published evidence without importing it as measurements made inside this daily window.
-
Conditions matter. The machine runs Ubuntu 26.04.1 with four 16GB cards behind a PCIe Gen5 x8 switch. Qwen3.8-Flash-Next INT4 uses fixed tensor parallelism, six configured seats, host-RAM KV and thinking disabled. The author interleaves two enabled and two disabled boots. This differs from the engine default pipeline mode for four cards. Six configured seats do not establish six simultaneous benchmark requests.
-
A decline and an increase use different denominators. The reported two-boot means show that disabling P2P reduced 10K-prompt prefill by 32.4% and narrative decode by 14.7% relative to enabled P2P. These figures apply to this rig and these phases. They are not an equal-percentage gain in the reverse direction, and narrative decode must not be paired with a separate 10K stress-test prefill to manufacture one workload record.
-
Transport logs are decisive. The disabled arm stages collectives through pinned host memory, rather than automatically switching to the copy engine. FLASH_COMM=ce was not tested, so the report does not establish that P2P beats every copy-engine implementation. The 90K prefill leg has only one measurement per boot.
Check negotiated links, topology and actual transport logs before changing switches. QEngine currently uses PolyForm Noncommercial 1.0.0; commercial deployment is not automatically permitted. This report helps design an experiment, but cannot validate another machine or justify a four-GPU purchase by itself.
Sources: https://github.com/noonghunna/club-3090/issues/1952 ; https://github.com/bpx03/QEngine/blob/main/LICENSE