Pull requests / #489

#489 bench/results: RTX A3000 12GB laptop + i7-12850HX (sm_86): hugepages, PCIe probe, calibration

closed · @yannickloth · 0 Kommentare · Auf GitHub

BenchmarksNVIDIA / CUDAModels & quantsDocumentation

Beschreibung

A results entry for a 12 GB Ampere laptop part with a 40 GiB CPU-side expert arena (engine 0.1.34, Qwen3.8-Flash-Next Swift 1.5 IQ3_XXS native pack).

- **Hugepages are all-or-nothing.** A 34 GiB `vm.nr_hugepages` silently gave 4 KiB pages for the 39.97 GiB arena; 44 GiB gives `hugetlb 2 MB pages`. Unlimited `RLIMIT_MEMLOCK` is required too (the systemd user manager caps units at 8 MiB).
- **The PCIe probe read this x16 Gen4 link at 5.8-18.5 GB/s across starts** (#485), which drove the calibration to `--pcie-frac 0.35` over the 0.55 default.
- `tools/calibrate.py` keeps `--pcie-frac 0.35`, `--spec-min-p 0.70`, `--pool-workers 8` (8 P + 8 E). Raw output in `calibrate.json`; decode numbers and their caveats are in the README.

Hardware/OS/engine details only; no usernames, hostnames or paths.

Mehr auf der Site

Links zu Install, Modellen, Releases.