改动 / #3076
#3076 bench: RTX 5060 Ti + RTX 4080 SUPER community results, before and after calibration
open · @Runedia · 0 评论 · 去 GitHub 看
BenchmarksSetup & installNVIDIA / CUDAModels & quantsWindows
说明
Two community benchmark reports from one Windows 11 machine (Ryzen 9 9950X3D, 61.6 GiB RAM, RTX 5060 Ti 16 GB + RTX 4080 SUPER 16 GB, layer split auto), measured the same day with `tools/strata_bench.py`. Results only; no engine changes. - Strata 0.1.45 (`477ac13c`), Flash-Next IQ2_XS, 262,144-token context - Run 1: fresh-install settings. Run 2: after `START-HERE.bat --calibrate` - Engine arguments that differ (from `engine_info` in both results.json): `--pcie-frac` 0.55 -> 0.90, `--spec-min-p` 0.50 -> 0.70, `--prefill-pipe` 0 -> 3072 | Median | Run 1 | Run 2 | | --- | ---: | ---: | | short chat decode tok/s | 45.2 | 80.4 | | 4K prompt tok/s | 750.1 | 1150.2 | | 16K prompt tok/s | 705.7 | 1707.5 | | agent session follow-up decode tok/s | 44.3 | 72.6 | Limitations: speed only, no correctness check. Not an interleaved A/B: calibration restarted the engine several times before run 2, so OS file cache and background load probably differed. Each run was done once. PCIe is x8 per GPU because lanes are shared with NVMe drives (idle link read only for the 5060 Ti).
本站相关内容
相关页面的快捷入口。