#850 bench: community report, Tesla V100 32 GB + RTX 4070 layer split with 16 GB of RAM (IQ2_XS)

描述

Community benchmark report: the full Qwen3.8-Flash-Next **IQ2_XS** split across a **Tesla V100 32 GB (sm_70, PCIe 3.0 x4) and an
RTX 4070 12 GB (sm_89)**, in both card orders, on a PC with **16 GB of RAM**. In `bench/results/2026-10-04-community-v100-rtx4070-split/`.
Results only, no code changes (companion to #823, the V100 alone).

- **Build:** engine 0.1.39 from source with `--cuda 12`, `"archs": [70, 89]`.
- **Configuration:** `"layer_split": "auto"`, low-RAM mode (`--mmap-experts`), context 131,072, KV int8, MTP `--spec 4`, no vision.
  Auto chose K=39 with the V100 first and K=11 with the 4070 first.
- **Method:** the 2x MI50 report's `benchmark.py`, unchanged (3 runs each at 4,096 / 32,768 / 128,000 fresh prompt tokens,
  256-token cap, reasoning none) + needles 32k/128k x 10/50/90.
- **Results (medians), V100 first:** decode 84.9 / 81.5 / 75.1 tok/s, prompt 996 / 1,602 / 1,552 tok/s. 4070 first: decode
  79.1 / 83.8 / 77.1, prompt 985 / 1,535 / 1,533. Needles 6/6 in both orders.
- **Mixed-arch split on long prompts:** every 32K/128K prompt went through the second card in both orders, no errors (the
  failure in #690 is on a mixed AMD pair).
- **PCIe:** the V100's x4 link peaks at ~3.6 GB/s (essentially full) during long-prompt reading.
- **Limitations:** one machine, one size, three runs per configuration, no vision in the speed runs.

The report was drafted with an AI assistant (Claude) from our measurements, and checked by me.