#3314 [Benchmark] 2x RTX 3090 + RTX 5060 Ti (Linux): UD-IQ4_XS on 0.2.0-beta, default and with the sm_86 table set by hand

本文

## Problem and result

A results-only community benchmark: two `tools/strata_bench.py --long` runs on one PC (Linux, Ryzen 9 5950X, 128 GB DDR4, 2x RTX 3090 + RTX 5060 Ti 16 GB) with the same config: UD-IQ4_XS, the layers split over the two RTX 3090 (`--layer-split auto --split-device 1`), the RTX 5060 Ti a helper expert cache (`--expert-cache-device1 auto --remote-expert-opt`), 200K context.

- `2026-10-20-community-2x-rtx-3090-plus-rtx-5060-ti-ud-iq4-xs`: as is. With the 5060 Ti visible the sm_86 table is not applied (#3313).
- `...-ud-iq4-xs-sm86-switches`: the same with the table's nine switches in the config's `env`.

| | default | sm_86 table by hand |
|---|---:|---:|
| short chat decode tok/s | 122.5 | 126.3 |
| agent turns decode tok/s | 118.0 | 123.9 |
| 4K / 16K / 32K prompt tok/s | 855 / 1,830 / 1,917 | 856 / 1,826 / 1,913 |
| agent cache hit, follow-up turns | 69.3% | 69.5% |

## Related discussion

#3313 (the sm_86 table is skipped when a card of another architecture is visible).

## Approach

Each folder has the report and raw requests `strata_bench` wrote, the config with the paths replaced by placeholders, and the engine log of the run (paths replaced the same way). The model name in the files is the quant's (my local config's own name replaced); nothing else was edited.

## Validation

**Passed:** `python tools/strata_bench.py --url http://127.0.0.1:8093 --config <the config> --long`, against `serve/server.py --engine strata --config <the config> --port 8093`, on v0.2.0-beta (3d1fb480, engine 0.1.45.3) built from source for `86;120`, CUDA 13.4, GCC 16.2.1, driver 615.78.08. 3 runs per test, no failed requests, no skipped tests.

**Failed:** `strata_bench.py` starting the server itself: `NameError: name 'port' is not defined` (#3188 / #3235), so the server was started separately and the bench given `--url`.

**Not run:** `--clients` (concurrency), `--think`, other quants, the desktop app.

## Limits

One PC and one run of the suite per configuration (3 runs per test inside it). The model load is not timed.