#2642 bench: community report - GTX 1080 Ti (sm_61 Pascal), 2x Xeon E5-2696 v4, IQ3_S at 262K

Description

## Summary

A community benchmark on a **GTX 1080 Ti (sm_61 Pascal)** - [TEST_REQUESTS.md](https://github.com/Niko1221/Strata/blob/main/docs/TEST_REQUESTS.md) asks for Pascal and the earlier answers it lists are Tesla P40s. A consumer card shares the fp32 compute path but has a different memory subsystem (GDDR5X, higher clocks), so it is a useful second data point for the row.

**Strata 0.1.44.1 official release engine** (the installed server - a release binary, not a source build), the standard `tools/strata_bench.py --long --runs 3` suite against the running server, so model loading is not timed.

## Results

| Configuration | Prompt tok/s median (range) | Decode tok/s median (range) | TTFT s median (range) |
|---|---|---|---|
| short chat | — | 27.4 (23.1-31.9) | 1.28 (1.08-1.32) |
| 4K prompt, fresh | **202.4 (201.1-202.8)** | — | 19.92 |
| 16K prompt, fresh | **194.6 (193.5-196.7)** | — | 82.59 |
| agent follow-up turns | ~115 (resent history) | 23.6 (19.1-24.5) | 25.60 |

Prompt throughput is flat from 4K to 16K and reproducible to better than 1% across the three runs of each.

**Conversation cache on agent turns:** 66.0% (63.2-68.8) of the resent history reused, **3 tokens a session read again** - the cache holds almost all of it. That part of the prompt runs at ~115 tok/s against ~202 tok/s fresh.

## What failed, honestly

The **32K prompt test has no numbers**: one agent session and all three 32K requests failed with the server down (the guard script restarted it). The engine log shows healthy requests up to the last successful one and no crash line. The report gives the likely cause without claiming it is proven: the agent sessions park their conversation snapshots (the log's last parking line: `bytes=3911781620` - 3.9 GiB against the configured 8 GiB budget) and the 32K fresh prompts need large buffers on top of that on an 11 GiB card. Worth a look for anyone with a small-card setup.

## Hardware and settings

GTX 1080 Ti 11264 MiB, 2x Xeon E5-2696 v4 (44c/88t, **no AVX-512** - the engine runs its expert kernels on AVX-2), 125.8 GiB RAM, NVMe. **PCIe Gen3 x8** although the card negotiates x16 elsewhere - nvidia-smi reports `pcie.link.width.current 8`, matching the link-tier observations in #2099. Stock 250 W power limit. Qwen3.8-Flash-Next GSQ-RCO **IQ3_S** native pack at **262,144** context, `--kv int8 --kv-resident 32768`, `--expert-cache auto --prefill auto --spec 4 --spec-min-p 0.5`, MTP on.

## Not tested

Contexts past 32K (the 32K test failed, above), `--batch` concurrency, the layer-split path, KV formats other than `int8`, and any correctness check - this is a speed suite. No thermal or clock capture.

Results only, no engine change. `README.md` and `results.json` (per-request data, sanitized by the tool and re-checked for host names and paths before pushing).

🤖 Generated with [Claude Code](https://claude.com/claude-code)