#3495 [Benchmark] 2x NVIDIA TITAN RTX / Qwen3.8-Flash-Next / IQ3_S / strata-qwen38-thinking-v1-draft.1

Description

Standard-benchmark result, results-only pull request per `docs/BENCHMARK_PROTOCOL.md` (Submit and review).

**Report:** [`bench/results/2026-10-20-standard-titan-rtx-015453d3/README.md`](bench/results/2026-10-20-standard-titan-rtx-015453d3/README.md)

Protocol `strata-qwen38-thinking-v1-draft.1` (draft, labelled as such). Status: **complete**. 21 of 21 required requests, 0 failed, 0 skipped. 12 of 12 sanity checks passed (instruction 4/4, retrieval 4/4, tool calls 4/4). The optional concurrency stage was **not** run.

| Case | Prompt tok/s | Decode tok/s | TTFT s |
|---|---|---|---|
| short | 106.1 (104.5-106.9) n=3 | 86.3 (82.2-86.8) n=3 | 0.75 (0.74-0.76) n=3 |
| prompt-4k | 955.3 (952.8-962.5) n=3 | 76.4 (76.1-84.2) n=3 | 4.31 (4.28-4.32) n=3 |
| prompt-16k | 1474.6 (1473.1-1480.5) n=3 | 79.5 (68.8-81.1) n=3 | 11.15 (11.11-11.16) n=3 |

**Hardware.** 2x NVIDIA TITAN RTX 24 GB (Turing, sm_75) on an HP Z440, Xeon E5-2696 v4 without AVX-512, 125 GiB RAM, NVMe. `nvidia-smi topo -m` reports `NV2` between the cards, but the engine does not use NVLink or peer-to-peer access, so the second card adds expert cache and pipeline depth rather than bandwidth.

**Software.** Strata v0.2.1, engine 0.1.46.1, commit `dea6920b818424c8972ae2c0daf5676bbddde214`, built from source with `CMAKE_CUDA_ARCHITECTURES=75` (the release ships no Linux engine for Turing). AlmaLinux 9.8, driver 615.71.09.

**Model.** Original Flash-Next IQ3_S, `ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF` at revision `ed59f92082b1e93c0e96d60a8b11aab089b52f09`, source revision `de4b8e4d43b917e7706784d8bb445c9af86a3540`, at the model's native 262,144-token context.

**Track: `custom`**, as the report states. The engine is a source build for sm_75, the vision encoder was built locally for sm_75 as well, and `--kv-resident 65536`, `--ple-io ram`, `--spec 4` and `STRATA_STAGE_TRIM=1` are deliberate choices. The full run config is in the report folder with the API key redacted.

**Evidence.** The folder holds the runner's own `README.md` and `results.json`, `outputs.jsonl` (every request with its raw usage and timings), `engine.log` (the start-up and the whole run), `server-config.json` (the run config, key redacted), `config.json` (the runner's view of the server), `provenance/` (the chat template and the corpus manifest) and `artifact-hashes.json` over all of them.

**Limitations**, as the report lists them: the expert cache stays warm across the run, so these are not cold-model measurements; the prompt-prefix reuse is the shared template prefix only; the runner cannot see what else ran on the machine; and the run used a server that was already listening on loopback, so it could not confirm the engine had not answered anything before the warm-up (it had not - the engine was started for this run alone).