#2791 bench/results: RTX 5070 Ti (16 GiB) with 31.6 GiB of RAM — IQ2_XS at 128K in low-RAM mode

Descrição

Adds a results-only report for an **RTX 5070 Ti (16 GiB) with 31.6 GiB of system RAM** — the same GPU as `2026-10-04-community-rtx-5070ti-windows` and `2026-10-06-community-rtx5070ti-5900x`, but at half their RAM, which in `bench/results/` is a combination that is not covered yet.

Only three files under `bench/results/2026-10-17-community-rtx-5070-ti-qwen3-8-flash-next-iq2-xs/` (`README.md`, `results.json`, `engine-log-excerpt.txt`). No engine or code changes, as the guide asks for a results-only submission.

## Hardware

- RTX 5070 Ti, 16,303 MiB VRAM, power limit 300 W (not reduced), PCIe Gen5 x16
- Ryzen 7 9700X (8C/16T, AVX-512), **31.6 GiB RAM**
- WD_BLACK SN770 1 TB NVMe (model files and OS on the same drive)
- Windows 11 Pro 10.0.26200, driver 616.64, CUDA 13.0
- Not an isolated machine (proxy, Steam, ComfyUI MCP bridge, antivirus; the GPU also drives the display)

## Model and configuration

- `ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF` @ `ed59f92`, IQ2_XS, both shards with their SHA-256
- Context **131,072**, KV `q4_0`, KV streaming off, **low-RAM mode (`--mmap-experts`)**, expert cache `auto` (~5,000 slots / ~6.8 GiB VRAM), `--prefill auto`, MTP `--spec 4 --spec-min-p 0.5`, `--vram-reserve-mib 1600`
- Strata 0.1.44.1 / engine 0.1.44.1, release binary, zip install (not a git checkout)
- Thinking off for the suite; the suite's requests set `temperature 0`

## Results (medians, ranges over 3 runs)

| Configuration | Actual prompt tokens | Runs | Prompt tok/s | Decode tok/s | TTFT s |
| --- | ---: | ---: | --- | --- | --- |
| short chat (100 tok cap) | 46 | 9 | not measured | 57.2 (36.8-87.0) | 0.54 (0.42-1.06) |
| 4K prompt, fresh | 4019 | 3 | 270.0 (226.5-270.2) | not measured | 14.94 (14.94-17.80) |
| 16K prompt, fresh | 16058 | 3 | 447.8 (446.7-518.1) | not measured | 35.94 (31.07-36.03) |
| agent session, follow-up turns | 14572 | 3 | see report | 45.2 (32.3-51.4) | 12.79 (10.82-17.17) |

Total suite time 420 s.

## What the report is about — and its limitation

This install cannot keep the 36 GB expert arena in 31.6 GiB of RAM, so unlike the two 64 GiB 5070 Ti reports it runs the low-RAM mode and serves the experts through the OS file cache. The measured boundary:

- **decode** lands at ~85-87 tok/s once the cache is warm, roughly half the 64 GiB figures (102 tok/s)
- **prompt processing** is 270-448 tok/s against 1,992-3,523 tok/s at 64 GiB — **4-8x slower**, because the prompt path re-reads lent experts from the SSD

So on a 32 GiB machine the prompt/TTFT side, not decode, is what decides whether the setup is usable (a 16K prompt is 36 s before the first token). The chat row's 36.8-87.0 range is the low-RAM cache warming across the nine back-to-back requests, not run-to-run noise; the report says so in the Results section.

Also recorded because the guide asks for them: storage, PCIe link, power limit and the background load are filled in, and two things the tool got wrong are corrected — it reports the OS as "Windows 10 (10.0.26100)" (actually Windows 11 Pro 10.0.26200) and its own README template asks for facts it cannot read (CUDA version, commit on a zip install).

## One tool bug found while preparing this

`tools/strata_bench.py:564` reads `CMakeLists.txt` with `read_text()` and no encoding, so on a Chinese/Japanese Windows (GBK code page) it raises `UnicodeDecodeError: 'gbk' codec can't decode byte 0xbf in position 2` on the file's UTF-8 BOM and exits **before any measurement**. `setup.py:4649` reads the same file with `encoding="utf-8"`. `PYTHONUTF8=1` works around it. Happy to file it as an issue with the traceback if that is more useful than fixing it in this PR — I left it out of this PR to keep it results-only.