#2472 bench: community report, RTX 4070 12 GiB Thunderbolt 4 eGPU + i9-12950HX, Flash-Next IQ3_S, 32K on 0.1.44.1

本文

## What this changes / what it leaves alone
Adds one community result set and one index row. It changes no engine code, no default, and nothing else in the docs.

- New: `bench/results/2026-10-14-community-rtx-4070-qwen3-8-flash-next-iq3-s/` (`README.md`, `results.json`)
- New: one row in the index at `bench/results/COMMUNITY.md`

## Hardware and software
- **GPU: NVIDIA GeForce RTX 4070, 12 GiB VRAM (12282 MiB) in a Thunderbolt 4 eGPU enclosure.** The card is not in a motherboard slot, so the host-to-device link is about PCIe 4.0 x4 rather than a full x16 slot.
- For this run the engine's own calibration kept the PCIe share small: `pcie_frac` 0.08 in `results.json`, so little expert traffic crosses the Thunderbolt link.
- CPU: 12th Gen Intel Core i9-12950HX (16 cores, 24 threads); installed RAM 62.5 GiB; storage not measured.
- OS: Linux 7.2.9-2-cachyos; driver 580.178.04; CUDA 13.0; Python 3.14.8.
- Strata project 0.1.44.1, commit 296b59d2, engine 0.1.44.1 (official build).

## Model and configuration
- `qwen3.8-flash-next-iq3_s` (IQ3_S, GSQ-RCO), two gguf files plus `--ple-gguf` and `expert-profile.bin`.
- 32768 context, 1 parallel slot, the install's own configuration unchanged; thinking off for the suite, temperature 0.

## Method
`python tools/strata_bench.py` suite 1, 3 runs per test, medians with ranges, against an already-running server (`started_server: false`), so model loading is not timed and there is no `engine-log-excerpt.txt`. One warm-up chat and a 1K prompt were discarded first; every fresh prompt starts with a unique tag. Suite total 543 s (elapsed 542.78). `clients: 0`, `long: false`, skipped 0, failures 0. Privacy scrub passed.

## Numbers
- Decode: 33.5 tok/s short chat (30.4-38.7), 35.8 tok/s agent turns (25.7-56.0).
- Prompt, fresh: 213.7 tok/s at 4K, 413.8 tok/s at 16K.
- TTFT: 0.94 s short chat, 18.83 s at 4K, 38.85 s at 16K.
- Agent conversation cache hit over follow-up turns: 68.6 %.

## Limitations
- Speed only: no correctness or needle test was run.
- Prompt sweep only 4K and 16K; no concurrency (`clients: 0`).
- Background load, power limits and storage were not recorded.
- No same-machine PCIe x16 comparison, so the cost of the Thunderbolt 4 link is not quantified here.

## How I checked
- `results.json` and the README table come from the same suite run; medians and ranges in the README were copied from the JSON.
- The hardware section quotes the server, `nvidia-smi` and the OS, and the eGPU link is stated as the submitter reports it.
- Index row in `bench/results/COMMUNITY.md` matches the folder name and the README.