#3408 [Benchmark] community report: RTX 4060 Laptop 8 GiB (Windows 11, 32 G…

说明

## What this adds

`bench/results/2026-10-20-community-rtx-4060-laptop-gpu-qwen3-8-flash-next-q2-0/`: the `README.md` the suite wrote, `results.json` with every request and the machine facts, and the engine start-up log. Nothing else in the tree is touched.

## Why it may be useful

- **This is the smallest VRAM tier in `bench/results`.** The smallest existing entry is a 12 GiB RTX 3060 (`2026-10-09-community-rtx-3060-0141`); 8 GiB is the tier where the expert cache holds about 6% of the (layer, expert) pairs - the GPU holds the dense weights, the draft head and the streamed KV window too, leaving 1,487 cache slots.
- **The host has 32 GB of RAM**, the tier `docs/TEST_REQUESTS.md` asks for ("a host with 32 GB of RAM or less", because the expert file tier and the resident mode depend on how much of the model fits in RAM). This one has 31.7 GiB visible.
- The CPU has **no AVX-512**, so the expert kernels run on AVX-2 - the same constraint the closest existing entry had.

## Hardware and software

- GPU: RTX 4060 Laptop 8 GiB (sm_89), PCIe Gen4 x8, 12.8 GB/s measured host->device; 80 W limit
- CPU: i7-14700HX (8 P-cores + 12 E-cores, 28 threads), no AVX-512
- RAM: 32 GB (31.7 GiB visible); Windows 11 10.0.26200; driver 580.97; CUDA 13.0
- Strata: engine 0.1.45.3 (0.2.0-beta), source build from upstream main `3fa1b524` plus three local patches: three SSE keep-alive events in `serve/server.py`, `-DCMAKE_CUDA_RUNTIME_LIBRARY=Shared` in `setup.py`, and the MSVC `tq_avx2.cpp` unroll fix (Gyro only; Q2_0 does not use it)

## Suite medians

`tools/strata_bench.py --config strata-q2_0.json --runs 3 --long`, thinking off, the install's own configuration (Q2_0 pack, `--kv q4_0 --kv-resident 20480`, `--resident-experts`, `--pool-workers 19`, 131,072 context):

| Test | Prompt tok/s | Decode tok/s | TTFT s |
| --- | ---: | ---: | ---: |
| short chat (3 questions, 100 tokens) | - | **37.4 (29.9-49.7)** | 0.97 (0.64-1.08) |
| 4K fresh prompt | **266.5 (244.0-267.7)** | - | 15.17 (15.09-16.54) |
| 16K fresh prompt | **282.2 (279.0-283.4)** | - | 56.99 (56.78-57.64) |
| 32K fresh prompt | **289.4 (288.6-296.9)** | - | 110.93 (108.05-111.19) |
| agent session, follow-up turns | 228.7 (last turn) | **27.5 (9.8-39.2)** | 12.40 (10.31-14.47) |

The agent session hit the conversation cache 69.2% of the time; the last turn carries 14,536 prompt tokens.

## The one axis the suite does not cover: free RAM at start

The expert cache is sized from the RAM available when the engine starts, and on a 32 GB host that number moves with what else is loaded. Across starts on this box the complement ranged 13.15-18.62 GiB of the 29.73 GiB of experts the VRAM cache does not hold (44%-62.7%), and Q2_0's decode stayed in the same 24-28 band - so for this quant the coverage in that range was not the limit; closing other programs before the start still buys the largest complement, and this run had it (62.7%).

## Closest sibling

`2026-10-06-community-rx-6800m`: 12 GiB RX 6800M, 31.4 GB RAM, no AVX-512, Q2_0, `--resident-experts` with `--kv-resident`, a self-built HIP engine. It measured **9-27 tok/s decode and 50-114 tok/s prompt** (its own README names the same two causes: the experts do not fit in RAM and are streamed, and the CPU has no AVX-512). Here: **24-28 decode on longer generations (37.4 on the short chats) and 266-289 prompt**. The differences - CUDA on sm_89 against HIP on gfx1031, 12.8 against 7.0 GB/s PCIe, i7-14700HX against 5900HX, KV 20,480 of 131,072 cells in VRAM against 32,768 of 65,536 - are consistent with the gap.

## Not tested

- correctness or quality (the suite is a speed check, not a quality test)
- vision (enabled, not exercised)
- concurrency (`--clients`), a long soak, and a second GPU
- power limits and background load (not recorded by the suite)

## How to rerun

```text
python tools/strata_bench.py --config strata-q2_0.json --runs 3 --long
```

The folder holds `README.md`, `results.json` and `engine-log-excerpt.txt`. `--config` is required here because the suite's default picks the newest `strata-*.json` of the folder, and this install also has configs for other quants.