#2479 Community report: RTX 3080 Laptop 8 GB, i7-11800H, 32 GB RAM, Linux (0.1.44.1)

Description

## Title
Community report: RTX 3080 Laptop 8 GB, i7-11800H, 32 GB RAM, Linux (Strata 0.1.44.1)

Issue: answers the "Linux host with 32 GB of RAM or less" test request (#1194). Not a fix, so nothing is resolved.

## Summary
Three `bench.sh --long` runs on an 8 GB laptop card with 32 GB of RAM, Strata 0.1.44.1 source build (`296b59d2`), 32,768-token context, int8 KV, setup's configs, performance power profile, one engine start per configuration, each from GPU ≤ 50 °C:

| | Coder IQ1_M, setup's config | same + `STRATA_PREFILL_CPU_SHARE=auto` | IQ2_XS, mapped |
|---|---:|---:|---:|
| chat decode tok/s, median (range), 9 chats | 24.0 (22.7-29.0) | 25.3 (22.3-35.0) | 32.6 (22.4-39.8) |
| 4K / 16K prompt tok/s, median of 3 | 294 / 283 | 541 / 628 | 226 / 196 |
| 16K prompt, first token | 56.8 s | 25.6 s | 82.2 s |
| failed requests | 0 | 0 | 0 |

## What changed
- `bench/results/2026-10-14-community-rtx-3080-laptop-gpu-qwen3-8-flash-next-coder-iq1-m/`: Coder IQ1_M, setup's config unchanged.
- `bench/results/2026-10-14-community-rtx-3080-laptop-gpu-qwen3-8-flash-next-coder-iq1-m-cpu-share/`: the same with `STRATA_PREFILL_CPU_SHARE=auto` in the server's environment (bench.sh does not record it; the README says so). `-cpu-share` suffix because bench.sh gave both the same name.
- `bench/results/2026-10-14-community-rtx-3080-laptop-gpu-qwen3-8-flash-next-iq2-xs/`: IQ2_XS with setup's mapped config.
- `docs/COMMUNITY_BENCHMARKS.md`: one line under "Community reports".

Each folder is bench.sh's output unchanged, plus a section at the end of its README headed "Added by the submitter, not measured by bench.sh" (machine, build, file hashes, power profile, heat, memory at start, what was not tested).

## Extra Notes
- Two observations from these runs are filed separately: https://github.com/Niko1221/Strata/issues/2477 (VRAM free after load is below the engine's own 512 MiB warning) and https://github.com/Niko1221/Strata/issues/2478 (the CPU prefill share is not armed on 32 GB; forcing it gives 1.8-2.2× prompt speed).
- One start per configuration: chat-decode differences under ~20% between the folders are within one start's spread. Prompt throughput is tight (±6%).
- Results only; no engine changes.

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

https://claude.ai/code/session_01G11swoAvpB3YhkFcCLfyAK