#3621 bench: add RTX 5080 16 GB community report (15 models)

Descripción

## Summary

Adds a results-only community report for 15 models on one RTX 5080 16 GB:

`bench/results/2026-10-21-community-rtx-5080-15-models/`

No engine, setup, server, tool or docs changes. The three local `*.json.bak` configs are not included.

**Protocol note:** this is the older `tools/strata_bench.py --long` community format, measured on Strata 0.1.45 (commit `477ac13c`), before the newer `strata bench run` standard protocol. It is a pre-protocol report, not a standard benchmark result. I can rerun on 0.2.x with `strata bench run` separately if that is preferred.

## Hardware and software

| Item | Value |
| --- | --- |
| GPU | NVIDIA GeForce RTX 5080, 16,303 MiB VRAM, one GPU (GPU 0) |
| CPU / RAM | AMD Ryzen 7 9800X3D (8C/16T), 93.7 GiB RAM |
| OS | Windows 11 build 26200 (the tool prints Windows 10 for this build) |
| Driver | 617.14, CUDA 13.4 runtime |
| Strata | 0.1.45, commit `477ac13c`, official Windows build |
| Storage | local NVMe; filesystem cache not flushed between runs |

PCIe width/speed, memory use, power and image inputs were not measured.

## Models and configuration

Twelve Qwen3.8-Flash-Next files:

- Qwen: Q2_0, IQ2_XS, IQ3_XXS, IQ3_S
- Coder: IQ1_M
- Swift 1.5: IQ2_XS, IQ3_XXS, IQ3_S
- Unsloth: UD-IQ4_XS, UD-Q4_K_XL
- NVFP4: Maximilian228 GPTQ pack, NVIDIA pack

All twelve use the same launch shape: `--max-context 262144 --kv int8 --kv-resident 32768 --vision --prefill auto --spec 4 --mtp rt`, the shared `expert-profile.bin`, and `setup.py --calibrate` settings. `Q2_0` keeps the engine default `--pcie-frac 0`; the other eleven use 0.35. Exact per-model commands and calibration values are in `model-manifest.json`.

Three Qwen3.6-35B-A3B files (UD-Q4_K_XL, UD-Q6_K_XL, Q8_0) use `--max-context 65536`, default fp32 KV, `--mtp 2` and `--vision`.

## Method

- Official `tools/strata_bench.py --long`, 3 runs per test, 540 per-request rows total.
- Chat: 3 questions x 3 runs, 100 new tokens. 4K/16K/32K: 9 fresh prompts each, 24 new tokens. Agent: 3 sessions x 6 turns, 64 new tokens per turn; 15 follow-up turns are measured.
- Actual prompts were 4,019 / 16,058 / 32,049 tokens for 4K/16K/32K and 43-51 tokens for chat, not nominal lengths.
- Recall: 32K/128K/200K at 10%/50%/90% depth for the twelve 256K files. Qwen3.6 ran 32K/64K plus a 60K fallback.
- Quality checks (not an official suite): GSM8K-100, MMLU-100 with seed 1234, HumanEval-50.
- Sampling: temperature 0; thinking off for the suite, on for GSM8K/HumanEval and off for MMLU.

## Key results

- Fastest chat decode: Q2_0 at 122.6 tok/s; IQ2_XS 118.3; slowest Qwen3.6 Q8_0 at 32.8 tok/s.
- Prompt read: about 2.5-2.85k tok/s at 128K and 2.28-2.33k tok/s at 200K.
- Every valid needle target was 3/3. All twelve Flash-Next files passed 32K/128K/200K.
- Qwen3.6 at 64K: all three files rendered 65,584-65,586-token prompts plus a 40-token cap against a 65,536-token context and got HTTP 400. The 60K fallback passed 3/3 for all three.
- Quality ranges: GSM8K 96-99%, MMLU 73-89%, HumanEval 94-100%. Coder IQ1_M has the lowest MMLU (73%).

## Limitations

- One machine, one model session per configuration. Expert caches warm within a session; results are not a cold-start comparison.
- The 128K rate is the median of two fresh requests; the 90%-depth request reused 16,384 prefix tokens and is excluded. Every 200K request reused 16,384 tokens, so that rate covers 183,936 freshly read tokens.
- Quality sets are small (100/100/50); MMLU has roughly +/-4 points of standard error.
- NVFP4 is a prepared pack that `--inspect` cannot read, so its weight/bits cells are marked unavailable.
- No memory, power, PCIe-link, concurrency or image measurements.

## Files

- `README.md`: report, tables, method, limitations
- `measurements.csv`: 540 official-suite and agent requests
- `long-context.csv`: 135 needle requests, including the rejected 64K rows
- `quality.csv`: model-level scores
- `model-manifest.json`: repositories, revisions, GGUFs, launch commands, calibration
- `runs/<model>/`: original `results.json`, quality JSON and engine-log excerpts
- `scripts/summarize.py`: rebuilds the speed and recall tables from `runs/`

If 15 models is too much for one review, I am happy to split this into a Flash-Next PR and a Qwen3.6 PR.