#3184 [Benchmark] RX 7700 XT / 5700X3D, IQ2_XS, Windows HIP 0.1.45

说明

Sharing measurements from an RX 7700 XT system, following docs/COMMUNITY_BENCHMARKS.md.

## Hardware and software

- GPU: AMD Radeon RX 7700 XT, 12 GiB VRAM, gfx1101
- CPU: Ryzen 7 5700X3D, 8 cores / 16 threads, AVX2
- RAM: 64 GB installed, 63.9 GiB reported
- OS: Windows 11, build 10.0.26300
- GPU driver: 32.0.31041.1004
- Engine: official prebuilt Windows HIP 0.1.45
- Repository commit: 477ac13c
- Model: Qwen3.8-Flash-Next-GSQ-RCO-IQ2_XS

A separate RAM-using background server and the desktop client remained open. Background load, clocks and power were not controlled.

## Configuration

- Configured context: 262144 tokens
- KV: int8, streaming with 32768 resident cells
- One serving slot
- Prefill: auto:8192
- MTP enabled; speculative decoding: 4, spec-min-p: 0.5
- Conversation cache: 8192 MiB, 4 slots, minimum 8192 tokens, 2560 MiB free-memory floor
- prompt-cache-spread enabled

Opt-in environment settings:

STRATA_PF_FUSED=1
STRATA_PF_GEMM=1
STRATA_PF_SWITCH_MIN_T=4096
STRATA_PF_PAD=1
STRATA_HC_UPMIX=1
STRATA_PREFILL_CPU_SHARE=auto
STRATA_PREFILL_CPU_SHARE_MAX=3072

## Method

Measured on 2026-10-19 using tools/strata_bench.py against an already-running local server.

Thinking was off and temperature was 0. Three runs per test; short chat used three questions per run, giving nine requests. Fresh prompts had unique tags and reported zero reused tokens.

This is a speed measurement, not a correctness or recall benchmark. Model loading was not timed.

## Results

| Test | Actual prompt tokens | Reused tokens | Median speed | Median time to first token |
| --- | ---: | ---: | --- | --- |
| Short chat, 9 requests | 43–51 | 0 | Decode 42.0 tok/s | 0.74 s |
| Fresh 4K | 4019 | 0 | Prefill 412.7 tok/s | 9.78 s |
| Fresh 16K | 16058 | 0 | Prefill 760.4 tok/s | 21.16 s |
| Agent follow-up turns | About 14.5K at final turn | 68.8% aggregate reuse | Decode 37.1 tok/s | 8.90 s |
| Fresh 32K supplement | 32050 | 0 | Prefill 665.7 tok/s | 48.207 s |

The main suite took approximately 307 seconds.

### Separate 32K supplement

Only the official Suite.calibrate() and Suite.run_prompt(32768, "32K") paths were added: two tiny calibration requests followed by three fresh prompts, capped at 24 output tokens. The other tests were not rerun.

| Run | Prompt tokens | Reused | Prefill tok/s | TTFT |
| --- | ---: | ---: | ---: | ---: |
| 1 | 32050 | 0 | 689.2 | 46.564 s |
| 2 | 32050 | 0 | 665.7 | 48.207 s |
| 3 | 32050 | 0 | 665.1 | 48.245 s |

## Runtime conditions and limitations

- Both engine starts accepted 2 MiB large pages.
- The main suite used 2650 PROFILE expert slots, 3.57 GiB VRAM, and automatic prefill chunks of 5376 tokens.
- The 32K supplement followed a Strata-only restart enabling effort_position=end. The new process selected 2560 expert slots, 3.45 GiB VRAM, and 5120-token chunks.
- These are separately labelled measurements, not a controlled before/after comparison.
- Windows commit limit was 124.6 GiB. Before the 32K supplement, available commit headroom was 46.7 GiB.
- An earlier log warning showing only 6.5 GiB commit headroom belonged to an older engine start. The page file had already been enlarged online before these measurements.
- No Windows reboot or controlled large-pages-off/on comparison was performed.
- Low-VRAM warnings occurred. GPU paging was not instrumented.
- No correctness, recall, 128K/200K sweep, or different-build A/B comparison was performed.

The main suite needed PYTHONUTF8=1 because project_version() read the UTF-8 BOM in CMakeLists.txt using the Windows default encoding. The local fix matches existing PR #3092; no duplicate fix PR is proposed.

[strata-rx-7700xt-community.zip](https://github.com/user-attachments/files/33673727/strata-rx-7700xt-community.zip)