#3383 [Benchmark] RX 7900 XTX 24 GB (gfx1100), Windows: 0.2.0-beta hip-rdna3 package vs 0.1.44.1, and the GPU vision encoder (Qwen3.8-Flash-Next IQ3_XXS)

本文

### What did you measure, or what went wrong

The `windows-x86_64-hip-rdna3` package of 0.2.0-beta on an RX 7900 XTX, against 0.1.44.1 on the same machine, model and config. Measured on 2026-10-20. Main limitation: two runs per variant, one card, and a large download was writing to the model drive during the runs (about 60 MB/s), so all numbers are ~5-10% below what this box does on a quiet night, on both sides.

**Hardware and software**

- GPU: AMD Radeon RX 7900 XTX 24 GB (gfx1100); CPU: Ryzen 9 7900X3D; RAM: 64 GB; model files on a PCIe 4 NVMe.
- OS: Windows 11 Pro 26100.
- A: Strata 0.1.44.1 (`strata-windows-x64-hip.zip` from the v0.1.44.1 release).
- B: Strata 0.2.0-beta, engine 0.1.45.3 (`strata-engine-0.1.45.3-windows-x86_64-hip-rdna3.tar.gz`, SHA-256 checked; BUILD.json: commit 18f7b90ca, rocm 10.2.0a20260930, hipblaslt_version 100500). The log shows `prefill gemm: hipBLASLt tuning enabled (32 rows, gfx1100, version 100500)`, so the tuned tables are in the package and used.
- Model: Qwen3.8-Flash-Next GSQ-RCO IQ3_XXS, native pack `iq3_xxs`, expert profile `data/expert-profile.bin` (24576 ranked pairs).

**Config (same for A and B)**

```
strata.exe --serve --pack <pack> --native <gguf 1 of 2> --ple-gguf <gguf 2 of 2> --expert-profile <expert-profile.bin>
  --expert-cache auto --prefill auto --spec 4 --spec-min-p 0.5 --mtp <mtp/rt> --max-context 131072 --kv int8
  --kv-resident 131072 --short-read 512 --vision --vram-reserve-mib 1536
  --conversation-cache-mib 4096 --conversation-cache-slots 2
  (+ the experimental speed projection control vector, layers 4-44)
```

**Method**

One fresh engine per session, A and B alternated (A, B, A, B). Each session: a 20-token warm-up, then fresh prompts (unique text, 0 reused), additions to the same conversation, a 300-token story, and in a second session a fresh 26K prompt and two pictures (an invoice, a chart). Times are client-side, speeds are the engine's `prompt ... tok/s` / `generated ... tok/s`.

**Results (run 1 / run 2)**

| | A 0.1.44.1 | B 0.1.45.3 rdna3 |
|---|---|---|
| fresh 9K prompt | 10.1 / 11.2 s (919 / 831 tok/s) | 11.0 / 10.4 s (852 / 900 tok/s) |
| +2.8K added to that conversation | 4.7 / 4.7 s | 4.6 / 4.7 s |
| fresh 11.7K prompt | 10.2 / 10.2 s (1187 / 1189 tok/s) | 11.4 / 10.1 s (1060 / 1183 tok/s) |
| fresh 26K prompt | (24.9) / 25.9 s (1072 / 1024 tok/s) | 25.2 / 25.9 s (1053 / 1022 tok/s) |
| 300-token story, decode | 61.1 / 58.2 tok/s | 57.8 / 59.4 tok/s |
| invoice picture, question | (15.0) / 16.8 s | 16.8 / 17.1 s |
| chart picture, question | (14.5) / 16.2 s | 15.2 / 15.6 s |
| expert cache | 7829 / 7831 slots | 7831 / 7831 slots |

No measurable difference between A and B on this card: the spread between two runs of the same build is as large as the gap between builds.

**GPU vision encoder of 0.2.0-beta** (`strata-vision-hip.exe`, log: `warmed up at 289 image tokens`) against the CPU encoder (12 threads), engine B, same config, alternated:

| | CPU encoder | GPU encoder |
|---|---|---|
| invoice | 16.2 / (15.5) s | 13.5 / 12.6 s |
| chart | 15.7 / (14.3) s | 12.1 / 12.4 s |
| photo | 11.5 / (6.5) s | 11.7 / 11.8 s |
| decode, 4 stories | 61.2 / 59.4 tok/s | 58.8 / 59.4 tok/s |
| expert cache | 7912 / 7788 slots | 7094 / 7094 slots |

About 3 s faster on documents, nothing on the photo, and ~750 fewer expert-cache slots, with no visible effect on decode here. Answers to the pictures were right with the GPU encoder.

**Poor results and failed requests**

- The CPU run 2 (numbers in brackets) came up in the bad start state of #2960: all three picture answers were about some other picture, so its times don't count. In the speed bench, run 1 of A (its long session, numbers in brackets) also answered the 26K prompt wrong, same defect. Details and logs in #2960.

### Reproduction status

Measured locally