#3475 [Benchmark] RX 9070 XT / 64 GB: IQ3 and Q4 performance and coding results

本文

Benchmarks from an RX 9070 XT 16 GB, Ryzen 7 9800X3D and 64 GB RAM running Strata's native Windows HIP engine 0.1.45.3.

| Measurement | IQ3_XXS | UD-Q4_K_XL |
| --- | --- | --- |
| Short-prompt generation | 59–71 tokens/s in warm repeats with expert skipping off | About 8–9 tokens/s |
| Fresh long-context prompt | About 136 s to first token at 130K input | 476–499 s at 262K input |
| Express coding task | 234–255 s; visible tests passed, independent concurrency check failed | 1502 s; visible tests and independent check passed |

IQ3 also attempted a WordPress plugin task and full-stack debugging. WordPress passed the visible tests but mishandled database read failures; the full-stack repair passed all 16 tests and five additional checks. A separate Q4 coding loop with 225K–230K input completed in 836 seconds, passing 15 tests and five independent checks.

Includes 39 individual inference records, configurations, prompt generators, coding fixtures, generated patches, and test evidence. Measurements were collected October 19 local time / October 20 UTC.

The long-context prompts use synthetic code, and the coding projects are small fixtures. Background applications and cache allocations varied between groups. The context sizes and agent setups differ, so these results show observed behavior rather than a controlled quantization-quality comparison. The report documents those settings alongside each measurement.