#1791 bench: RTX 5090 IQ3_S 262K, honest baseline 160-195 tok/s

Beschreibung

## Summary

Results-only community report measured on 2026-10-10: RTX 5090 32GB (Blackwell, sm_120), AMD Ryzen 7 9800X3D 8-Core, 120GB DDR5-6000, Qwen3.8-Flash-Next IQ3_S (GSQ-RCO, 84GB), Strata v0.1.41 self-compiled, 262144 context.

- Decode: 162.5 tok/s (short English prose), 195.8 tok/s (8059-token cold prompt)
- Prefill: 6710.2 tok/s (8059-token cold prompt, cache_n=0)
- Hot-cache artifact: ~230 tok/s on fully cached repeated prompt — NOT representative, documented below
- Concurrency: total_slots=1, sequential queuing (NOT parallel batching)

Related: #1722 (same GPU+CPU, different quant IQ3_XXS and 1M context; no cross-report speedup claim).

## Honest Baseline

**160–195 tok/s** for real-world single-user workloads.

The ~230 tok/s figure (226.0 / 230.6 / 229.6 / 230.3 / 230.7) comes from:
- 8054/8059 tokens already in KV cache (99.9% hit rate)
- Only 5 new tokens to process
- Short output, MTP-friendly repeated content (spec=6)

This is a benchmark artifact. Rule of thumb: if `cache_n / prompt_n > 95%` with short output, the tok/s number is inflated. Cold-cache numbers only for honest comparison.

Short-prompt prefill (~200 tok/s) was retracted as invalid measurement on tiny inputs.

## Configuration

```
--spec 6
--spec-min-p 0.7
--prefill auto:16384
--max-context 262144
--kv int8
--kv-resident 32768
--pcie-frac 0.0
--vision-cpu
```

## Hardware

| Component | Spec |
|---|---|
| GPU | NVIDIA RTX 5090 32GB (Blackwell, sm_120) |
| CPU | AMD Ryzen 7 9800X3D 8-Core Processor |
| RAM | 120GB DDR5-6000 (2x64GB) |
| OS | Ubuntu 26.04.1 |

## Notes for Blackwell Optimization

- sm_120 kernels are the main opportunity; current kernels are not Blackwell-specific
- PCIe Gen5 x16 underutilized (model fits in 32GB VRAM, minimal offload)
- MTP (spec=6) contributes significantly to the 160+ tok/s figure
- int8 KV cache stable at 262K context

## Validation

- All numbers measured via Strata HTTP API on real hardware
- No estimated or extrapolated figures
- Hot-cache numbers explicitly labeled as non-representative

---
*All numbers measured on real hardware, not estimated.*