Decode · tok/s
158.3
Prefill
3149
tok/100W
22.6
2× RTX 3090s. Just 280W power cap per GPU. Strata v0.1.43 running Qwen3.8-Flash-Next IQ3_S: Peak decode: 158.3 tok/s Sustained decode: 122+ tok/s Peak prefill: 3,149 tok/s 48GB VRAM + 64GB RAM | 262K context | INT8 KV | CPU/RAM offload | pipeline-windows=2
First-person 2× 3090 update on v0.1.43. Rank is the quoted peak 158.3. Sustained 122+ stays in the note. Prior pair on the same cards was 153.4 peak on v0.1.42.
