RQ2 · Scaling knobs — Strix Halo throughput

Two actionable results: keep sequence length as short as the task allows (the biggest free throughput knob), and stay at batch size 1 — larger per-device batches reduce tok/s because the shared LPDDR5X (~215 GB/s) is the bottleneck, not compute. LoRA rank and packing are nearly free to pick on throughput grounds.

Window note: the secondary sweep uses a shorter measurement window than the primary A/B, so its absolute baseline reads a few % higher (820→~872 tok/s for the same config). Relative effects — what this page is about — are unaffected; see Methods.