Skip to content

Latest commit

 

History

History
23 lines (18 loc) · 2.19 KB

File metadata and controls

23 lines (18 loc) · 2.19 KB

Evidence Library

Auto-generated index of xdna-top experiment reports. Each entry is a self-contained, reproducible benchmark grounded in xdna-top evidence artifacts. Add a new docs/experiments/<slug>.md with front-matter and re-run tools/build_experiments_index.py — it appears here automatically.

Experiment Date Platform Headline
Structured-Output Reliability Across Sampling Configs (NPU LLM stack) 2026-06-14 AMD RYZEN AI MAX+ 395 (Strix Halo, gfx1151), kernel 6.17.0-35-generic, XRT 2.21.75 Greedy decoding is the most reliable for structured (JSON) output across small models; aggressive repetition penalties (rep≈1.3) collapse valid output to near zero. Failures are structural, not creative.
NPU vs CPU Background-Job Contention on Strix Halo 2026-06-13 AMD RYZEN AI MAX+ 395 (Strix Halo, gfx1151), kernel 6.17.0-35-generic, XRT 2.21.75 For a latency-tolerant background generation job next to an interactive iGPU model, the NPU does ~3x the throughput at ~3x better total-board perf/watt and contends slightly LESS with the main lane (~3.8% vs ~4.8% decode loss) than spare CPU cores.

Raw artifacts