npu-bench and Laya on NPU
Tokens-per-second-per-watt benchmarking and a custom single-kernel matrix engine for AMD XDNA2 NPUs on Linux.
Intention
Tokens-per-second-per-watt benchmarking and a custom single-kernel matrix engine for AMD XDNA2 NPUs on Linux.
What exists today
Working code and measured results on one laptop, with a priority-queue benchmarking daemon.
Technical approach
Custom IRON kernels and a measurement harness that records power alongside throughput.
Evidence and known limitations
- Not yet public. Numbers will be published with exact conditions or not at all.
- One machine, one chip generation.