Reddit
npunlock runs custom C kernels on Intel Core Ultra NPUs by rebuilding the path Intel's public stack leaves out
Intel's NPUs contain programmable SHAVE cores, but the public stack exposes only graph-level programming. npunlock reconstructs the path from C source to a runnable kernel and ships a Python API where npu.custom() places a C kernel in a graph. Its README example is an FP32 GELU checked against NumPy. It is verified on Windows x64 with Meteor Lake NPU3720, and a 2026-09-23 update runs mixed-precision custom branches in one graph. The Show HN has 45 points. This is reverse engineering on one NPU generation, but it opens NPU compute beyond the ops Intel ships.
↳ Follow the thread