Reverse-engineered NPU engine format to run GGUFs 1.5× faster than vendor runtime

woolcoxm · reddit · 2026-08-28

Author reverse-engineered the engine format of the Axera AX8850 NPU, discovering that int8 weights are stored as two nibble planes. They implemented a method to patch GGUF weights directly into precompiled engines at load time, eliminating the need for model conversion. The project also fixed a supposedly broken batched-prefill path, achieving 1.5× faster decode speed (24.5 t/s) and 716 t/s prompt processing on a Raspberry Pi 5 compared to the vendor's runtime. The analysis reveals the NPU's performance is primarily memory-bandwidth bound.

Original post →

More from Infra

Infra channel →