Reverse-Eng the RK3588 NPU: Hacking Memory Limits to Run Vision Transformers(amohan.dev)
amohan.dev
Reverse-Eng the RK3588 NPU: Hacking Memory Limits to Run Vision Transformers
https://amohan.dev/blog/2025/shard-optimizing-vision-transformers-edge-npu/
2 comments
awesome write-up! i was surprised there wasn’t more built-in/intuitive mechanisms for ignoring the optimization. the solution in the blog post is pretty clever.