Search papers, labs, and topics across Lattice.
1
2
3
LLMs struggle to generate efficient NPU kernels, achieving only ~10% average vectorization across a new benchmark of 102 common ML operators, despite promising results on select kernels.
