AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-21
LEAF: LUT-aware Efficient Attention on FPGAs
Deploying transformers on FPGAs that are accurate, compact, and fast enough for sub-microsecond real-time systems such as those at CERN is difficult, because the attention mechanism that makes these models expressive is costly to implement in hardware. LEAF converts the query, key, and value projections and the output feed-forward layer into FPGA-native layers through LUT-aware training, tabulating the attention block's constant matrix-vector multiplications. On a standard jet tagging benchmark it improves accuracy by 1.8 percentage points while using 2.9 times fewer LUTs than the previous sub-microsecond state of the art, and its most efficient linformer reaches 80.4% classification accuracy at 28k LUTs and 35ns with 64 input particles.