Accelerating Transformer Neural Networks on FPGAs for High Energy Physics Experiments

About

The ever-increasing data rates and ultra-low-latency requirements of particle physics experiments demand innovations for real-time decision-making. Transformer Neural Networks (TNNs) have demonstrated state-of-the-art performance in classification tasks, including jet tagging, but implementations on CPUs and GPUs fail to meet the constraints for real-time triggers. This work introduces two novel TNN architectures optimized for Field-Programmable Gate Arrays. The first one prioritizes latency, achieving a speedup of >1000× over GPU implementations while maintaining accuracy. The second one explores trade-offs between latency and accuracy through design-space exploration and a custom post-training quantization, which identifies optimal bit-widths, yielding significant reductions in hardware resource utilization with negligible accuracy degradation. Experiments demonstrate the effectiveness of these designs, setting new benchmarks for real-time machine learning in high-energy physics..

0:00:00 Slide 1

0:01:41 Slide 2

0:05:39 Slide 3

0:07:39 Slide 4

0:09:23 Slide 5

0:10:31 Slide 6

0:12:05 Slide 7

0:13:07 Slide 8

0:16:22 Slide 9

0:18:06 Slide 10

0:19:38 Slide 11