A transformer on FPGA for low latency applications in particle physics and beyond

About

To identify physics phenomena that are difficult to detect by standard triggers, real-time modern AI triggers are needed. One of the most promising novel trigger classes is anomaly detection triggers, capable of flagging rare events without relying on a full theoretical understanding of the underlying process. By leveraging modern, heterogeneous accelerator platforms, like the AMD Versal Adaptive Compute Acceleration Platform, which combines classic Field Programmable Gate Array (FPGA) logic with higher-level AI Engines, we aim to achieve ultra-low-latency inference at a high throughput, necessary for operations in the CMS trigger and scouting system. We present a machine learning architecture based on a Transformer model, specifically adapted for low-latency inference. As the Transformer architecture has shown great success in interpreting complex relationships across many areas of research, we believe that a fast, flexible, high-throughput Transformer model, implemented 'on the edge', has wider-reaching applications across technology and industry.

0:00:00 Slide 1

0:00:03 Slide 2

0:00:19 Slide 3

0:00:48 Slide 4

0:01:11 Slide 5

0:01:28 Slide 6

0:02:46 Slide 7

0:03:10 Slide 8

0:03:30 Slide 9

0:03:51 Slide 10

0:04:18 Slide 11