Aachen 2019 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

T: Fachverband Teilchenphysik

T 42: Experimentelle Methoden I

T 42.8: Talk

Tuesday, March 26, 2019, 17:50–18:05, ST 1

Highly performant, deep neural networks with sub-microsecond latency for trigger FPGAs — •Noel Nottbeck, Volker Büscher, and Christian Schmitt — Johannes Gutenberg-Universität Mainz

Artificial neural networks are becoming a standard tool for data analysis, but their potential remains yet to be widely used for hardware-level trigger applications. Nowadays, high-end FPGAs, as they are also often used in low-level hardware triggers, offer enough performance to allow for the inclusion of networks of considerable size into these systems for the first time. Nevertheless, in the trigger context, it is necessary to highly optimize the implementation of neural networks to make full use of the FPGA capabilities.

We optimized and implemented the processing and control flow of typical NN layers for use within FPGAs, such that they can run efficiently in a real-time context with e.g. the ATLAS data rate of 40 MHz and latency limits of at most few hundred nanoseconds for entire networks. Significant effort was put especially into the 2D convolutional layers, to achieve a fast implementation with minimal resource usage.

A Python-based toolkit has been developed that makes implementing a neural network into an FPGA as easy as executing a few lines of code on an already trained Keras network. Results are presented, both for individual layers as well as entire networks created by the toolkit.