Please visit Jefferson Lab Event Policies and Guidance before planning your next event: https://www.jlab.org/conference_planning.

May 8 – 12, 2023
Norfolk Waterside Marriott
US/Eastern timezone

Acceleration of a CMS DNN based Algorithm

May 11, 2023, 3:15 PM
15m
Hampton Roads VII (Norfolk Waterside Marriott)

Hampton Roads VII

Norfolk Waterside Marriott

235 East Main Street Norfolk, VA 23510
Oral Track 9 - Artificial Intelligence and Machine Learning Track 9 - Artificial Intelligence and Machine Learning

Speaker

Barbone, Marco (Imperial College London)

Description

The search for exotic long-lived particles (LLP) is a key area of the current LHC physics programme and is expected to remain so into the High-Luminosity (HL)-LHC era. As in many areas of the LHC physics programme Machine Learning algorithms play a crucial role in this area, in particular Deep Neural Networks (DNN), which are able to use large numbers of low-level features to achieve enhanced search sensitivity. Sophisticated algorithms however present computing challenges, especially looking forward to the data rates and volumes of the HL-LHC era. Accelerated computing, using heterogeneous hardware such as GPUs and FPGAs alongside CPUs, offers a solution to the computing challenges, both in offline processing and realtime triggering applications. Demonstrating state-of-the-art algorithms on such hardware is a key benchmark for developing this computing model for the future.

The studies presented describe the implementation of a DNN based LLP jet-tagging algorithm, published by the CMS experiment, on an FPGA. Novel optimisations in the design of this DNN are presented, including the adoption of cyclic random access memories for a simplified convolution operation, the reuse of multiply accumulate operations for a flexibly selectable tradeoff between throughput and resource usage, and storing matrices distributed over many RAM memories with elements grouped by index as opposed to traditional storage methods. An evaluation of potential dataflow hardware architectures is also included. It is shown that the proposed optimisations can yield performance enhancements by factors up to an order of magnitude compared to other FPGA implementations. They can also lead to smaller FPGA footprints and accordingly reduce power consumption, allowing for instance duplication of compute units to achieve increases in effective throughput, as well as deployment on a wider range of devices and applications.

Consider for long presentation Yes

Primary authors

Ourida, Tarik (Imperial College London) Prof. Tapper, Alexander (Imperial College (GB)) Prof. Luk, Wayne (Imperial College London)

Presentation materials

Peer reviewing

Paper