Project Goal

The objective of this project is to optimize the training and inference of fast simulation models as well as to test the capabilities of the new hardware and software tools developed by Intel. This can help CERN better evaluate the capabilities and benefits of the generative deep learning models. The work is a joint effort of CERN openlab, Intel (CA, USA), and SURF (NL).

Background

Detector simulations are an essential part of the HEP research. However, the current simulation techniques based on the Monte Carlo sampling are very time consuming. The number of simulations scales up with each upgrade of the LHC, unlike the available computing resources. The deep learning algorithms can provide detector simulations in a fraction of time and thus help overcome this bottleneck. Yet, the efficiency of the deep learning approach is significantly dependent on the hardware and software being used.

Progress & Achievements

To continue the work started by the end of 2022, the focus in 2023 was mainly on optimizing the inference of the 3DGAN model for calorimeter simulations. We examined the quantization of the 3DGAN generator to INT8 precision while keeping the FP32 accuracy with the Intel® Neural Compressor tool. It enabled an automated partial quantization of the model (3 layers out of 7 converted to INT8) without any increase in the loss value. The inference performance of both version of the model (FP32 and partially quantized) was tested using the Intel® Xeon® Max Series CPUs. Using the Intel Neural Compressor, we obtained a 1.9x speedup on the inference. This results in a speedup of more than 8500x compared to the Geant4 simulations.

Later in 2023, the focus was on examining the possibility to integrate the 3DGAN model into a Geant4based application simulating the entire detector. Conducting this study is relevant as it will provide the community with more detailed evaluation of gains from using the deep learning generators for detector simulation in a more realistic setup where we consider the exchange of information between the Geant4 simulator and the 3DGAN generator.

 

Project Coordinator: Sofia Vallecorsa

Technical Team: Emma Call (Intel), Adel Chaibi (Intel), Valeriu Codreanu (SURF), Soumyadip Ghosh (Intel), Kristina Jaruskova, Duncan Kampert (SURF), Cai Maxwell (SURF), Hans Pabst (Intel), Damian Podarean (SURF), Vikram A. Saletore (Intel), Kalliopi Tsolaki, Sofia Vallecorsa

Intel Collaboration Liaisons: Vikram A. Saletore

In partnership with: Intel