FNAL: Ultrafast Neural Networks for On-Detector Edge Processing in Resource-Constrained Extreme Radiation Environment

Fermi National Accelerator Laboratory will demonstrate that a neural network (NN) autoencoder model can be implemented in a radiation-tolerant application-specific integrated circuit (ASIC) to perform lossy data compression alleviating the data transmission problem while preserving critical information of the detector energy profile. This AI algorithm enables specialized compute capability and has been optimized for data compression in the trigger path of the High-Granularity Endcap Calorimeter (HGCal), an upgrade for the Compact Muon Solenoid (CMS) experiment for the high-luminosity LHC (HL-LHC). The implementation of a complex neural network algorithm demonstrates the effectiveness of Catapult High-Level Synthesis (HLS) based design automation flow utilizing the hls4ml framework for developing design IP for ASICs. The low-power, low-latency hardware accelerator is designed to explore the use of unsupervised machine learning methods to obtain 7x to 16x data compression at inference rates of 40 MHz. The objective encoding can be adapted based on detector conditions and geometry by updating the trained weights. The design has been implemented in an LP CMOS 65 nm process. It occupies a total area of 2.5 mm2, consumes 80 mW of power and is optimized to withstand approximately 200 MRad ionizing radiation.