Speaker
Description
The Electron-Ion Collider (EIC) under construction at Brookhaven National Laboratory is a next-generation facility designed to systematically investigate the gluonic and spin structure of nucleons and atomic nuclei. To fulfill the stringent physics requirements of the EIC ePIC detector, the dual-radiator Ring Imaging Cherenkov (dRICH) detector provides particle identification (PID) over a wide momentum range, using aerogel and heavy gas radiators and an array of Silicon Photomultipliers (SiPMs). However, the high radiation environment is expected to significantly degrade SiPM sensor efficiency over the experiment lifetime, driving Dark Count Rates (DCR) up to 300 kHz per single channel. This generates a thermal noise floor that threatens to saturate the 3 Tbps aggregate dRICH readout bandwidth.
To resolve this critical communication constraint, this work presents an FPGA-accelerated real-time data reduction framework designed to suppress DCR noise-driven throughput while maximizing physics signal sensitivity. The system is deployed as a hierarchical Artificial Intelligence (AI) pipeline distributed across 30 front-end Data Aggregation Module (DAM) boards and a centralized Trigger Processor (TP) board.
To enable deployment on resource-constrained dRICH Data Acquisition (DAQ) system FPGAs, the networks underwent heterogeneous fixed-point Quantization Aware Training (QAT) via the HGQ2 framework. Converted into C++ High-Level Synthesis (HLS) code via the hls4ml framework, the quantized model correctly classified an 83.8 ± 0.4% of the physics events while providing a data reduction factor of ~5 at the maximum 300 kHz DCR.
Furthermore, by integrating the da4ml library into the synthesis flow, dense Constant Matrix-Vector Multiplications (CMVM) operations were converted into Distributed Arithmetic (DA) structures mapped directly onto FPGA Look-Up Tables (LUTs). This eliminates the use of dedicated Digital Signal Processing (DSP) blocks, saving vital resources for standard ePIC Data Acquisition (DAQ) routines.
The fully synthesized DSP-free architecture was validated on different AMD Versal Premium VPK180 hardware testbeds, achieving a sustained, fully pipelined throughput of ~100 MHz at an Initiation Interval of II = 1, matching the ~10 ns bunch-crossing rate of the ePIC experiment.