Optimizing FPGA-based Sparse Coding for Real-time Processing in X-ray Instrumentation

As detectors improve, large physics experiments generate more data. However, real-time raw data processing near the sensors can be challenging due to the high data velocity and the resource-constrained location. A practical solution involves using low-latency data reduction to lower the incoming throughput. Sparse representations and machine learning are popular methods for data reduction, with the Learned Iterative Shrinkage-Thresholding Algorithm, dubbed LISTA, combining both techniques to produce fast approximations of sparse coding. In previous work, we proposed a multi-stage strategy to compress the data at the edge. We targeted the billion-pixel X-ray camera, an X-ray imaging system with a data rate of 10 TB/s. The compression leverages a LISTA-like architecture followed by quantization and entropy coding, achieving a 100:1 compression ratio with minimal image artifacts. Using high-level synthesis, we fitted the fully connected LISTA on an FPGA. Although the inference time was around 1 μs, the number of parameters per layer of the network was limited, imposing a low input size and restricting the FPGA throughput. In this work, we explore and compare solutions for a larger input size that maintain similar performance. We studied splitting the initial linear layers into smaller ones using low-rank matrix factorization. Preliminary results indicate that the factorization can achieve a 77% increase in the input size, with increased throughput at the cost of added resources. Finally, we discuss the current limitations of the compression strategy and potential improvements moving forward.