Deep learning (DL) is crucial for various applications, but the rapid evolution of DL models and systems integration challenges custom chip creation. Field-programmable gate arrays (FPGAs) offer customizable hardware execution, accelerating DL inference more efficiently than CPUs and GPUs. FPGAs allow for tailored processing pipelines, lower latency, and higher energy efficiency. Academic and industrial enhancements in FPGA architecture for DL are discussed, from model-specific dataflow to software-programmable overlays. DL-specific improvements to logic blocks, circuitry, and on-chip memories are highlighted. Hybrid devices combining processors and accelerator blocks show promise for future research. FPGAs are evolving to meet the increasing demand for high DL performance.
https://arxiv.org/abs/2404.10076