Lightweight programmable DSP block overlay for streaming neural network acceleration

Implementations of hardware accelerators for neural networks are increasingly popular on FPGAs, due to flexibility, achievable performance and efficiency gains resulting from network optimisations. The long compilation time required by the backend toolflow, however, makes rapid deployment and prototyping of such accelerators on FPGAs more difficult. Moreover, achieving high frequency of operation requires significant low-level design effort. We present a neural network overlay for FPGAs that exploits DSP blocks, operating at near their theoretical maximum frequency, while minimizing resource utilization. The proposed architecture is flexible, enabling rapid runtime configuration of network parameters according to the desired network topology. It is tailored for lightweight edge implementations requiring acceleration, rather than the highest throughput achieved by more complex architectures in the datacenter.