CNN-on-AWS: Efficient Allocation of Multikernel Applications on Multi-FPGA Platforms

Multi-FPGA platforms, like Amazon AWS F1, can run in the cloud multikernel pipelined applications, convolutional neural networks (CNNs), with excellent performance and lower energy consumption than CPUs or GPUs. We propose a method to efficiently map these applications on multi-FPGA platforms maximize application throughput. Our methodology finds, for given resources, optimal number of parallel instances each kernel pipeline their allocation one more among available FPGAs. obtain this by formulating solving mixed-integer, nonlinear optimization problem, which we model component duration phases accelerated computation be split into, namely: 1) data transfer from host CPU DDR memory FPGA; 2) FPGA on-chip memory; 3) 4) DDR; 5) host. Finding solution using mixed-integer programming (MINLP) solver is often highly inefficient. Hence, provide fast heuristic that according our experiments much efficient MINLP finds comparable results. For larger problems (more CNN layers), quickly find (several thousand times faster) better solutions solver, even if latter very long time.