Table 3.

Comparison of hardware accelerators for the implementation of federated learning in edge computing.

Name	Owner	Pros	Cons
CPU/GPU	NVIDIA and Radeon	High memory, bandwidth, and throughput	Consumes a large amount of power
FPGA	Intel	High performance per watt of power consumption, reduced costs for large-scale operations, excellent choice for battery-powered devices and on cloud servers for large applications	It requires a significant amount of storage, external memory and bandwidth, and computational resources on the order of billions of operations per second
ASIC	Intel	Minimizes memory transfer, most energy efficient compared to FPGAs and GPUs, and best computational speed compared to FPGAs and GPUs	Long development cycle, Lack of flexibility to handle varying DL network designs