Implementation of Reconfigurable Deep Learning Accelerator (RDLA) on PolarFire SoC

In neural networks and other computationally demanding applications, general-purpose CPUs are slow and ineffective. To run such applications, it is better to create specialized hardware capable of doing several multiply-accumulate operations quickly and effectively. For a wide range of neural network applications, the Reconfigurable Deep Learning Accelerator (RDLA) architecture has been developed. The fundamental unit of the RDLA is composed of a variety of Multiply-Accumulate (MAC) units, registers, and Address Generation Units (AGU). On the PolarFire SoC, RDLA was tested and implemented with a clock frequency of up to 62.5MHz for data processing. This paper shows the results testing with different images for a custom MNIST model with 4 layers with accuracy of 97.49% with power consumption of 1.85W.