Combining Reconfiguration and Instruction Computations with an Asynchronous Crossbar

Abstract

With the rapid development of artificial intelligence (AI) technology, a more flexible and efficient computing architecture is required. The coarse-grained reconfigurable architectures (CGRAs) as a co-processor that combines the flexibility and specificity of AI would be an appropriate solution. In this paper, we introduce an innovative asynchronous CRGA architecture that combines reconfiguration and specific instructions. We aim to accelerate a common computation with memorized calculation and keep the reconfiguration for rare ones, so that the whole efficiency could be enhanced. To show the feasibility of our method, we choose the sigmoid function that is essentially and widely used in AI algorithms, which can achieve about 40Mhz configuration throughput and about 669ns to finish the whole calculation.