부산대학교 도서관

Binarized Neural Networks (BNNs) offer the promise of low power and high throughput, but this is difficult to achieve on regular processors. A considerable amount of research has been devoted to mapping BNNs to specialized hardware, especially FPGAs, setting aside the flexibility of instruction-set processors. This paper introduces a configurable VLIW processor with a specialized instruction set to efficiently compute the inference of Look-Up-Table based artificial binary neurons in a single clock cycle. Our experiments show that the processor achieves an increased throughput of 2994 when compared to the inference done on a base processor. ×