Abstract

The applications of the Artificial Intelligence have been increasingly used with huge datasets for many purposes. The beyond Von Neumann architectures (like digital and analog in-memory computing) are proposed to mitigate the data-movement bottleneck. However, they are struggling with the limitations of the conventional data representations: either the computation complexity of the digital binary domain or the interfacing and scalability issues of the analog domain; Meanwhile, the stochastic computing domain suffers from the generation complexity bottleneck which degrades the benefits of its computation simplicity. This paper presents a new Bent-Pyramid system which acts as a quasi-stochastic data representation. The new Bent-Pyramid system utilizes two complementary fixed sets of bitstreams to perform deterministic multiplication. The Bent-Pyramid inherits the same multiplication simplicity of the stochastic computing while avoiding the stochastic number generation complexity. The Vector-Matrix Multiplication benchmarking shows that the 10bit Bent-Pyramid system has a comparable accuracy to the 16bit stochastic computing. The generation circuit of the 10-bit Bent-Pyramid reduces the energy and the latency of the 16-bit stochastic counterpart by 15.15x and 16.0x respectively.