Cost-effective stochastic MAC circuits for deep neural networks.

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Hardware Acceleration

Background:

• Stochastic computing (SC) offers a path for future hardware, but its imprecision and latency hinder deep neural network (DNN) efficiency.
• Existing SC-based DNNs often sacrifice accuracy for performance, making them less competitive than fixed-point implementations.

Purpose of the Study:

• To develop a novel, highly efficient, and accurate SC-MAC algorithm for DNNs.
• To demonstrate the applicability of the new SC-MAC for accelerating convolutional neural networks (CNNs).

Main Methods:

• Proposed a new SC-MAC algorithm, extending it to a vector version for broader CNN layer acceleration.
• Implemented and evaluated SC-based CNNs on MNIST and CIFAR-10 datasets.
• Validated the approach through FPGA prototypes.

Main Results:

• The new SC-MAC algorithm is orders of magnitude more efficient and accurate than prior SC-MACs.
• SC-based CNNs achieved 40–490x greater energy efficiency in convolution layers compared to conventional SC methods.
• Achieved lower area-delay product and energy consumption than optimized fixed-point implementations without accuracy loss.

Conclusions:

• The proposed SC-MAC algorithm significantly enhances the viability of SC for efficient DNN hardware.
• SC-based CNNs demonstrate competitive or superior performance to fixed-point designs, validated by FPGA prototypes.