Core Concepts
A reinforcement learning-based framework, RL-MUL, is proposed to efficiently optimize the design of multipliers and fused multiply-accumulators (MACs) by leveraging matrix and tensor representations to enable seamless integration of deep neural networks as the agent network.
Abstract
The paper proposes RL-MUL, a reinforcement learning-based framework for optimizing the design of multipliers and fused multiply-accumulators (MACs). The key highlights are:
RL-MUL utilizes matrix and tensor representations to characterize the multiplier architecture, enabling the seamless integration of deep neural networks as the agent network.
A Pareto-driven reward mechanism is introduced to encourage the RL agent to learn Pareto-optimal designs, balancing the trade-off between area, delay, and power.
The framework is extended to support the optimization of fused MAC designs, where the accumulation is integrated into the partial product stages of multiplication.
To improve search efficiency, RL-MUL leverages a parallel training methodology to enable faster and more stable training.
Experimental results demonstrate that the multipliers and MACs produced by RL-MUL outperform various baseline designs in terms of both area and delay. Applying the optimized multipliers and MACs to a larger computation module also results in improved power, performance, and area.
Stats
Multiplication can constitute over 99% of operations in standard deep neural networks.
The ratios of MAC computations in various neural networks range from 90% to 100%.