Automatic Hybrid-Precision Quantization for MIMO Detectors

TL;DR

This paper introduces an automatic hybrid-precision quantization framework for MIMO detectors, significantly reducing bitwidth and improving efficiency without sacrificing performance, verified through CMOS implementation and outperforming state-of-the-art methods.

Contribution

The paper proposes a novel automatic hybrid-precision quantization framework (AHPQ) that optimizes quantization for MIMO detectors using PDF and deep reinforcement learning, enhancing efficiency and flexibility.

Findings

Achieves up to 58.7% lower average bitwidth compared to uniform quantization.

Realizes 2.97× higher throughput-to-area ratio with 19.3% lower energy dissipation.

Outperforms state-of-the-art in throughput (17.92 Gb/s) and energy efficiency (7.93 pJ/b).

Abstract

In the design of wireless systems, quantization plays a critical role in hardware, which directly affects both area efficiency and energy efficiency. Being an enabling technique, the wide applications of multiple-input multiple-output (MIMO) heavily relies on efficient implementations balancing both performance and complexity. However, most of the existing detectors uniformly quantize all variables, resulting in high redundancy and low flexibility. Requiring both expertise and efforts, an in-depth tailored quantization usually asks for prohibitive costs and is not considered by conventional MIMO detectors. In this paper, a general framework named the automatic hybrid-precision quantization (AHPQ) is proposed with two parts: integral quantization determined by probability density function (PDF), and fractional quantization by deep reinforcement learning (DRL). Being automatic, AHPQ demonstrates high efficiency in figuring out good quantizations for a set of algorithmic parameters. For the approximate message passing (AMP) detector, AHPQ achieves up to $58.7\%$ lower average bitwidth than the unified quantization (UQ) one with almost no performance sacrifice. The feasibility of AHPQ has been verified by implementation with $65$ nm CMOS technology. Compared with its UQ counterpart, AHPQ exhibits $2.97\times$ higher throughput-to-area ratio (TAR) with $19.3\%$ lower energy dissipation. Moreover, by node compression and strength reduction, the AHPQ detector outperforms the state-of-the-art (SOA) in both throughput ($17.92$ Gb/s) and energy efficiency ($7.93$ pJ/b). The proposed AHPQ framework is also applicable for other digital signal processing algorithms.