BROCA: A 52.4-to-559.2mW Mobile Social Agent System-on-Chip with Adaptive Bit-Truncate Unit and Acoustic-Cluster Bit Grouping computation. The ACE PE processes input data in a bit-serial manner, where four 8b weights are accumulated in parallel, leading to a decrease in compute energy proportional to the input bitwidth. As a result, an average reduction of 4.4b in input bitwidth and a 4.7× speedup can be achieved, reducing computation energy by 44.5% at the RG.

📋 论文概要

论文提出了一款名为BROCA的移动社交代理系统级芯片（SoC），通过自适应位截断单元（ACBU）和声学聚类（Acoustic Cluster）技术，在52.4至559.2mW功耗范围内实现运行时比特宽度缩减，有效降低音频处理能耗。该芯片针对音频特征图的通道维度局部性，通过聚类窄范围特征图元素进行高位截断，从而优化计算效率。

💡 主要创新点

• 提出ACBU（自适应位截断单元），利用音频特征图在通道上的幅度局部性，将窄范围元素聚成声学簇并执行MSB截断，实现运行时位宽动态缩减以减少计算能耗。
• 实现了一种面向移动社交代理的低功耗SoC架构，在宽功耗范围内（52.4-559.2mW）提供可伸缩的音频处理能力。

🏷 关键词

📄 原文摘要

Dongseok Im, Sangyeob Kim, Sangjin Kim, Taekwon Lee, Hoi-Jun Yoo Figure 23.7.4 illustrates the proposed ACBU, which achieves runtime bitwidth reduction of the vocoder input feature map to reduce computation energy in the AF. The motivation for ACBU originates from the longitudinal characteristics of audio, which demonstrate channelwise confined magnitude locality input feature maps. ACBU exploits this locality by grouping the narrow-ranged feature map elements into an acoustic cluster (AC) for each channel and performing MSB truncation. For a new timestep element, the overflow processing unit (OPU) first checks if incorporating the element would require an additional bit. If so, a new AC is initiated starting from this element, and no additional elements are added to the old cluster. Subsequently, the narrow-ranged combination tracking unit (NCTU) examines if including the new timestep can yield a narrow-ranged (NR) combination that represents the data with

👥 作者与机构

Wooyoung Jo, Seongyon Hong, Jiwon Choi, Beomseok Kwon, Haoyang Sang,