A 28nm 27.5TOPS/W Approximate-Computing-Based Transformer Processor with Asymptotic Sparsity Speculating and Out-of-Order Computing

📋 论文概要

该论文提出了一款基于近似计算的Transformer处理器，采用渐进式稀疏推测和乱序执行技术，在28nm工艺下实现了27.5TOPS/W的能效。通过BESA PE（位级近似计算）降低MAC能量40.8%，在GPT-2情感分析任务中达到89.6%准确率，仅比精确计算低0.4%。

💡 主要创新点

• 提出近似计算方法，利用6位MSB级联OR/AND生成多位0信号以禁用压缩器，降低能耗。
• 采用渐进稀疏推测和乱序执行机制，提高Transformer推理效率。
• 引入BESA PE（位级近似计算处理单元），在保证高准确率的同时大幅减少MAC能量消耗。

🏷 关键词

📄 原文摘要

code generator performs cascaded OR (AND) for positive (negative) data via 6b MSBs to generate a 6b signal with more 0-bits for small values to disable compressors. For exact mode, it modifies 2 BVs to “0” to make approximate PPs be exactly “0”, ensuring exact results for approximate compressors. The result is compensated with addition if a modified bit changes from “1” to “0” or subtraction for “0” to “1”. The BESA PE reduces 40.8% of the MAC energy for the attention block of GPT-2 with 89.6% accuracy for sentiment analysis, only 0.4% less accuracy than entirely exact computing. 1 2 Recently, Transformer-based models have achieved tremendous success in many AI fields, from NLP to CV, using the attention mechanism [1-3]. This mechanism captures

👥 作者与机构

Yang Wang1, Yubin Qin1, Dazheng Deng1, Jingchuan Wei1, Yang Zhou1,

Yuanqi Fan1, Tianbao Chen2, Hao Sun1, Leibo Liu1, Shaojun Wei1, Shouyi Yin1 Tsinghua University, Beijing, China Tsing Micro, Beijing, China approximate mode, a self-gating