神經網路和記憶體內運算架構共同設計

Abstract

記憶體內運算架構已證明其有效解決記憶體牆瓶頸的能力，神經網路結構搜索致力於自動化機器學習模型的設計。然而，若欲整合記憶體內運算架構至神經網路結構搜索中，將出現重大的挑戰。在記憶體內運算加速器上部署神經網路，會引入與硬體相關的因素，導致大量額外的模擬負擔。本論文透過結合量化和裝置感知的準確率預測器，介紹了一種超高效的記憶體內運算和神經網路的架構搜索框架。此外，我們邁出了第一步，在定量分析資訊如何於記憶體內運算的神經網路加速器中傳播，以及額外的記憶體內運算因素如何影響該訊息傳播。另一方面，本論文介紹了第一個利用記憶體內運算的最佳化機會，來解決記憶體效率低下問題的點雲深度學習分析加速器，而後我們也基於所提出的記憶體內運算架構，利用神經網路結構搜索的技術來探索最佳的點雲模型。

Computing-in-memory (CIM) architecture has demonstrated its ability to address the memory wall bottleneck effectively. Neural architecture search (NAS) endeavors to design machine learning models automatically. However, integrating CIM into NAS presents a significant challenge. Deploying neural networks on CIM accelerators introduces hardware-related factors, resulting in substantial additional simulation overhead. This dissertation introduces an ultra-efficient CIM-NAS framework by incorporating a quantization and device aware accuracy predictor. In addition, we take the first step in quantitative analysis of how information propagates in CIM neural accelerators and how additional CIM factors influence that information propagation. On the other hand, this dissertation introduces the first deep point cloud (PC) analytics, an emerging machine learning application, accelerator that leverages CIM optimization opportunities to address memory inefficiency. We also explore optimal PC models based on the proposed CIM architecture using NAS techniques.