可變電阻式記憶體類神經網路加速器: 效能與耗能模擬框架

ReRAM-based Neural Network Accelerator: Performance and Energy Consumption Simulation Framework

林蔚廷(Wei-Ting Lin)

指導教授 ： 楊佳玲

國立臺灣大學/電機資訊學院/資訊網路與多媒體研究所/碩士(2021年)

https://doi.org/10.6342/NTU202100826

摘要

可變電阻式記憶體之神經網路加速器利用記憶體式運算技術計算類神經演算法，記憶體式運算技術能使得記憶體不僅能夠儲存神經網路之權重，還能夠實現向量與矩陣乘法，因此可以提升系統的能源效率。由於不同類神經網路的權重配置、排程以及硬體設置皆會影響加速器的效能和耗能，為了設計高效率或低耗能的可變電阻式記憶體之神經網路加速器，我們會需要一套模擬框架分析不同設計對系統的效能與耗能的影響。
此篇論文中，我們提出了一個可變電阻式記憶體之神經網路加速器模擬框架，此框架可根據使用者選擇的權重配置方法、排程方式與硬體配置當作輸入參數，模擬加速器的效能與耗能，此模擬框架由數個模組組成，使用者除了使用預設選項當模擬器的輸入外，還能夠改寫模擬框架的模組，來彈性地達成不同的權重配置、排程。
我們使用多個不同的卷積神經網路為討論對象，以證實此模擬框架可提供使用者設計觀點，幫助設計可變電阻式記憶體之神經網路加速器。

關鍵字

神經網路 ； 可變電阻式記憶體 ； 加速器架構 ； 模擬器

並列摘要

ReRAM-based Neural Network (NN) accelerator utilizes in-memory computing technology to perform NN algorithm. The in-memory computing enables memory can not only store the weights of NN but also perform matrix-vector multiplication, which improves the energy efficiency of system. Different mapping of NN's weights, scheduling policy and hardware configuration setting affect performance and energy consumption of accelerator. To design a high performance and low energy consumption ReRAM-based NN accelerator, a flexible simulation platform is needed.
In this paper, we proposed a simulation framework of ReRAM-based NN accelerator. The users could select the mapping policy, scheduling policy and hardware configuration as input parameter to simulate the performance and energy consumption of ReRAM-based NN accelerator. The simulation framework includes multiple modules. Beside using default input options, users can rewrite the module of simulation framework to flexibly achieve different weight mapping and scheduling.
We use multiple convolutional neural networks as case studies to show that the simulation framework can provides insights for users to design the ReRAM-based NN accelerator.

並列關鍵字

Neural network ； ReRAM ； Accelerator architecture ； Simulator

參考文獻

A. Shafieeet al., “Isaac: A convolutional neural network accelerator with in-situ analogarithmetic in crossbars,” inISCA, 2016, pp. 14–26.

P. Chiet al., “Prime: A novel processing-in-memory architecture for neural network com-putation in reram-based main memory,” inISCA, 2016, pp. 27–39.

W. H. Chenet al., “A 65nm 1mb nonvolatile computing-in-memory reram macro with sub-16ns multiply-and-accumulate for binary dnn ai edge processors,” inISSCC, 2018, pp. 494–496.

F. Suetal., “A 462gops/j rram-based nonvolatile intelligent processor for energy harvestingioe system featuring nonvolatile logics and processing-in-memory,” inVLSI-TSA, 2017,T260–T261.

P. Y. Chenet al., “Neurosim: A circuit-level macro model for benchmarking neuro-inspiredarchitectures in online learning,”IEEE TCAD, pp. 1–1, 2018.