應用於非揮發記憶體內運算之高阻態單元偏好神經網路權重量化演算法

High-Resistance-State-Favored Quantization Algorithm Based on Non-Volatile Computing In Memory for Neural Network

劉大維(Ta-Wei Liu)

指導教授 ： 張孟凡

國立清華大學/電機資訊學院/電機工程學系所/碩士(2020年)

摘要

隨著人工智慧的浪潮愈演愈烈，深度學習演算法也愈趨成熟，因此神經
網路(NN) 硬體加速器的研究也如雨後春筍般出現，而記憶體內運算便是其
中一種打破傳統逢紐曼架構的記憶體導向特殊硬體。記憶體內運算的優勢
是同時具有運算以及儲存的功能，在記憶體端先進行運算再傳出結果，減
少移動的資料量，節省所需時間與能源。本篇論文針對一篇已經發表過的
非揮發記憶體內運算(nv-CIM) 作為實驗平台，其電路有著一個可以增進讀
取良率的電路(in-situ HRS-C)，而本篇論文利用了其電路的附帶特性，設計
一種特殊的神經網路權重量化流程，這個流程可以被有效加入到網路中，
使得在使用這個硬體進行AI 辨識時，在付出些微的精準度下降後，能使
TOPS/W 得到4.6 到10.51 的提升，同時也在網路訓練的過程中加入了CIM
的行為模型，這個行為模型能模擬感測放大器的行為，使訓練出來的網路
能包容因為CIM 中的類比運算誤差。進一步達到與純軟體接近的準確度。

關鍵字

非揮發記憶體內運算 ； 神經網路 ； 權重量化 ； 人工智能 ； 友善量化 ； 降低功耗

並列摘要

With the rapid development of artificial intelligence and a more mature deep
learning algorithm, as a result, the number of DNN accelerator research has sprung
up, and In-Memory Computing is one of them, which is used to overcome the Von
Neumann bottleneck. The advantage of in-memory computing is that it has both
computation and storage functions. It performs operations on the memory side and
then transmits the results, reducing the amount of data moved, saving time and
energy. This paper aims at a published non-volatile memory in-memory computing
(nvCIM) as an experimental platform. It has a unique circuit (in-situ HRS-C)
that can improve the read yield. However, this thesis applies to this circuit’s additional
characteristics and provides a neural network parameter training process.
This training flow can be effectively added to the neural network. When this hardware
is used for AI classification, the proposed training flow makes it possible to
achieve a reasonable tradeoff between energy efficiency and inference accuracy.
It can improve TOPS/W from 4.6 to 10.51 with a negligible inference accuracy
drop. At the same time, the CIM behavior model also applies to the network training
process. This behavior model can simulate the sensing amplifier’s behavior.
The trained network can accommodate the CIM’s analog error and further achieve
accuracy close to pure software.

並列關鍵字

Non-Volatile Computing In Memory ； Neural Network ； Quantization Algorithm ； artificial intelligence ； friendly quantization ； low power

參考文獻

10.1109/ISSCC19947.2020.9063078, pp. 244–246, 02 2020.

10.1109/ISSCC.2018.8310400, pp. 494–496, 02 2018.

10.1109/ISSCC.2019.8662395, pp. 388–390, 02 2019.

of a tensor processing unit,” 10.1145/3079856.3080246, pp. 1–12, 06 2017.

10.1109/ISSCC.2019.8662302, pp. 142–144, 02 2019.