講演抄録/キーワード |
講演名 |
2023-08-08 13:25
大規模確率的演算システムに向けた超伝導ストカスティックメモリの設計 ○羅 文輝(横浜国大)・竹内尚輝(産総研)・陳 オリビア(東京都市大)・吉川信行(横浜国大) SCE2023-7 |
抄録 |
(和) |
Stochastic computing-based neural networks (SC-NNs) have been developed for energy-efficient neuromorphic computing. In SC, a stochastic number generator (SNG) is used to convert a binary input into a corresponding stochastic bit stream, namely a stochastic number (SN). In general, an SNG comprises a digital comparator and random number generator (RNG), which requires a relatively large hardware overhead and becomes unsuitable for large-scale SC systems. In the present study, we propose the superconducting stochastic memory (SSM) as a hardware-efficient SNG for SC-based adiabatic quantum-flux-parametron (AQFP) circuits. The SSM has a simple structure comprising a flux storage loop and a QFP, which can generate stochastic bit streams with an arbitrary probability distribution, controlled by the stored flux quanta. We experimentally demonstrate that the probability of 1s in the generated bit stream (i.e., stochastic number) can be controlled by the number and polarity of stored flux quanta. Furthermore, we show that the SSM occupies much less power dissipation of the entire SC system compared with the conventional comparator-based SNG. |
(英) |
Stochastic computing-based neural networks (SC-NNs) have been developed for energy-efficient neuromorphic computing. In SC, a stochastic number generator (SNG) is used to convert a binary input into a corresponding stochastic bit stream, namely a stochastic number (SN). In general, an SNG comprises a digital comparator and random number generator (RNG), which requires a relatively large hardware overhead and becomes unsuitable for large-scale SC systems. In the present study, we propose the superconducting stochastic memory (SSM) as a hardware-efficient SNG for SC-based adiabatic quantum-flux-parametron (AQFP) circuits. The SSM has a simple structure comprising a flux storage loop and a QFP, which can generate stochastic bit streams with an arbitrary probability distribution, controlled by the stored flux quanta. We experimentally demonstrate that the probability of 1s in the generated bit stream (i.e., stochastic number) can be controlled by the number and polarity of stored flux quanta. Furthermore, we show that the SSM occupies much less power dissipation of the entire SC system compared with the conventional comparator-based SNG. |
キーワード |
(和) |
/ / / / / / / |
(英) |
stochastic computing / superconducting stochastic memory / / / / / / |
文献情報 |
信学技報, vol. 123, no. 153, SCE2023-7, pp. 34-38, 2023年8月. |
資料番号 |
SCE2023-7 |
発行日 |
2023-08-01 (SCE) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
SCE2023-7 |
|