- 【Updated on May 12, 2025】 Integration of CiNii Dissertations and CiNii Books into CiNii Research
- Trial version of CiNii Research Automatic Translation feature is available on CiNii Labs
- Suspension and deletion of data provided by Nikkei BP
- Regarding the recording of “Research Data” and “Evidence Data”
66 TOPS/W FeFET CiM with Multiply-Accumulate by 32 ML & 1024 AL Parallel Source-follower Read and Charge-sharing
Bibliographic Information
- Other Title
-
- ソースフォロワ読み出し・チャージシェアリングにより32 ML & 1024 AL並列で積和演算を行う66 TOPS/W強誘電体FET CiM
Description
読み出し電流の On/Off 比が高い強誘電体 FET (FeFET) を用いた,電圧センス型 Computation-in-Memory (CiM) を提案した.AND フラッシュ回路のように接続された複数の FeFET セルに,ニューラルネットワークの重みを保存する.前段の pre-neuron の発火であるニューラルネットワークの入力は FeFET のソースに与える.Phase 1) ソースフォロワ読み出しによって入力と重みの積を 32 本並列に Multiply-line (ML) に読み出し,Phase 2) 1024 本の Accumulate-line (AL) の配線容量を並列にチャージシェアすることで積の結果を合計し,積和演算結果を得る.従来の電流センス型 CiM と異なり,提案の FeFET を用いた電圧センス型 CiM は積演算時に DC 電流が流れず,和演算時に電力を消費しないため,並列に MAC 演算を行うことができ,66 TOPS/W の高スループット・高エネルギー効率を実現できる.
A voltage-sensing computation-in-memory using ferroelectric FETs (FeFETs) with high on/off current ratio. To store the weights of the neural networks, multiple FeFETs are connected like AND-type flash array. As pre-neurons information of the neural networks, the input is given to the source-line of the FeFET. Multiply-accumulate (MAC) result is obtained by Phase 1) multiplying inputs and weights by source-follower read to 32 multiply-lines (MLs) in parallel and Phase 2) accumulating capacitance of parallel 1024 accumulate-lines (ALs) by charge-sharing. Unlike the conventional current-sensing CiM, DC current flows in multiply-phase, and no active power is consumed in accumulate-phase in the proposed voltage-sensing FeFET CiM. As a result, the proposed FeFET CiM achieves 66 TOPS/W high throughput and high energy efficiency by parallel MAC operation.
Journal
-
- DAシンポジウム2021論文集
-
DAシンポジウム2021論文集 2021 61-62, 2021-08-25
情報処理学会
- Tweet
Keywords
Details 詳細情報について
-
- CRID
- 1050855522098374528
-
- NII Article ID
- 170000185145
-
- Text Lang
- ja
-
- Article Type
- conference paper
-
- Data Source
-
- IRDB
- CiNii Articles