Performance Benchmarking of Spin-Orbit Torque Magnetic RAM (SOT-MRAM) for Deep Neural Network (DNN) Accelerators
2022 IEEE International Memory Workshop (IMW), 2022•ieeexplore.ieee.org
In this paper, the system level evaluation is performed for DNN inference engines using SOT-
MRAM, which includes compute-in-memory (CIM) paradigm and near-memory systolic
array. The write performance of the SOT materials is projected to 7nm with a macrospin
model. For read-intensive CIM, SOT-MRAM with increased on-resistance can achieve 51%
and 93% higher energy efficiency than 8T-SRAM at 22nm and 7nm nodes, respectively. For
write-intensive systolic array at 7nm node, SOT-MRAM with PtCu track shows 17% higher …
MRAM, which includes compute-in-memory (CIM) paradigm and near-memory systolic
array. The write performance of the SOT materials is projected to 7nm with a macrospin
model. For read-intensive CIM, SOT-MRAM with increased on-resistance can achieve 51%
and 93% higher energy efficiency than 8T-SRAM at 22nm and 7nm nodes, respectively. For
write-intensive systolic array at 7nm node, SOT-MRAM with PtCu track shows 17% higher …
In this paper, the system level evaluation is performed for DNN inference engines using SOT-MRAM, which includes compute-in-memory (CIM) paradigm and near-memory systolic array. The write performance of the SOT materials is projected to 7nm with a macrospin model. For read-intensive CIM, SOT-MRAM with increased on-resistance can achieve 51% and 93% higher energy efficiency than 8T-SRAM at 22nm and 7nm nodes, respectively. For write-intensive systolic array at 7nm node, SOT-MRAM with PtCu track shows 17% higher energy efficiency than SRAM global buffer, respectively.
ieeexplore.ieee.org
以上显示的是最相近的搜索结果。 查看全部搜索结果