Abstract
Magneto-resistive RAM (MRAM) is an emerging non-volatile memory technology with a very wide range of potential applications. In particular, Spin-Transfer-Torque (STT) memories, which leverage spin polarized current as the switching mechanism, offer advantages in almost all aspects and memory performance metrics. These include direct integration with CMOS technology, very good density and scalability, near-zero static power consumption, high access times, and long endurance. STT memories have already been utilized in commercial applications, primarily as caches, by major hardware manufacturers. More recently, STT-based In-Memory Computing systems have also been developed. The majority of these designs utilize bit-wise digital processing inside the STT cells, by enabling multiple cells at the same time and modifying the peripheral sensing circuits. Modifications to the elementary MTJ cell in order to facilitate processing capabilities have also been proposed, but require further technological improvements to be applicable. Despite the low resistance ratio between parallel and anti-parallel states, and due to the attractive properties of the STT memories, particularly in terms of energy consumption and technology maturity (as compared to other emerging resistive memories), analog IMC systems that utilize STT memories have also been demonstrated.
Original language | British English |
---|---|
Title of host publication | In-Memory Computing Hardware Accelerators for Data-Intensive Applications |
Publisher | Springer Nature |
Pages | 57-79 |
Number of pages | 23 |
ISBN (Electronic) | 9783031342332 |
ISBN (Print) | 9783031342325 |
DOIs | |
State | Published - 25 Sep 2023 |
Keywords
- Data-centric computing
- Efficient computing
- In-memory computing
- Machine learning
- MRAM
- Near-memory computing