High-Accuracy, High-Performance In-Memory Computing With High-Resistance Spin-Orbit Torque (SOT) Magnetic Memory

In-memory computing (IMC) has emerged as a promising solution for artificial intelligence (AI) accelerators due to the reduced data movement and improved parallelism in the crosspoint array. However, IMC faces several limitations such as the device variations affecting the computing accuracy, the area- and energy-consuming peripheral circuitry, and the time-consuming high-voltage program-verify operations of the nonvolatile memory (NVM). In addition, the relatively large summation currents cause IR drop within the array, leading to further loss of accuracy. To overcome these issues, this work presents 1-bit quantized networks based on spin-orbit-torque magnetoresistive random access-memory (SOT-MRAM) with high resistance-area (RA) product. We provide a detailed statistical characterization of SOT-MRAM arrays and develop quantization-aware training of various neural networks. Our results indicate that SOT-MRAM enables: 1) high inference accuracy, thanks to excellent uniformity; 2) negligible input-dependent IR drop nonlinearity, thanks to high resistance; and 3) high-speed, low-power reconfiguration, thanks to fast device programming. These results support high-RA SOT-MRAM for digital-like and reconfigurable IMC accelerators of edge AI.