BEOL compatible high retention perpendicular SOT-MRAM device for SRAM replacement and machine learning

Spin orbit torque-MRAM devices are promising both for high performance cache replacement and for low power neural networks requiring non-volatile weights, such as analog in-memory compute (AiMC). Using a new free layer design, we demonstrate a BEOL compatible perpendicular SOT device with high retention and excellent switching efficiency. ∆ > 75 kBT is reported for 50nm devices at 125°C operating temperatures with critical current < 400µA at 1ns. This concept offers a flexible way to adjust ∆ and to exploit advanced SOT material both for high performance and machine learning (ML) applications.