Numem to supply its STT-MRAM to a NASA AI core project

High-performance STT-MRAM developer Numem announced that it has been selected for a NASA AI project, for which the company will supply its Numem NuRAM MRAM-based Memory. Numem says its memory enables a 2-3x smaller memory area and 20x to 50x lower standby power compared to SRAM.

The project, titled “DNN Radiation Hardened Co-processor as companion chip to NASA’s upcoming High-Performance Spaceflight Computing Processor” will develop a reconfigurable DNN Engine with multiple compute units which can support a wide range of DNN models and frame rates.

Superlattice and half-metallic magnets used to developed SS-MRAM, an ultra-high performance MRAM device

Researchers from the National Taiwan University developed an ultra-high performance MTJ, using a superlattice barrier and half-metallic magnets. The so-called superlattice-MTJ can be the basis of a new class of STT-MRAM (which the researcher call SS-MRAM) that achieves ultra-low power RA and write operations, high writing speed and unlimited endurance.

Geometric structure of a three-cell superlattice MTJ (National Taiwan University)

SS-MRAM adopts a superlattice barrier that replaces the MgO layer used in common STT-MRAM. The MgO layer is unstable and also suffers from a very large RA which results in high power consumption for writing operations. The superlattice has higher spin polarization than MgO and so the SS-MARM can provides not only ultra-high MR ratio but also ultra-low RA for high-speed and low power writing.