In-Memory Computing has emerged as one of the most promising techniques to realize next-generation low-power AI processing. Researchers have tested In-Memory Computing technology using most available types of non-volatile memory. However, the potentially most complementary memory tech for this purpose, MRAM (Magnetoresistive Random Access Memory), hasn't been viable due to its low resistance. Samsung's researchers now say they have successfully demonstrated MRAM for In-Memory Computing (opens in new tab) using a novel 64 × 64 crossbar array based on MRAM cells.
According to the paper 'A crossbar array of magnetoresistive memory devices for in-memory computing,' published by Nature (opens in new tab) on Wednesday, the key architectural innovation by the Samsung research team was to test a 64 × 64 crossbar array based on MRAM cells "that overcomes the low-resistance issue with an architecture that uses resistance summation for analogue multiply–accumulate operations." Demonstrations of Samsung's novel technology used this 64 x 64 array integrated with readout electronics in 28nm CMOS technology.
The teams from Samsung Advanced Institute of Technology (SAIT) in close collaboration with Samsung Electronics Foundry Business and Semiconductor R&D Center, performed tests with the new MRAM solution, and checked its AI computing performance. In tests, the new device worked well in AI computing, achieving an accuracy of 98% in classification of hand-written digits and a 93% accuracy in detecting faces from scenes.
MRAM has some very attractive qualities for In-Memory Computing, namely its non-volatile nature, its operational speed, endurance, and economy of large-scale production. However, the Samsung scientists point out that it was impractical to use MRAM previously, due to its low resistance being a mismatch for the architecture. It must be highlighted that MRAM technology is by no means new, and Tom's Hardware has discussed its features and applications several times over the last 15 years, with practical MRAM development starting in earnest back in 1998.
Not only can this new MRAM technology application be used for in-memory computing, but scientists also believe it can serve as a platform to download biological neuronal networks. The Samsung researchers comment that "In-memory computing draws similarity to the brain in the sense that in the brain, computing also occurs within the network of biological memories, or synapses, the points where neurons touch one another."
Samsung plans to move forward, building on this MRAM in-memory computing success, to cement its leadership in memory technology and processor development. It is hard to say how long it will be before this type of technology drips down into consumer devices like PCs, tablets and smartphones.