Analog In-memory computing with multilevel memristive devices for high performance computing

Описание: Glenn Ge, TBD - TetraMem Inc.

The von Neumann architecture's intrinsic bottleneck in data transfer between processor and memory units hinders performance as data sets continue to grow. TetraMem's memristive devices-based analog in-memory computing significantly boosts throughput and energy efficiency in deep learning. Our approach utilizes pre-trained synaptic weights from cloud-based training, directly programming them into multi-level memristors/RRAMs for edge deployment and enabling post-tuning to accommodate specific scenarios. High-precision programmability ensures uniform performance across memristive networks by necessitating numerous distinguishable conductance levels in each device. This advancement benefits applications like neural network training and inference computing. By achieving stable 8 bits and above multi-levels conductance in individual memristor devices (up to 11 bits/cell, as featured in main journal publication, Mar 2023), we enable monolithically integrated semiconductor chips, featuring large crossbar arrays on complementary metal-oxide-semiconductor (CMOS) circuits in the commercial foundry, suitable for diverse applications including high performance computing (HPC). Our arbitrary precision computing based on analog computing work is now in press with main journal and available to public very soon.