Operating AI chips at the edge places multiple strains on memory systems, to cope with the signature high-throughput, quasi-continuous, parallel processing. Whilst off-chip memories induce latency, power consumption, and cost, embedded SRAM is showing its limits. SOT is the only embedded memory technology that combines all the requirements for feature maps operation, from the mandatory infinite endurance to high read and write speed and zero standby power, all with a moderate added processing cost.
With neural networks today stuck in the cloud to cope with the ever-increasing power hunger of AI systems, SOT creates opportunities for chip vendors to develop edge AI chips with improved performances and energy efficiency.