As the demand for AI computing surges, the challenge for the semiconductor industry is shifting toward processing inference with less power rather than higher performance. Amid this trend, the U.S. AI semiconductor company Mythic is accelerating the development of low-power AI semiconductors by incorporating Silicon Storage Technology (SST)'s memBrain technology into its next-generation Analog Processing Units (APUs).

Microchip Technology's subsidiary SST announced on March 18 that Mythic has adopted its memBrain neuromorphic hardware IP. The two companies plan to implement an analog in-memory computing (aCIM) architecture using SST's SuperFlash embedded non-volatile memory bitcells, and thereby develop an AI inference accelerator applicable from edge devices to enterprise and data center environments.

The key is a structure that performs operations within memory. The memBrain cell can store up to 8 bits of data per bit cell and supports low read currents at the single-digit nanoampere level. In addition, it is designed to meet the requirements for low-power AI computation by featuring 10 years of data retention, 100,000-cycle repeatability, multi-state write control, and single-cycle multiplication-accumulation (MAC) operation capabilities.

The Mythic APU, to which this technology will be applied, aims for inference performance of 120 TOPs per watt. The company explained that it is aiming for up to 100 times higher energy efficiency compared to existing digital GPUs. The plan is to expand the scope of application to include industrial equipment with severe power constraints, automotive electronics, and AI sensor fusion systems, as well as data centers requiring large-scale computation.

SuperFlash, the underlying technology, is an embedded non-volatile memory solution that has already been widely used in the industrial, automotive, consumer, and computing sectors. To date, shipments of products based on this technology have exceeded 150 billion units, and it is licensed to all of the world's top 10 semiconductor foundries. Based on this foundation, SST has been developing memBrain for 40nm and 28nm processes and plans to expand to the 22nm process in the future.

This collaboration is interpreted as an example demonstrating the shift in AI semiconductor competition from a performance-centric approach to a power-efficiency-centric one. In particular, the key factor is expected to be how competitive in-memory computing—which reduces data movement between computation and storage—will be in real-world applications.