Everspin unveils UNISYST MRAM for unified edge AI memory

Everspin has introduced a new MRAM product family called UNISYST, positioned as a unified memory for embedded systems that combines code storage and data memory in a single non-volatile architecture.

The announcement comes as system designers hit limits with NOR flash, particularly as software images and embedded AI models grow. In many systems, flash also constrains write endurance and update speed, affecting development cycles and field updates.

Unified memory

UNISYST extends Everspin's magnetoresistive RAM into roles typically handled by NOR flash and higher-density persistent storage. Everspin describes it as a "code-and-data MRAM" architecture that fits between configuration memory and bulk non-volatile storage in embedded designs.

"System designers are running into the physical and performance limits of NOR flash, especially as process nodes move below 40 nanometers and workloads become more demanding," said Sanjeev Aggarwal, President and CEO of Everspin Technologies. "With UNISYST, we are extending our MRAM roadmap to higher densities while giving customers a practical way to start with PERSYST today and migrate to a code-and-data MRAM architecture as soon as it is available."

UNISYST builds on Everspin's existing PERSYST MRAM platform and is intended to provide a migration path from current serial MRAM devices to higher-density unified memory without changes to system architecture or software.

Device details

The initial UNISYST range is planned with densities from 128 megabits to 2 gigabits, using an xSPI interface up to octal SPI at 200MHz.

For demanding environments, the devices are planned to include AEC-Q100 Grade 1 qualification and a minimum 10-year data retention specification at extreme temperature.

Planned performance includes read bandwidth of up to 400 MB/s and write bandwidth of about 90 MB/s. Everspin also says write speed will be more than 400 times faster than NOR flash, with write endurance up to 10 times higher than typical NOR.

Edge AI focus

Everspin tied the launch to the growth of larger AI workloads outside the data centre, where storage and update cycles can become bottlenecks. UNISYST is aimed at edge AI, industrial systems, and mission-critical designs that need non-volatile memory with predictable behaviour over temperature and time.

OpenMV, which works on embedded machine vision, highlighted update speed as a practical issue when models move from cloud training to devices in the field.

"As generative AI models move from the cloud to embedded systems, we're suddenly dealing with assets that are tens or even hundreds of megabytes in size," said Kwabena W. Agyeman, President and Co-founder of OpenMV. "Storing those models is only part of the challenge - updating them quickly during development and deployment is equally important. High-speed, non-volatile Everspin UNISYST MRAM changes what's practical for edge AI systems by removing the write bottlenecks associated with traditional flash."

Everspin says the architecture supports fast boot and rapid updates, with more predictable performance than flash-based designs that can vary based on write and erase behaviour. The company also positions UNISYST for software-defined systems that undergo frequent reconfiguration and require data integrity across power cycles.

Target markets

Alongside edge AI, Everspin lists military and aerospace as target areas, including FPGA configuration and code storage in mission-critical systems such as low-Earth orbit satellites. It also points to frequent over-the-air updates as a common operational requirement.

Automotive is another intended segment, where the planned Grade 1 qualification and long-term data retention are positioned as a fit for control systems, logging, and configuration memory. Everspin also cites industrial and casino gaming for high-traffic logging and configuration workloads that demand fast writes and high endurance.

Everspin has supplied MRAM for mission-critical storage applications for close to two decades. With UNISYST, it describes the launch as expanding its portfolio into a broader market for non-volatile memory in embedded designs.

Engineering samples of UNISYST are expected in the fourth quarter of 2026, with additional densities and options planned after that.