QuInAs Technology, developer of the compound-semiconductor memory technology Ultraram, announced the publication of its latest research in the Journal of Applied Physics.The work links device-level physics – including resonant tunnelling and floating-gate dynamics – directly to AI system performance through compact modelling and hardware-aware benchmarking, addressing a key limitation in how emerging memory technologies are typically evaluated.

The company will also present this work at the International Symposium on Quality Electronic Design (ISQED) 2026, taking place in San Francisco from 8–10 April.

Click to enlarge

ULTRARAM 4×4 array for in-memory computing

The presentation, scheduled for Friday 10 April at 9:20am (Session 5A), will focus on system integration and design considerations, bringing Ultraram into the electronic design automation (EDA) and system design community.

The paper, titled “Artificial synapse based on Ultraram memory device for neuromorphic applications”, demonstrates how Ultraram can be modelled and evaluated as a synaptic memory element for next-generation AI hardware.

Developed in collaboration with IIT Roorkee and Lancaster University, the work introduces a physics-based compact modelling framework that links device-level behaviour – including resonant tunnelling and floating-gate charge dynamics – to circuit- and system-level performance.

This enables, for the first time, hardware-aware evaluation of Ultraram in neuromorphic and in-memory computing architectures, using crossbar array simulations and DNN+NeuroSim benchmarking on tasks such as CIFAR-10 classification.

“Much of today’s AI hardware research evaluates memory technologies under idealised assumptions,” said James Ashforth-Pook, CEO, QuInAs Technology. “This work takes a different approach – connecting real device physics directly to system-level performance. That’s essential if we are to build practical, energy-efficient AI systems.”

The research shows that Ultraram can achieve competitive accuracy while offering advantages in energy efficiency and area compared to conventional SRAM-based approaches, highlighting its potential as a platform for future AI hardware.

“By integrating physics-based modelling with system-level benchmarking, we can better understand how emerging memory technologies behave in real AI workloads, rather than relying on idealised models,” added Abhishek Kumar, lead author.

Ultraram is based on III–V compound semiconductor heterostructures and leverages resonant tunnelling to enable ultra-low energy switching and long data retention, positioning it as a candidate for neuromorphic and in-memory computing applications.

The paper is available at: https://doi.org/10.1063/5.0314826.