Packing more power into less space: TI on power modules
Weebit Nano on ReRAM and in-memory compute for Edge AI

Weebit Nano on ReRAM and in-memory compute for Edge AI

Weebit Nano on ReRAM and in-memory compute for Edge AI Weebit Nano on ReRAM and in-memory compute for Edge AI

At electronica 2026 in Munich, Electronic Specifier’s Consulting Editor Mark Gradwell spoke with Eran Briman, VP of Marketing and Business Development at Weebit Nano, about why resistive RAM could transform AI at the Edge.

By the third day of electronica, Briman reported plenty of traffic and “a lot of interest” in Weebit Nano’s booth. Much of that interest centred on the company’s newly announced development kit for in-memory compute, built on its resistive RAM (ReRAM) technology.

The real bottleneck: moving data

In AI inference, Briman explained, the bottleneck is often not processing at all. “It’s actually moving data in and out from memory onto these AI accelerators,” he said. “In some cases, we see AI accelerators actually sitting idle waiting for this data to come in for about 50% of the time.”

That suggests the industry may be “focusing on the wrong thing”. The aim should be to minimise, or even eliminate, data movement, and that is where in-memory compute comes in. “In-memory compute basically means you compute where the data is stored,” he said. The memory cell not only stores the data but also runs the computation in place.

A development kit for researchers

For several years, researchers around the world have asked to get their hands on Weebit’s ReRAM. Over the past year, the company has built a research development kit, Weebit Matrix, that lets them use its hardware to develop their own algorithms, AI systems, interfaces, and neuromorphic applications.

The kit was announced the week before electronica, but five customers were already using the platform ahead of the official launch, and one has built a working demonstrator.

Why ReRAM is “a perfect fit”

In-memory compute has been researched for years using various technologies. ReRAM stands out, Briman said, because of the physics it relies on: each memory cell is “basically a resistor that can change its resistance”.

AI inference spends around 99% of its time on vector-matrix multiplications, which break down into vast numbers of multiply-accumulate operations. Apply a voltage to a resistor and the resulting current is the product of that voltage and the cell’s conductance. With the AI weights stored as conductances in the ReRAM array, “the multiplication is done automatically without any clock cycles, without moving the data anywhere”.

Summing the currents along a column completes the accumulate step. Arrange the cells in a crossbar array, and a matrix programmed once into the cells, and reprogrammed if needed, will turn an input voltage vector into an output vector “instantaneously”.

One early-access customer has built a demonstrator, shown on Weebit’s booth, that runs image processing functions and compares an NVIDIA GPU with Weebit’s in-memory compute platform. “It’s about three orders of magnitude lower power consumption running on our in-memory compute compared to a GPU,” Briman said.

He attributes that to two factors. First, the GPU constantly moves coefficients and data in and out of memory, which in-memory compute eliminates. Second, results are available immediately, with no clocking on the chip, unlike a GPU running at hundreds of megahertz. “Altogether more than 1,000x – this is quite significant,” he said, adding that Weebit expects more researchers to produce prototypes and demonstrations.

Embedded NVM remains the core

For all the excitement around in-memory compute, Briman was clear that Weebit’s main focus “was and still remains for the foreseeable future” ReRAM as embedded non-volatile memory, replacing embedded flash.

Embedded flash, he said, is “running into a wall” below 28nm on power consumption and cost, requiring seven, 10, or sometimes 15 additional masks compared with just two for Weebit’s ReRAM.

The development kit is “just a first step” for in-memory compute. Within two to three years, Weebit hopes to offer a commercial IP solution for in-memory compute alongside its embedded NVM platform.

Keep Up to Date with the Most Important News

By pressing the Subscribe button, you confirm that you have read and are agreeing to our Privacy Policy and Terms of Use
Previous Post
Packing more power into less space: TI on power modules

Packing more power into less space: TI on power modules