SpinEdge AI accelerator

Technology

Our world is analog - it generates analog data like colors, sounds and radio waves. At SpinEdge, we naturally process this data in an analog way, avoiding its costly digitalization.

SpinEdge technology is based on the array of non-volatile spintronic memory cells, which stores matrices of neural network weights. The cells array acts not like digital memory storing digits - it's rather a network of resistive elements, which adds and multiplies electric currents.

​

The analog data processing occurs according to the Kirchhoff’s law when the input data is fed into the crossbar in the form of voltages, then converted to currents and propagates through the cells, the output is instantaneously read out as electric currents. Actually, this principle is already applied in the best processor ever built – the human brain. That's why this technology is called "neuromorphic".

​

SpinEdge AI accelerator enables virtually instantaneous inference of pre-trained neural networks – at the speed of data input.

The neural network weights are permanently stored in the SpinEdge memory cells and are instantly available for the new incoming data. This makes SpinEdge a perfect accelerator for AI on the Edge, where instant, event-driven boot-ups are needed amidst general energy-saving hibernation of an Edge device. This feature also enables SpinEdge to retain the data after sudden power-downs, which often happen at the Edge. 

The SpinEdge approach offers high processing speed and efficient TOPS/W.

​

While a pre-trained neural network is permanently stored in the SpinEdge crossbars, it's possible to remotely update or re-purpose a SpinEdge accelerator functions by re-writing the neural network data.

​

SpinEdge hardware innovation – focus on energy efficiency and low latency

Our patented memory array configuration and custom-optimized spintronic cell, allow to produce large memory computational arrays for higher energy savings, computing speed and versatility. The SpinEdge analog architecture requires no energy-expensive data digitalization and processor <–> memory data exchange, as the major processing operations are done in memory.

​

Up to 1000X gain in energy efficiency and up to 100X performance gain are achievable compared to leading CPU architectures.

​

SpinEdge spintronic memory cells are faster, smaller, more durable and energy-efficient - the next step in neuromorphic computing, compared to conventional NAND Flash or SRAM cells.