Scaling the Next Generation AnalogML™: The AML200


Aspinity next generation analogML coreAnalogML™ is moving to 22nm for the next generation processor. And while there is agreement that there are significant power and performance benefits to using analog circuitry, there are many who would say that its functionality is limited because it can’t be successfully scaled to this more cutting-edge technology node.  This common view stems from the fact that analog performance relies on geometry, and when geometries shrink, even small deltas in line widths across a wafer can cause a big performance mismatch from one device to another. Some companies prevent small variation errors from compounding across operations by limiting the functions performed by analog circuitry.

So how can Aspinity take advantage of scaling when the chip’s processing is performed entirely within analog circuitry?

Our last blog post discussed the importance of high-precision analog parameters being stored at the location of the compute for the successful deployment of Aspinity’s all-analog machine learning technology. In that article we discussed that analogML™ technology embeds analog weights for the analog neural network, as well as analog biases and activations for other circuits within the computational circuits themselves. Most importantly for scaling to smaller geometries, the analog values are stored at the high precision needed to trim out any variations in the analog circuit performance that arise from manufacturing or the environment.

We are taking advantage of our proprietary architecture to scale our next generation analogML core, the AML200, to leverage the efficiency and RF support of 22nm technology. The AML200 supports up to 125,000 parameters and a wide range of model types for efficient complex machine learning applications such as RF and image processing, enabling a typical system power consumption <100μW for battery-operated edge AI.

The AML200 is the only edge AI processor to deliver TOPS level performance in a battery-operated power footprint.

Aspinity has fundamentally achieved the best of both worlds –  leveraging the high performance and near-zero always-on system power that can be achieved with an all-analog processing technology, while also being scalable and flexible to address more applications with higher performance and efficiency.