Most AI accelerators currently rely on CNN (convolutional neural network) to perform AI inference in a much faster and efficient way than on CPU cores, or even GPUs. But there’s another type of neural network, namely spiking neural networks (SNN) that uses the timing of spikes in an electrical signal to perform pattern recognition tasks in a way similar to neurons in the brain.
The claims in terms of efficiency are quite unbelievable, with up to 10,000 times more performance per watts than in microprocessor and digital accelerometer, 500 times lower energy, and 100 times shorter latency.
Several companies are working on neuromorphic AI accelerators for spiking neural networks, with notably Prophesee focusing on image processing, and Innatera that is working on an ultra-low-power AI accelerator handling audio, health, and radar for sound and speech recognition, vital signs monitoring, elderly person fall sensors, etc…
Innatera recently provided additional information about their solution, so let’s focus on that one. Talking to EETimes, Marco Jacobs, Innatera VP marketing and business development, explains the sensors have time series data, instead of images which are very parallel, and the processing in the three applications targeted by the accelerator typically happens in the sensor node, which may be battery powered.
In Innatera’s audio tests, each spike event (each neuron firing in response to input data) required less than 200 femtoJoules in TSMC 28nm process, which is getting closer to the amount of energy used by biological neurons and synapses. Since a typical audio keyword spotting application requires under 500 spike events per inference, sub-milliWatt inference is possible, and Innatera AI accelerator enables always-on pattern recognition capabilities directly in battery-powered sensor nodes.
Innatera neuromorphic AI accelerator is expected to become available by the end of 2021.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.