The chip requires custom software training but achieves 55 TOPS/W inferencing, 20x to 100x higher than the efficiency of many alternatives. This begs the question what in the case of the Ergo chip is an operation?
Perceive has not provided detail of the interior architecture of the chip but states it achieves 4TOPS sustained and can run multiple large neural networks simultaneously for such applications as video object detection, audio event detection and speech recognition. This would make the chip a good fit for security cameras. Similarly Perceive has not indicated what data types are supported or which ones are supported efficiently.
Perceive has provided one benchmark result claiming that Ergo can run YOLOv3 at up to 246 frames per second while consuming about 20mW.
The chip is manufactured for Perceive in the 22nm FDSOI manufacturing process 22FDX by Globalfoundries and is capable of processing large neural networks in 20mW. It supports a variety of advanced neural networks including convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory NNs (LSTMs), and more.
Ergo requires no external RAM and its 7mm by 7mm package makes it suitable for use in consumer electronics such as phones and cameras. The chip is shipping to consumer equipment makers and is aimed at applications that include: security cameras, smart appliances, mobile phones, action cameras and wearables.
Along with the chip Perceive provides reference boards and standard imaging and audio inferencing applications. Customer can tune the applications or create novel applications with support from Perceive.
Karl Freund, an analyst with Moor Insights, has observed that the Ergo chip is NOT compatible with standard Tensorflow training output. The need to let Perceive use its tool chain to train the network – using large computer resources based on GPUs – is the trade off for the claimed increase in performance.
Next: Unsubstantiated claims of novelty