By Maria Deutscher 2020-03-31
Chip startup Perceive Corp. today emerged from stealth mode to introduce Ergo, an artificial intelligence processor for edge devices that it says is 20 to 100 times more power-efficient than competing products in the category.
San Jose, California-based Perceive was incubated for two years at publicly traded semiconductor firm Xperi Corp., which is also the startup’s majority owner. Xperi sells chip technologies under license to customers such as Samsung Electronics Co. Ltd.
Percieve’s Ergo chip (pictured) is a seven-by-seven-millimeter die designed for use in edge devices with power and space constraints. The startup is initially targeting consumer products such as smartphones, Chief Executive Officer Steve Teig told SiliconANGLE. The chip’s compact size also allows it to run AI models inside smaller devices, including home security cameras, smart appliances and wearables.
Perceive claims Ergo provides sustained performance of more than 4 trillion operations per second, or TOPS, for AI applications. That’s enough horsepower to run multiple neural networks simultaneously or one neural network at a trot. Perceive claims Ergo can run YOLOv3, a popular object recognition model developed at the University of Washington, at up to 246 frames per second, four times the refresh rate of the display in Samsung’s flagship Galaxy S10 phone.
Teig emphasized that the 4 trillion-plus operations Ergo performs a second are done with floating-point values rather than lighter integer values.
“TOPS is an often-quoted number for inference processors,” Teig explained. “The TOPS number can often refer to INT8 (8-bit integer) TOPS, or FLOAT (floating point) TOPS, et cetera.”
Wiith Ergo, he said, “we are highlighting that this number is sustained versus peak TOPS, is floating point and not integer TOPS, and is calculated based on the amount of GPU-equivalent work Ergo can perform, such as frames per second of a popular network like YOLOv3, and not some theoretical calculation of operating frequency times computing units.”
Ergo can run YOLOv3 using as little as 20 milliwatts of power in certain scenarios. The chip’s performance-to-electricity ratio works out to 55 TOPS per watt, according to Perceive, which the startup says is 20 to 100 times better than other edge AI processors.
A more power-efficient chip theoretically allows the device in which it’s running to last longer on a single battery charge. Ergo’s lower energy footprint also means it gives off less heat, which simplifies system design for hardware makers.
Perceive is currently sampling Ergo to customers and plans to begin mass production in the second quarter. The startup is fabricating the chip on Globalfoundries Inc.’s 22FDX process, the same process Google LLC supplier Synaptics Inc. uses to make silicon for the latest Nest Mini.
The edge AI market holds big opportunities for chip makers since there are billions of smartphones, smart speakers and other connected devices that could potentially benefit from having an onboard machine learning processor. Perceive is far from being the only player eyeing a piece of the pie. Other startups such as Hailo Technologies Ltd. have introduced tiny AI processors of their own, while Google offers the penny-size Edge TPU as part of its Coral product family.