Perceive Reduces AI to a Math Problem
By Mike Demler 2020-05-12
Two-year-old startup Perceive has shattered power-efficiency benchmarks for edge-AI processors. Its first chip, Ergo, needs just 70mW to deliver 250fps on Yolo v3, all without connecting to DRAM. On the slower 30fps video streams customers will typically employ, it needs 20mW to process full-frame images. Unlike other inference engines, Ergo has no large on-chip SRAM, reducing storage requirements to “a few megabytes,” according to the startup.
Perceive says it achieved this incredible performance by developing a new method that eliminates the multiply-accumulate (MAC) arrays common to such devices, replacing them with more-efficient (but undisclosed) math operations. To match Ergo’s performance, a standard MAC-based inference engine would need to sustain four trillion operations per second (TOPS) at a power efficiency of 55 TOPS per watt.
By sipping just tens of milliwatts, Ergo is well suited to a wide range of battery-powered devices, including cameras, drones, and wearables. Its ability to simultaneously recognize objects and sounds makes it an excellent fit for security systems. ADASs require a more powerful application processor, but including the chip in automotive cameras allows processing to occur right next to the sensor, reducing bandwidth requirements for in-vehicle networks. By combining its proprietary hardware and software, Perceive is making low-power edge devices a lot smarter, as well as more secure.