Page 13 - EE Times Europe Magazine | June2020

EE|Times EUROPE 13

Startup Reinvents Neural Network Math, Launches 20-mW Edge AI Chip

carried with the signal, making mainstream neural networks very large and
making them susceptible to adversarial examples and other tricks. Perceive’s technology is
“The more you can be mathematical about figuring out which parts based on reinventing neural
need to be kept and which parts are just noise, the better job you can network math using
do at generalization and the less other overhead you have to carry techniques from information
with you,” Teig said. “I would claim even current neural networks are theory. (Image: Perceive)
extracting signal from noise; they’re just not doing it in as rigorous a
way, and as a result, they’re carrying extra weight with them.”
This information-theoretic point of view is the basis for Perceive’s plus the machine-learning technology that’s able to find this represen-
machine-learning strategy, which represents neural networks in a tation of the networks and to train the networks in a way that makes
new way. “Really, this is a marriage between an information-theoretic them compatible with what the chip wants to see.”
perspective on how to do machine learning and a chip that embodies
those ideas,” Teig said. IMAGE AND AUDIO
Ergo can support two cameras and includes an image-processing unit
CHIP ARCHITECTURE that works as a pre-processor, handling things like de-warping fisheye
With Teig’s background as CTO of Tabula, you might expect hardware lens pictures, gamma correction, white balancing, and cropping.
based on programmable logic, but that’s not the case here. “It’s not fancy, but the pre-processing that’s obviously useful to do
“I’ve been strongly influenced by thinking about programmable logic in hardware, we do in hardware,” Teig said. “And we have the audio
for a decade and how to build rich interconnect architectures to enable equivalent, too — we can take multiple stereo microphones and do
high-performance, very parallel computation, because much of what beamforming, for example.”
happens on an FPGA is also massively parallel and very intensive in its There is also a Synopsys ARC microprocessor with a DSP block that
interaction between computation and memory,” he said. “That work can also be used for pre-processing, plus a security block, also from
has definitely influenced my work at Perceive, but what we have is not Synopsys.
programmable logic per se. It’s been influenced by that way of thinking, “One of the things we’ve done is to encrypt absolutely everything in
but the architecture itself is around neural networks.” order to maintain a level of security in an IoT setting,” Teig said. “We
Perceive’s neural network fabric is scalable, with initial chip Ergo hav- encrypt the networks, encrypt the code that runs on the microproces-
ing four compute clusters, each with its own memory. While exact details sor, encrypt the interfaces, encrypt everything.”
are still under wraps, Teig did say that these clusters are significantly The chip features appropriate I/Os for sensors outside image and audio,
different from anything found in other AI accelerators, which typically and it supports an external flash memory and/or microprocessor, which
use arrays of MAC units to compute dot products of vectors and matrices. enables over-the-air updates. This could be used to update the neural
“We are not doing that,” Teig said. “We do not have an array of MACs. networks loaded on the chip or load different networks as required.
As a result … we are 20× to 100× as power-efficient as anything else on Ergo is sampling now along with an accompanying reference board.
the market. The reason for that is that everybody else is doing the same Mass production is expected in Q2 2020. ■
thing and we’re not. Our representation of the networks is quite new,
and that’s what’s allowed us to achieve such great efficiency — that, Sally Ward-Foxton is a staff correspondent at AspenCore.

www.eetimes.eu | JUNE 2020