Page 39 - september22_082722.indd
P. 39
TAP TIMES -39- SEPTEMBER 2022
New chip (from 38) on different tasks. They found that it’s
99% accurate in letter recognition from
The memory technology that NeuRRAM the MNIST dataset, 85.7% accurate on
used is resistive random-access memory image classification from the CIFAR-10
(RRAM). dataset, 84.7% accurate on Google speech
command recognition and showed a 70%
Non-volatile memory reduction in image-reconstruction error on
It is a type of non-volatile memory— a Bayesian image recovery task.
memory that retains data even once pow-
er is off — that has emerged in commer- From hardware to software
cial products. “Efficiency, versatility, and accuracy are all
important aspects for broader adoption of
RRAM can store large AI models in a small the technology,” said Wan. “But to realize
area footprint, and consume very little them all at once is not simple. Co-optimiz-
power, making them perfect for small-size ing the full stack from hardware to soft-
and low-power edge devices. ware is the key.”
Even though the concept of CIM chips is Right now, NeuRRAM is a physical proof-
well established, and the idea of imple- of-concept but needs more development
menting AI computing in RRAM isn’t before it’s ready to be translated into ac-
new, “this is one of the first instances tual edge devices.
to integrate a lot of memory right onto
the neural network chip and present all But this combined efficiency, accuracy, and
benchmark results through hardware mea- ability to do different tasks showcases the
surements,” said Wong, who is a co-senior chip’s potential.
author of the Nature paper.
Simple to complex tasks
The architecture of NeuRRAM allows the “Maybe today it is used to do simple AI
chip to perform analog in-memory compu- tasks such as keyword spotting or human
tation at low power and in a compact-area detection, but tomorrow it could enable a
footprint. whole different user experience. Imagine
real-time video analytics combined with
Designed with UC San Diego speech recognition all within a tiny device,”
It was designed in collaboration with the said Wan.
lab of Gert Cauwenberghs at the Universi-
ty of California, San Diego, who pioneered “To realize this, we need to continue im-
low-power neuromorphic hardware design. proving the design and scaling RRAM to
more advanced technology nodes.”
The architecture also enables reconfigu-
rability in dataflow directions, supports “This work opens up several avenues of
various AI workload mapping strategies, future research on RRAM device engineer-
and can work with different kinds of AI ing, and programming models and neural
algorithms – all without sacrificing AI com- network design for compute-in-memory, to
putation accuracy. make this technology scalable and usable
by software developers”, said Priyanka
To show the accuracy of NeuRRAM’s AI Raina, a co-author and assistant professor.
abilities, the team tested how it functioned stanford.edu
