Page 38 - september22_082722.indd
P. 38
TAP TIMES -38- SEPTEMBER 2022
New chip ramps-up the memory itself, thereby eliminating
AI computing efficiency the separation between the compute and
memory units.
STANFORD, Calif. — AI-powered edge
computing is already pervasive in our “Compute-in-Memory” chip
lives. Their “compute-in-memory” (CIM) chip,
called NeuRRAM, is about the size of a
Devices like drones, smart wearables, and fingertip and does more work with limited
industrial IoT sensors are equipped with battery power than what current chips can
AI-enabled chips so that computing can do.
occur at the “edge” of the internet, where
the data originates. This allows real-time “Having those calculations
processing and guarantees data privacy. done on the chip instead of
sending information to and
from the cloud could enable
faster, more secure, cheap-
er, and more scalable AI
going into the future, and give more peo-
ple access to AI power,” said H.-S Philip
Wong (above), the Willard R. and Inez
Kerr Bell Professor in the School of Engi-
neering.
The data movement issue
“The data movement issue is similar to
The NeuRRAM chips is twice as energy efficient as the state-of- spending eight hours in commute for a
the-art. David Baillot, University of California, San Diego) two-hour workday,” added Weier Wan, a
recent graduate at Stanford leading this
Limited functionality project. “With our chip, we are showing a
However, AI functionalities on these tiny technology to tackle this challenge.”
edge devices are limited by the energy pro-
vided by a battery. Therefore, improving They presented NeuRRAM in a recent ar-
energy efficiency is crucial. ticle in Nature. While compute-in-memory
has been around for decades, this chip is
In today’s AI chips, data processing and the first to actually demonstrate a broad
data storage happen at separate places – a range of AI applications on hardware,
compute unit and a memory unit. The fre- rather than through simulation alone.
quent data movement between these units
consumes most of the energy during AI pro- To overcome the data movement bottle-
cessing, so reducing the data movement is neck, researchers implemented what is
the key to addressing the energy issue. known as compute-in-memory (CIM), a
novel chip architecture that performs AI
Stanford University engineers have come computing directly within memory rather
up with a potential solution: a novel resis- than in separate computing units.
tive random-access memory (RRAM) chip
that does the AI processing within next page
