Page 19 - EE Times Europe Magazine – November 2023

EE|Times EUROPE 19

Could IBM’s AI Chip Reinvent Deep Learning Inference?

is a type of nonvolatile random-access mem- terized errors in a lot of detail. Our precision NEXT STEPS
ory that switches between a low-conductive is sufficient for neural networks.” Since publishing their results in Nature
amorphous phase and a high-conductive Intel Labs research scientist Hechen Wang Electronics, the IBM Europe researchers have
crystalline phase when heated. Devices based has also been working with analog in-memory made clear their confidence that their work
on these materials can harness the phase computing for many years, and he concurs proves analog AI can deliver the necessary
change, encoding the changes in conduc- that the approach can achieve exceptional compute precision to rival conventional
tance as synaptic weights that are then used energy efficiency. “Researchers started to look digital accelerators, but with far greater
to compute operations. Critically, recording at analog in-memory around five years ago, energy efficiency. With the rise of AI-based
this continuum of values—rather than just and now we have IBM, imec, GlobalFoundries, technologies set to make energy-efficient
the 1s or 0s of digital devices—works out well TSMC, Samsung and other companies and and accurate inference hardware essential,
for deep neural network operations, as IBM’s academic groups starting to research [the the researchers’ aim is to create analog
latest results indicate. technology],” he said. “If we want to do very in-memory chips that can execute end-to-
When benchmarked against other chips efficient [AI] computing, we need to put the end inference operations.
based on similar technology, including processing unit inside the memory array, or In the meantime, IBM Research Europe
NeuRRAM and those developed by Mythic and even the memory cells.” told EE Times Europe that it intends to take
TSMC, IBM’s tech could perform matrix- Intel Labs is pursuing several avenues for advantage of the high synaptic densities
vector multiplications—fundamental to AI in-memory computing and exploring a host of that can be reached on PCM devices and
operations—at least 15× faster with a compa- memory technologies, Wang said. “We haven’t build bigger chips that can run entire net-
rable energy efficiency. Notably, the chip also yet drawn a conclusion as to which memory is work operations to more than rival digital
proved to be more accurate than the other the right direction to take.” accelerators.
chips at image recognition when tested using Wang nonetheless believes the latest “Once we’ve really shown the promise of
AI-training color image database CIFAR-10, analog in-memory developments from IBM this technology and more people want to
challenging the notion that analog and elsewhere are having a positive impact invest [in the field], then we could have teams
in-memory computing is energy-efficient but on what has been a “heated” field. “IBM’s of hundreds of researchers working on this so
can be prone to calculation errors. research has been published in Nature papers, we can get the chip into production mode,”
“The strength of phase-change memory is and to be honest, I never dreamed this would the IBM scientist said. “So for now, we’re
that it’s sufficiently stable to do some rela- happen,” he said. “Many people read these going to continue working at this.” ■
tively accurate computations,” said the IBM publications, and I hope [these results] will
scientist. “We have developed techniques for open their minds and attract even more Rebecca Pool is a contributing writer for
accurate programming and have also charac- attention to the field.” EE Times Europe.

We help fill in the outlines

to make your idea real.

When you’re an innovator dreaming on the edge of
tomorrow, Arrow’s engineering services can help you
take your idea from imagination to innovation.
See how at arrow.com/fiveyearsout