- The energy demands could be reduced by large single variety chips
- Researchers say they can exceed the limitations faced by GPUs
- Brains and Tesla already use these huge chips, with special cooling systems to manage heat
The engineers of the University of California Riverside are exploring a new approach for artificial intelligence hardware that could address performance and sustainability.
In an article reviewed by pairs, published in the magazine DeviceThe team investigated the potential of accelerators at wafering scale: giant computer chips operating with whole silicon wafers instead of the small chips used in today’s GPUs.
“The wafering scale technology represents a great advance forward,” said Mihri Ozkan, a professor of Electrical and Informatics Engineering at UCR and main author of the document. “It allows the AI models with billions of parameters to function faster and more efficiently than traditional systems.”
Like Monorraíla
These chips, such as the headache engine of the brain wafer 3 (Wse-3), which we have previously covered, contain up to 4 trillions of transistors and 900,000 nuclei focused on AI in a single unit. Another processor at the scale of wafers, Dojo D1 of Tesla, houses 1.25 billion transistors and about 9,000 nuclei per module.
Processors eliminate common energy delays and losses in systems where data between multiple chips travels.
“By keeping everything in a wafer, avoids delays and energy losses from Chip’s communication to Chip,” Ozkan said.
Traditional GPUs remain important due to their lower cost and modularity, but as the models of AI grow in size and complexity, chips begin to find the performance and energy barriers.
“The computing of AI is no longer just speed,” Ozkan explained. “It’s about designing systems that can move massive amounts of data without overheating or consuming excessive electricity.”
Scale systems also have important environmental benefits. Wse-3 brains, for example, can perform up to 125 quadrillones of operations per second, while using much less energy than GPU settings.
“Think of the GPUs as occupied, effective roads, but the jams waste energy,” Ozkan said. “Wofleter scale engines are more like Monorraíl: direct, efficient and less polluting.”
However, there is still a great challenge: the old heat problem. Oscalo’s scale chips can consume up to 10,000 watts of power, almost everyone becomes heat, which requires advanced cooling systems to avoid overheating and maintain performance.
The brains use a glycol cooling loop integrated into the chip, while Tesla has a liquid system that extends a refrigerant uniformly on the chip surface.
Through Tech xplore
