- AI energy requests could be lowered by large single Wafer chips
- The researchers say that these can overcome the limitations encountered by the GPUs
- Cerebras and Tesla already use these huge chips, with special cooling systems to manage heat
Engineers at the University of California Riverside are exploring a new approach to artificial intelligence equipment that could approach performance and sustainability.
In an article evaluated by peers, published in the journal DeviceThe team has studied the potential of accelerators on the scale of the verses – giant computer fleas that operate on whole silicon slices rather than on the small chips used in today’s GPUs.
“Technology at the scale of payments represents a major jump forward,” said Mihri Ozkan, professor of electrical and computer engineering at the UCR and the main author of the newspaper. “It allows AI models with billions of parameters to operate more quickly and more efficiently than traditional systems.”
Like monorails
These chips, such as the engine at the Cerebras 3 brochure (WSE-3), which we have previously covered, contain up to 4 billions of transistors and 900,000 nuclei focused on a single unit. Another brochure processor, Tesla’s Dojo D1, is home to 1.25 billion of transistors and nearly 9,000 cores per module.
Processors delete common delays and energy losses in systems where data move between several chips.
“Keeping everything on a brochure, you avoid delays and loss of power of chip for chip,” said Ozkan.
Traditional GPUs are always significant because of their cost and lower modularity, but as IA models increase in size and complexity, fleas are starting to meet performance and energy barriers.
“AI computer science is no longer a question of speed,” said Ozkan. “It is a question of designing systems which can move massive quantities of data without overheating or consuming excessive electricity.”
Systems on the scale of versions also have significant environmental advantages. The Cerebras WSE-3, for example, can perform up to 125 quadrillion operations per second, while using much less energy than GPU configurations.
“Consider GPUs as highly frequented highways – but traffic jams waste energy,” said Ozkan. “Motors on a brochure scale look more like monorails: direct, efficient and less polluting.”
However, there remains a major challenge – the secular heat problem. Passes -scale chips can consume up to 10,000 watts of power, which turn into heat almost, requiring advanced cooling systems to prevent overheating and maintain performance.
Cerebras uses a glycol -based cooling loop integrated into the chip, while Tesla has a liquid system that uniformly distributes the coolant on the surface of the chip.
Via Tech Xplore
