Aug 24 (Reuters) – Cerebras Systems, the Silicon Valley startup making the world’s largest computer chip, said on Tuesday it can now assemble nearly 200 chips to dramatically reduce the power consumed by artificial intelligence work .
Cerebras is one of several startups making specially designed chips for AI and aiming to challenge current market leaders Nvidia Corp (NVDA.O) and Alphabet Inc (GOOGL.O) Google. The company has raised around $ 475 million in venture capital and made deals with pharmaceutical companies GlaxoSmithKline Plc (GSK.L) and AstraZeneca Plc (AZN.L) to use its chips to accelerate drug discovery.
Traditionally, hundreds or even thousands of computer chips are made on a 12 inch (30 cm) silicon disc called a wafer, which is then cut into individual chips. Cerebras, on the other hand, uses the entire blister pack. The huge Cerebras chip can hold more data at once.
But artificial intelligence researchers now have AI models called “neural networks” that are too big for a single chip to fit, so they have to spread them across multiple chips. Today’s largest neural networks are still only a fraction of the complexity of a human brain, but they use much more energy than human brains because the systems that run them become less energy efficient as more bullets are added.
Cerebras said on Wednesday it could put 192 of its chips together to form huge neural networks, but energy efficiency would remain the same as more chips are added. In other words, Cerebras can double the amount of computing power on its chips to double the power, unlike current systems that need more than twice as much power to double their computing capacity.
The current AI systems “are in the realm where you talk about tens of megawatts of power, and you do it over months. You use the equivalent of the power of a small town to form these networks,” said the general manager of Cerebras. Andrew Feldman told Reuters. “Power is extremely important.”
Reporting by Stephen Nellis in San Francisco Editing by Matthew Lewis
Our standards: Thomson Reuters Trust Principles.