SAN FRANCISCO, March 13 (Reuters) - Artificial

intelligence startup Cerebras Systems announced a new version of

its dinner-plate-sized chips on Wednesday, claiming the hardware

will offer twice the performance for the same price as its

predecessor.

WHY IT IS IMPORTANT

Santa Clara, California-based Cerebras' AI chips compete

with the advanced hardware produced by Nvidia ( NVDA ) that help

OpenAI develop the underlying software that powers apps such as

ChatGPT. Instead of stitching together thousands of chips to

build and run AI applications, Cerebras has bet that its roughly

foot-wide chip can outperform Nvidia's ( NVDA ) clusters of chips.

KEY QUOTE

"So the largest chip that we made was our first generation.

People said we couldn't make it," Cerebras CEO Andrew Feldman

said to reporters on Tuesday. "Eighteen months later we did it

in seven nanometer. Eighteen months (after that), we've

announced a five-nanometer part. This is the largest part by

more than three and a half trillion transistors."

CONTEXT

Power consumption is a critical problem for AI processing.

Cerebras' third-generation chip uses the same amount of energy

to achieve superior performance, when power costs to build and

run AI applications have soared. Cerebras does not sell the

chips by themselves, but says the systems constructed around

them are a more efficient method of building AI applications, a

process called training.

BY THE NUMBERS

The new Wafer-Scale Engine 3 (WSE-3) has 4 trillion

transistors capable of performing 125 petaflops of computing. It

was built on Taiwan Semiconductor Manufacturing Co's ( TSM )

5nm manufacturing process.

Feldman said Cerebras is cash flow positive

WHAT IS NEXT

Cerebras also said on Wednesday that it planned to sell its

WSE-3 systems together with Qualcomm AI 100 Ultra chips

to help run artificial intelligence applications, a process

known as inference.