OpenAI Taps Cerebras Chips to Boost
ChatGPT Speed and Inference Capacity
The agreement with the start-up Cerebras is the latest in a series intended to expand the
A.I. company’s computing power.
OpenAI
has signed a new agreement with Cerebras, a
California-based A.I. chip start-up, as part of its broader push to expand
computing power for its artificial intelligence systems. Under the deal, OpenAI
will begin using Cerebras chips for inference—serving
A.I. models to users—eventually consuming about 750 megawatts of electricity.
The
partnership adds to OpenAI’s growing roster of chip relationships, which
already include Nvidia and AMD chips and custom chip design work with Broadcom.
Together, these efforts reflect the massive scale of OpenAI’s infrastructure
buildout as it and other tech giants invest hundreds of billions of dollars in
A.I. data centers across the United States.
Cerebras is known for its unconventional approach
to chip design, including a dinner-plate-sized processor that keeps data on a
single massive chip to reduce bottlenecks and increase speed. OpenAI said the
collaboration would help make ChatGPT faster and more responsive, underscoring
how competition in A.I. is increasingly driven by access to specialized,
high-performance computing hardware.
After
signing deals to use computer chips from Nvidia and AMD and design its own chips
with Broadcom, OpenAI has reached an agreement with yet another chip maker.
OpenAI,
the maker of ChatGPT, said on Wednesday that it would begin using chips from Cerebras, a start-up in Sunnyvale, Calif. OpenAI said the number
of Cerebras chips it eventually used would consume 750
megawatts of electricity, an amount that could power tens of thousands of households.
The
agreement is the latest in a series with various partners as OpenAI works to expand
the computing power needed to build and deploy its artificial intelligence technologies,
including ChatGPT.
OpenAI
is one of many tech companies that are spending hundreds of billions of dollars
on new data centers for A.I. OpenAI, Amazon, Google, Meta
and Microsoft plan to spend more than $325 billion combined on these facilities
by the end of this year alone. The company is building data centers
in Abilene, Texas, and plans additional computing facilities in other parts of Texas,
New Mexico, Ohio and the Midwest.
OpenAI
previously said it would deploy enough Nvidia and AMD chips to consume 16 gigawatts
of power, which could power millions of households. The chips that the company is
designing with Broadcom are slated to consume 10 gigawatts.
As
OpenAI and its partners pack new chips into data centers,
some will be used to create its A.I. technologies. Others will serve these technologies
to people and businesses around the globe — a process that industry insiders call
inference. The Cerebras chips will be used for inference.
“This
partnership will make ChatGPT not just the most capable but also the fastest A.I.
platform in the world,” Greg Brockman, OpenAI’s president, said in a statement.
Co-founded
in 2015 by Andrew Feldman, a chip industry veteran who previously sold a start-up
to AMD, Cerebras is among a number of start-ups that have
spent years building chips just for A.I.
In
2019, the company unveiled what it called the largest chip ever built. As big as
a dinner plate — about 100 times the size of a typical chip — it would barely fit
in your lap.
A.I.
systems are typically powered by many chips that work together. But moving data
between chips can be slow, limiting how quickly A.I. software operates. While some
chip makers are broadening the pipes that run between chips, Cerebras has taken a different approach: Keep all the data on
one giant chip so a system can operate faster.