Inside a Cerebras Data Center in Oklahoma City

SUMMARY

In 30 seconds

Matthew Berman describes a visit to a Cerebras data center, where he says the company’s CEO showed him its infrastructure. He highlights the size of the wafer-scale chip, on-chip memory, liquid cooling and protection against tornadoes; the short video does not supply a benchmark for its speed claims.

1A dinner-plate-sized chipBerman introduces the Cerebras wafer-scale engine and says its size enables memory to be kept on the chip.
2A facility built around the chipsHe describes two large halls with hundreds of racks in each.
3Cooling and condensationA closed-loop system cools the racks, while water is warmed when necessary to avoid moisture forming on the chip.
4Concrete protectionBerman says the facility is protected with substantial concrete because of the region’s tornadoes.

Takeaways

Cerebras builds a wafer-scale chip that Berman says is about 50 times the size of a conventional chip.
Berman attributes reduced latency to the chip’s on-chip memory and claims up to 30 times the speed of a traditional chip, without specifying a benchmark.
The facility’s racks use a closed-loop cooling system, with water temperature controlled to prevent condensation.
Berman says the building uses substantial concrete protection against tornadoes.

TopicsInference Infrastructure & deployment AI cost & performance

FULL TEXT

Read the edited text by chapter

Show transcript

Transcript only

The wafer-scale engine

Berman introduces the large Cerebras chip and its on-chip memory. He makes a speed comparison without giving the benchmark or workload.

SLIDE 1 / 2A wafer-scale engineThe chip is held up during the introduction, showing its physical scale.

What if a chip were the size of a dinner plate? That is the idea behind the Cerebras wafer-scale engine. I say it is about 50 times the size of a traditional chip, with enough space to keep memory on the chip itself and reduce latency. I also say it can be up to 30 times faster than a traditional chip; this brief video does not specify the workload, baseline or benchmark behind that comparison.

Inside the data center

Berman describes a tour of the Oklahoma City facility and its two halls of racks. He highlights their AI inference use.

Cerebras built a data center in Oklahoma City to house these systems, and its CEO took me on a tour. I describe two enormous halls, each containing hundreds of racks equipped with wafer-scale chips for AI inference. The video calls the infrastructure extremely fast, but does not establish a universal ranking against other data centers.

Cooling and protection

He describes the closed-loop cooling system and the need to avoid condensation. The building also uses concrete protection against tornadoes.

SLIDE 2 / 2Data center cooling infrastructureThe footage shows equipment and piping as Berman describes the closed-loop cooling system.

The racks are connected to a closed-loop cooling system. I say the system cools each chip to 70 degrees, although the clip does not state the temperature scale. The water must sometimes be warmed because the cooler could otherwise lead to moisture forming on the chip. The facility also uses substantial concrete protection against tornadoes common in the area.