How Does Cerebras Build the Fastest Data Center?

TL;DR

Cerebras has built the fastest AI infrastructure in the world, located in Oklahoma, leveraging a wafer-scale engine that significantly surpasses traditional chip sizes. The data center uses advanced liquid cooling systems and is powered by natural gas with backup generators to ensure consistent uptime. This facility represents a breakthrough in AI computing, promising advancements in fields like medicine and education.

Transcript

You open your AI chatbot. You type in your prompt and you hit enter. What happens next? We're pulling back the veil on the hidden backbone behind every AI response you see. Beneath the Oklahoma sky sits an unassuming concrete building. An AI factory built for one purpose. Speed. Heat. Heat. I'm standing in front of Cerebrus' brand new data center w... Read More

Key Insights

• Cerebras' data center in Oklahoma is the fastest AI infrastructure in the world, delivering 44 exaflops of compute power.
• The facility utilizes a wafer-scale engine, the largest processor ever built, which is significantly larger than traditional chips.
• Cerebras' chips keep memory on the chip, eliminating off-chip latency and providing a substantial performance advantage over traditional GPUs.
• The data center employs liquid cooling, starting from 2017, which enhances energy efficiency and maintains optimal operating temperatures.
• Power is primarily sourced from natural gas, with backup generators ensuring nearly uninterrupted service during outages.
• The facility's construction considers local weather conditions, using reinforced concrete for tornado resilience.
• Cerebras manufactures its systems in the United States, emphasizing domestic production and assembly.
• The data center's rapid expansion includes a second data hall, OKC2, which will add 20 exaflops of compute capacity.

Questions & Answers

Q: How does Cerebras' data center achieve its speed?

Cerebras' data center achieves its speed through the use of a wafer-scale engine, the largest processor ever built, which significantly surpasses traditional chip sizes. By keeping memory on the chip, Cerebras eliminates off-chip latency, providing a performance advantage over traditional GPUs. This design allows for faster data access and processing.

Q: Why was Oklahoma chosen for Cerebras' data center?

Oklahoma was chosen for Cerebras' data center due to its reasonable labor and power costs, and the availability of space for expansion. The facility's construction also considers local weather conditions, using reinforced concrete for tornado resilience. These factors make Oklahoma a strategic location for Cerebras' operations.

Q: What cooling system does Cerebras use in its data center?

Cerebras uses an advanced liquid cooling system in its data center, which began in 2017. This system involves chilled water circulating through the facility to manage the heat generated by the processors. The cooling system is energy efficient and maintains the chips at optimal operating temperatures, crucial for high-speed computing.

Q: What is the primary power source for Cerebras' data center?

The primary power source for Cerebras' data center is natural gas, which is converted into electricity. The facility also has backup generators that run on diesel or liquid natural gas, ensuring nearly uninterrupted service during power outages. This setup provides a reliable power supply essential for high-speed compute operations.

Q: How does Cerebras' chip design differ from traditional GPUs?

Cerebras' chip design differs from traditional GPUs by keeping memory on the chip, which eliminates the off-chip latency that slows traditional GPUs during inference. This design allows Cerebras' processors to access and use data much faster, providing a substantial performance advantage and enabling faster AI computations.

Q: What are the future expansion plans for Cerebras' data center?

Cerebras plans to expand its data center with the addition of a second data hall, OKC2, which is nearing completion. This expansion will add about 20 exaflops of compute capacity, significantly increasing the facility's ability to handle high-speed workloads. The expansion reflects Cerebras' commitment to scaling its AI infrastructure rapidly.

Q: Why does Cerebras manufacture its systems in the United States?

Cerebras manufactures its systems in the United States to emphasize domestic production and assembly. The company believes in being a good citizen in the economy by supporting local manufacturing. This approach also allows Cerebras to maintain control over the production process and ensure high-quality standards for its systems.

Q: What societal changes does Cerebras' technology aim to impact?

Cerebras' technology aims to impact societal changes in fields like medicine and education. In medicine, faster AI models can reduce the drug design process from 17-19 years to under 10 years. In education, AI offers opportunities to revolutionize teaching methods, moving beyond traditional approaches that have remained unchanged for centuries.

Summary & Key Takeaways

• Cerebras' Oklahoma data center is the fastest AI infrastructure globally, offering 44 exaflops of compute power. It features a wafer-scale engine, keeping memory on-chip to reduce latency and enhance performance. The facility uses liquid cooling and natural gas power, with backup generators ensuring uptime.

• The choice of Oklahoma for the data center was driven by factors like reasonable labor and power costs and the ability to expand. The facility is built with tornado resilience in mind, using reinforced concrete. Cerebras emphasizes domestic manufacturing, with systems assembled in California.

• Cerebras' breakthrough in AI computing is set to impact fields like medicine and education. The data center's rapid expansion includes a second hall, OKC2, which will significantly increase compute capacity. The company's commitment to innovation overcame initial challenges, resulting in a world-leading AI infrastructure.