Nvidia's blueprint to build gigawatt-scale AI data centres faster
Design virtually and optimise power use - all before construction begins.
Nvidia today unveiled a blueprint to build more efficient, gigawatt-scale AI data centres in less time. Here's what it is really about.
Nvidia today took the wraps off its Omniverse DSX, which it called a "comprehensive, open blueprint for designing and operating gigawatt-scale AI factories." Here's what it means for the data centres of the future.
Building data centres
Data centres used to take a long time to build. From land acquisition, design, approvals, piling, construction, to fit-out, a data centre would take two to four years to complete.
Then the post-COVID data centre boom happened. With prefabricated parts and new techniques driven by demand, time-to-market was significantly reduced. Indeed, Johor probably has one of the shortest data centre construction times for new data centres in the world outside of China, down to less than a year in some instances.
Fast and faster
But then came the rush for AI data centres. In the race to AGI or world-changing AI, even a year is too long. Defying conventional approaches, Elon Musk built his Colossus AI data centre in a record 122 days.
The incredible build times of this and other AI data centres in the US offer a glimpse of what is possible when enough effort, ready access to GPUs, and money are available.
Despite multiple successes, however, fast build times remain difficult to replicate consistently. This is what Nvidia's Omniverse DSX seeks to address.
Omniverse DSX way
The Omniverse DSX offers a blueprint to build AI data centres that operate at maximum efficiency and collaborate with hundreds of suppliers to build it.
Nvidia focused on the following outcomes: build AI data centres much faster, optimize for data centre-wide efficiency, and grid-to-rack design for maximum power use.
Building virtually
How does Nvidia achieve what others couldn't do? Two main strategies: using advanced digital twins and rallying the data centre ecosystem.
The new AI data centre is designed virtually, with validated components from suppliers simulated working together before an order is even placed. When operational, the digital twin is used with AI software to optimize power consumption across the entire data centre to save money.
What does it mean for traditional data centres? I reckon the impact won't be immediate, but it will surely accelerate the adoption of digital twins and facility-wide power management.
What do you think?
