Microsoft and Meta have been working on a new open rack design for AI data centers which separates power and compute into different cabinets.
Known as Mount Diablo, the architectural specifications of the system are being made available to members of the Open Compute Project.
Explaining why Mount Diablo has been developed, Jason Adrian, Microsoft’s general manager for Azure platform architecture, said the introduction of AI systems “has brought forth novel challenges and disruptions to the infrastructure that supports hyperscale datacenters," which make putting power and compute in a single cabinet problematic.
In a blog co-authored with colleagues Laurentiu Olariu and Banha Sok, Adrian said: “While compute and storage systems for the cloud usually have rack power densities below 20kW, AI systems are increasing rack power to hundreds of kW. To adapt to this fast-changing segment, we began to look at every layer of our infrastructure to optimize for these changes.
“Our solution is to separate the single rack into a server rack and a power rack, each optimized for its primary function.”
Adrian said Mount Diablo brings a range of benefits, including making better use of space in data centers: “Disaggregated power enables the entire server rack to be used for AI accelerators and scale up network switches to enable larger pods,” he said. “This optimization is crucial for performance and efficiency, enabling up to 35 percent more AI accelerators in each server rack.”
The blog does not detail how the cabinets are connected. Adrian said that, in future, the disaggregated power racks will allow AC inputs to be converted into 400Vdc. Current power solutions convert into 48Vdc, and Adrian argues 400V will be crucial for building more powerful and efficient AI systems. “With 400V we expect improvements and incremental evolution in improved efficiency, like what we have seen in the 48Vdc conversion space,” he said.
In his blog, Adrian said that industry-wide standardization, in areas such as connectivity and AC to DC converters, will be required if 400V is to take off.
“The disaggregated power rack enables scalability and flexibility in a time where innovation and time to market is of paramount importance,” Adrian said. “In an effort to move fast and shift the industry to HVDC power distribution, it’s critical to foster a healthy ecosystem and partnerships to drive commonality.”
It is not clear whether the design is in use in Microsoft or Meta’s data centers yet. DCD has approached the companies for further information.
Meta said Mount Diablo's "innovative design allows more AI accelerators per IT rack, significantly advancing AI infrastructure," adding that the company was proud to continue its collaboration with Microsoft and the Open Compute Project.