MITRE promises a cute little 17-PFLOPS AI super for the rest of Uncle Sam's agencies
No child process left behind
Later this year, MITRE is getting its hands on a modest supercomputer and is planning to use it to divvy out AI computing time to US government agencies.
MITRE is a not-for-profit corporation tasked with managing tax-funded research and development for Uncle Sam, though it's perhaps best known among Register readers for overseeing the CVE database of security vulnerabilities.
The organization's upcoming Federal AI Sandbox is basically an Nvidia DGX SuperPOD – a modular supercomputer built from 256 Nv H100 GPUs, with each chip capable of achieving roughly four petaFLOPS of performance at 8-bit precision for a total of one exaFLOPS of AI-level performance.
The system – set to be deployed in Ashburn, Virginia – marks a substantial upgrade to MITRE's existing AI infrastructure, boosting its compute resources by two orders of magnitude, the group claims.
But, as supercomputers go, the 32-node system is tiny compared to those operated by the US Department of Energy. For reference, Oak Ridge National Laboratory's Frontier machine is the number one ranked system in the Top500 – at least until the next results drop next week, anyway – and features 37,888 AMD MI250X GPUs. Its peak FP64 performance – the metric by which supercomputers are usually compared – exceeds one exaFLOPS.
With roughly 17 petaFLOPS of FP64 oomph, the DGX SuperPOD MITRE plans to deploy wouldn't even rank in the Top500's 50 most powerful publicly known beasts.
- Amazon to blow $11B on cluster of Indiana bit barns
- Google will pump more than $100B into AI, says DeepMind boss
- Cheyenne supercomputer sells at auction for just $480K
- IBM quantum system elbows into Arm-powered Fugaku supercomputer
That said, most AI workloads can get away with far lower precision than those ranked in the Top500, hence the claimed AI exaFLOP of performance. And in this case any ML compute capacity is probably better than none as MITRE notes most US agencies simply don’t have the resources necessary to train novel models. Well, until now.
"The recent executive order on AI encourages federal agencies to reduce barriers for AI adoptions, but agencies often lack the computing environment necessary for experimentation and prototyping," MITRE CTO Charles Clancy explained in a statement, referring to this push by President Biden in October for more trustworthy artificial intelligence.
"Our new Federal AI Sandbox will help level the playing field, making the high-quality compute power needed to train and test custom AI solutions available to any agency."
MITRE argues the machine will provide adequate compute to develop new AI applications, including large language models and multimodal perception systems. The latter, the group contends, would be capable of processing multiple data feeds ranging from images, audio, text, radar, environmental, and medical sensors.
Though the non-profit is well known for its interest in cybersecurity and vulnerability tracking, MITRE plans to open its SuperPOD up to agencies as a secure test bed to run workloads encompassing a wide range of things, including national security, healthcare, transportation, and climate research.
US agencies with existing contracts with any of MITRE's research and development centers will have access to the machine when it comes online at the end of the year. ®