MIT ORCD to bolster computing cluster with over 200 NVIDIA B200 GPUs
Most of these new chips will be available to the entire Institute
MIT’s Office of Research Computing and Data (ORCD) is set to deploy over 200 NVIDIA B200 graphics processing units (GPUs), some of the most advanced GPUs for training AI models, following a $31 million matching grant from Massachusetts Governor Maura Healey.
Most of these new state-of-the-art GPUs will join hundreds of older advanced NVIDIA devices, including H200 and L40 GPUs, already maintained by ORCD on the Engaging cluster, a shared computing resource available to the entire MIT community. These machines are hosted off campus in the Massachusetts Green High Performance Computing Center (MGHPCC).
The Office of Research Computing and Data was established in Sept. 2022 to address the “computing and data needs from the entire MIT research and education community.” Its services are especially useful for training or running large AI models, which generally require distributing computationally intensive tasks across multiple GPUs.
In late January, ORCD Head and Thomas A. Frank (1977) Professor of Physics Peter Fisher and ORCD Executive Director Chris Hill announced that ORCD would begin adding a fee-based tier to “provide PIs the opportunity to run longer” and more intensive jobs for research. However, they noted that the office would still continue to offer free “base level” usage of GPUs and CPUs to all MIT community members.
Per ORCD’s January newsletter, the office is targeting around “100 public B200s on Engaging by mid-year” with access across free and paid tiers. However, supply chain issues surrounding chip manufacturing have added uncertainty to these timelines. “We are routinely seeing months of delays on commodity parts and vendor price fluctuations of 100% or more on some components,” the newsletter said.
The B200 model, according to NVIDIA, provides “three times the training” and “fifteen times the inference performance” compared to “previous-generation systems.” In a statement, MIT Vice President for Research Ian Waitz expressed optimism about the new devices.
“Scholarship and innovation are increasingly fueled by computation,” said Waitz. “The broad access to hardware at MGHPCC will support the pursuit of new artificial intelligence discoveries, practices and solutions that can benefit the state and the nation.”
The MGHPCC also consumes “100% carbon-free” power, according to Massachusetts AI Hub Director Sabrina Mansur.