Some Summer GPU Statistics
Even though summer is not as hectic in some ways on campus, the GPUs are busy working. Of the 988 researchers using the ORCD Engaging cluster for Slurm sessions during July, 418 used a GPU in their sessions.
In sessions requesting specific GPUs
- 111 people requested H200 GPUs
- 197 people requested L40S GPUs
- 76 people requested H100 GPUs
- 34 requested A100 GPUs
By GPU type the number of GPU sessions run amounted to
- 7,463 jobs were run on H200 GPUs
- 16,810 jobs were run on L40S GPUs (our current default)
- 2,497 jobs were run on H100 GPUs
- 4,009 jobs were run on A100 GPUs
The month total came out to just under 200,000 GPU hours of computation - of various sorts! By comparison, estimates of V100 GPU hours for ChatGPT-3 are around 3 million hours.
AMD MI300 GPU Projects
AMD has credits available for researchers interested in dedicated cloud access to new 8-way MI300X GPU servers. The AMD team is looking for 5-10 initial projects to work with over an initial 3 month period. If you are interested and have a possible project please reach out to orcd-help@mit.edu with a brief summary of a project and the team involved.
ENI, a founding member of the MIT Energy Initiative, is also soliciting proposals from projects interested in new AMD GPUs.
New ORCD Scratch Storage
We sent an email to active Engaging cluster users earlier this month, and we wanted to remind everyone that Engaging's Lustre filesystem (/nobackup1, including /nobackup1b and /nobackup1c) will be retired at the end of the Summer on August 15. Any data left on nobackup will be gone when Lustre is turned off.
In place of nobackup we have set up a new scratch filesystem. Everyone now has their own scratch directory, which is available through the path ~/orcd/scratch. We recommend using Globus for large file transfers. See our documentation for details on the new scratch filesystem and file transfers. Contact orcd-help@mit.edu for help.
What We’re Reading
- National AI strategy from OSTP, including enthusiastic references to NAIRR academic infrastructure activities.
- Comments on NIST800-234 accepted until August 3.
- A reminder why those primary producers matter, hug a tree…
- MIT materials discovery leveraging AI on AWS NOVA from Wojciech Matusik and collaborators featured at AWS Summit
- Write-up of AI accelerating engineering of AI, nibbling its tail a little at least…
- Farewell Tom Lehrer (We asked ChatGPT to write a ditty about the AI data center boom in Tom Lehrer’s style. It no doubt plagiarized a bit, but it got all the elements.)
ORCD Summer Training Schedule
ORCD has one more on-campus summer training session coming up. Our available class spots have filled up, but you can reach out to orcd-help@mit.edu to request to be added to the waitlist.
- Introduction to Engaging: August 27, 2-4pm
See more details on ORCD’s website.
Want to schedule an onsite ORCD training session for your DLCI? Send an email to orcd-help@mit.edu and we can work with you to schedule a session.