I’m sharing this here to come back to later, as I host a podcast called Environment Variables and I recently did an interview with some folks who are working on getting a standards group off the ground focussed on how datacentre hardware and software work together. There was one claim that during the interview I didn’t really spend much time sense checking and I wish I did, so I’m leaving a note here.
What am I sense checking?
Here’s the podcast, Environment Variables – Hardware Standards Working Group, and here’s the specific clip from the podcast.
Here’s a summarised version of the quote:
Modern AI/ML workloads create extreme power fluctuations (up to 200 MW in 40ms), equivalent to a quarter-million people suddenly appearing on the electrical grid
At the time, I was trying to translate it into something listeners might be able to understand, more than try to visualise what kind of facility size you would need for a such a 200MW swing to be possible. Swinging 100% power draw like seems incredibly high, as that’s not that far off the swing in the grid that caused the grid outage affecting the entire Iberian Peninsula earlier this year.
How big a datacentre would you need for this swing?
I’ve spoken to people who run the LUMI supercomputer in Finland, and they told me they’ve seen power swings of a 1-2 megawatts, on a supercomputing cluster than draws around 7 megawatts. So that’s around 25%, on a system that isn’t that GPU heavy.
With more GPU heavy facilities I can imagine the power swing being larger, but I don’t have direct access to data to develop an intuition about how much larger.
Would you need a single cluster pulling something like… 500 MW of power for it to swing 200MW?
Trying to learn more
I’ve emailed the people on the podcast to learn more, but in the meantime, I’d be very grateful for any articles, papers or similar that would make for extra reading because this throwaway remark has stayed with me since the recording.