Lockmiller: Synchronized GPU clusters create extreme power spikes like breathing
Chase Lockmiller · Frontier Forum: An energy-first approach to data centers · Oct 22, 2024 · at 19:00
Crusoe CEO Chase Lockmiller explains the power management challenges that arise when training frontier AI models at massive scale.
0:00 / 0:33exact quote · 33.8s
720p mp4 · rendered on demand · StarZero watermark
“The cluster itself is like moving in, in, in sync with one another. So, you know, you're sort of, ah, all of the data is sort of being broadcast out across this high performance network, you know, to the GPUs. They're running their compute workloads and then publishing the results to everything. So it's almost like the cluster is breathing. Like you have these like big spikes in power and then sort of low, low spikes in power and then big spikes in power. And it's like basically every GPU is sort of moving in sync. And that frequency, and, you know, they're doing this hundreds of times or thousands of times a second.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Chase Lockmiller
Lockmiller: Power is now the primary constraint scaling AI
“Power is the main constraint right now.”
Lockmiller: Half an H100 per US adult would require 250 GW
“If every, you know, American adult used half an H 100 as, You know, a form of co-pilot for their, you know, daily workflows and, you know, social interactions, et cetera. Like, that would require 250 gigawatts of power, right?”
Lockmiller: Crusoe is developing compute at an upstate New York Alcoa site
“You know, other markets like upstate New York you know, we're working on a former Alcoa factory that's you know, powered by a large hydro dam that's, you know, being currently massively underutilized.”
Lockmiller: Remote AI training in clean-energy hubs improves on aluminum smelting model
“Training a large language model is a far better manifestation of that trade where, you know, you can move data across subsea cables to get to a data center based in Iceland. You can run your back propagation and large scale pre-training on significant amounts …”
Lockmiller: AI data centers drive massive shortages of electricians, welders, and plumbers
“We're experiencing massive shortages of electricians, of welders, of plumbers, like these large scale liquid cooling architectures that we've described that require tremendous amounts of plumbing, right? So you're building all these different pipes to move wat…”