Hacker News Reader

Comment by didgetmaster | Hacker News Reader

What makes a data center an 'AI data center' vs other kinds? I am sure that certain workloads are better suited for a particular server rack vs another; but can't a data center built for other computing needs also do AI and vice-versa?

View on HN

Data center mech eng here - from our perspective it's higher rack densities typically due to GPUs. It's certainly possible to have high densities due to CPUs as well but I've seen a significant spike in rack densities in the last couple of years which has caused a switch from air cooling to liquid to chip.

One side effect of higher density is less footprint on the building to exhaust the heat, which is one reason (the main one being efficiency) that cooling towers and indirect evaporative cooling are favoured over air cooled condensers which leads to large amounts of water consumption.

Cooling towers are also much quieter than air cooled condensers which is a significant factor near any residential areas. It would be great to see more use of data center waste heat for process or district heating to save on water consumption.

Another issue with AI training in particular is huge (multi-MW) swings in power consumption at the start and end of each training run which must be a nightmare for the sparkies.

It is a nightmare for sparkies! See Meta's creative solution:

https://github.com/pytorch/pytorch/pull/132936/changes#diff-...

Isn't that just HPC?

The distinction is scale. "AI Datacenters" are a new level of scale with new levels of power consumption and heat generation. Sure you could run regular compute and w/e in them but it's not practical to build these mega sites for regular compute. GPU Compute / AI workloads require network/interconnect bandwidth and latencies where distance matters so you're forced to solve problems you wouldn't otherwise have to. Those problems are mostly solved with money.

Different I/O, power and cooling requirements for majority GPU workloads?

GPUs have been in high demand since cryptocurrency became a thing? Are you saying that something built for AI can't be used for other workloads?

This strikes me as a combination of semantics and false equivalence. You might as well argue that a new crowd of people illegally dirt-biking in a public park isn’t a meaningful change because people with baby strollers are have also technically been violating the “no vehicles” sign for years.

Not nearly with this density and power.

The power an "AI data center uses" in a single rack used to be, or is still in many cases, the power draw of an entire room or even floor.

Going from a few megawatts to ~10GW.

Did crypto workload ever take over an entire data center?

Yes.

https://www.riotplatforms.com/bitcoin-mining/corsicana/

> Riot Platforms has initiated a large-scale, 1 gigawatt development to expand its Bitcoin mining and hosting capabilities in Navarro County, Texas with its new Corsicana Facility.

> Development of the Corsicana Facility has begun with an initial 400 megawatts of capacity on a 265-acre site. The substation was energized in April 2024 and mining operations have begun.

https://comptroller.texas.gov/economy/fiscal-notes/archive/2...

> In 2008, the city of Rockdale lost about 80 percent of its workforce following the closure of the Alcoa steel plant. Today, the old Alcoa plant is occupied by Riot Blockchain’s Whinstone facility, believed to be the largest single Bitcoin mining operation in North America. As an industry that relies on high levels of electricity, the company was drawn to the facility due to its existing power infrastructure, including valuable high-voltage transmission lines and large substations.

Bitcoin Mining is 138–205 TWh annually. Surely that's more than a few data centers.

Reply on news.ycombinator.com