Microsoft Azure CTO: US data centers will soon hit limits of energy grid

semafor.com

11 points·jonbaer··5 comments

The premise of this article is that when you train a model all cores need to be on a single cluster in the same physical location. Is that actually the case? I figured these jobs are massively parallelized and could be distributed to compute clusters around the world.

Isn't that what the article talks about in the section that begins with this?

>Given their AI ambitions, a solution could be building data centers in multiple locations to avoid overloading any one region’s power grid. It would be technically challenging, but it may be necessary, Russinovich told Semafor.

As if training an IA model required a whole datacenter .. my god, what a joke

Let's say, 1 GPU per RU on average (using the 8RU 8GPU from supermicro, and excluding a couple of RU for top of racks etc)

How much GPU can you run in an ENTIRE DATACENTER ? Is power even an issue ? In France, we have many datacenters located near power supply, for obvious reasons

To this point, honestly, and considering that datacenters runs millions of CPU today, this entire article seems like typical microsoft bullshit

You would have thought we would have transcended physical locations by now. It’s a totally valid argument. A GPU doesn’t care if it’s in Japan or Brazil or New York you would have thought you could distribute training work load across the world.

Unrelated, but I really liked the website design. The background colour and text font are nice.

Reply on news.ycombinator.com