Schneider Electric warns that existing datacenters aren't buff enough for AI The infrastructure behind popular AI workloads is so demanding that Schneider Electric has suggested it may be time to reevaluate the way we build datacenters. In a recent white paper [PDF], the French multinational broke down several of the factors that make accommodating AI workloads so challenging and offered its guidance …
Simply, don't do it. Don't be that first mover, bide your time and move in quickly when it starts to actually become useful. The first movers will have burned a ton of cash on powering their data centres.
Also, "a 500-horsepower autonomous vehicle"... that scares the shit out of me. Would not trust that at all, just a way to bring death far more quickly. Perhaps a 50-horsepower, and limited to 20 mph with current technologies and I'd feel comfortable with it. Anyone else?
It has three phases worth of 240v to power three PDUs (hopefully checking the 120/240 switch on the servers you are plugging into those PDUs is appropriately set first lol)
I assume that's what he's talking about, since I've never seen anything made to fit in a standard rack with a power supply that could accept more than 220/240/250v (obviously there are things that require more power that were not in standard racks like mainframes or large disk arrays, but they were typically directly wired not plugged in via a PDU)
The US is Fed 240 volts and the voltage is split at the inverter coil down to 120V so premises by default are supplied on the 240 volt grid and can operate at 240V. just consumer electronics are dropped to 115V/120V for safety reasons as you get less dead with 115V across the heart as opposed to 230V/240V. American logic there somewhere
So more compute power needs more power to run it and the ancillaries like cooling. Tell us something i didnt know 20 years ago.
What is shocking is this is bringing real world figures to the table rather than theory. for essentially something we dont exactly need.
Lets be honest AI was born out of the need to answer customer service queries and calls. DOH for computer systems , Lets automate call handling !!!!!!!!!!!! and lets let computers start answering said calls.
From the article it seems an underlying assumption is that the main AI workload envisaged is the large scale repeated running of LLM style training loads, hence the need for performance. If you work on the assumption you only occasionally need to run training loads then you don’t need the density etc.
Also what was missing from this was any consideration of computer architecture.- Cray seemed to have solved the power, cooling and performance problems rather elegantly with the Cray-1, okay it doesn’t fit a 19-inch rack…
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
