It’s impossible not to happen across some amazing worthwhile AI derived solutions to problems. Medical diagnosis, language translation, statistical analysis/prediction; benefitting many of us daily. These solutions are the result of thousands of hours of expensive energy hungry compute time, both in their creation but also their continued operation. Requiring always-on servers filled with specialist hardware. For valuable beneficial products this is surely a small price when compared to the myriad of benefits to you and I.
Unfortunately the availability of this new technology isn’t entirely being used to better humankind. The ability to generate an image of a cat paragliding, or a country song about how much you like cheese requires the same expensive training routines and operating costs. Typically running in the same data centers and hardware as your favourite AI translation engine. Often fighting for time on that hardware, driving the demand for more powerful servers, in new larger data centers.
“AI” being the hot new craze in business is one thing, using buzz words rather than functionality to drive sales. But seeing countries jumping aboard AI as a magic elixir to all ailments while at the same time championing their dedication to greener energy and combatting climate change is nothing short of insane.
Most typical hosting or datacenter workloads are demand based. Ie. someone requests a webpage or to start streaming a film and the datacenter responds, sending data when it’s required. AI training workloads on the other hand consume an enormous amount of power across many servers until the job is complete. This can be turned on/off to consume power when it’s cheapest or to pick up spare green energy that isn’t required by homes/businesses at the time.
Making use of potentially wasted green energy sounds like a positive thing. Except we’ve already solved that problem. Dozens of high altitude reservoirs across the globe act as efficient ways to store this unused energy for use later on by pumping water to a higher reservoir during periods of excess power generation. We then let it run back down again through turbines to generate electricity. These pumped hydro plants also perform a very important role in filling spikes in energy generation.
All this infrastructure that is absolutely necessary for achieving any clean energy future disappears if we have AI training workflows waiting to gobble up cheaper energy as it becomes available.
This leaves us at a bit of an impasse:
Embrace technological growth at the expense of exponential increases in energy usage?
Enforce limits on how much on-demand energy can be consumed by data centers or limit their size/growth. Both causing an increase in the cost of bringing AI services to market.
Ben Ford
Chief Technology Officer (TCW)