With the attention gained by Artificial Intelligence such as Chat GPT performing impressive creative tasks, it is easy to overlook the significant amounts of power needed to enable these tools.
Some researchers involved in the development of these platforms are already saying that the use of the current Large Language Modules (LLMs) is unsustainable in terms of energy consumption.
Where do these platforms sit? In data centres of course!
The power consumed by LLMs and the increasing number of AI applications are now being recognised as adding to data centre power needs.
This is inevitably going to be a growing problem for the current models, some of which rely on existing ‘unused’ processing capacity, and are therefore currently hidden in terms of increased data centre power consumption.
As power consumption rises, this will inevitably result in increasing calls from poorly informed media and governments to curb data centre power usage, while failing to understand where this demand comes from – all of us using Chat GPT!
Additionally, high-power demands of the new AI GPU Chipsets from the likes of Nvidia are going to be increasingly difficult to host in traditional data centres.
This is due to the dramatically increased power density requirements, and individual cabinet power consumption being closer to High Performance Computing (HPC) loads than more typical lower colocation loads.
A single AI chipset can consume almost 1 kW, with some individual servers reportedly consuming up to 17 kW. This produces both problems and opportunities for data centre operators, ironically some of which may be solved using AI!
Generative AI is very different to what might be considered true intelligence though. This is described as Artificial General Intelligence (AGI).
AGI emulates human intelligence and remains theoretical, with some even doubting that it will ever be truly possible; others have suggested it may be possible by the end of the decade.
Generative AI is very much with us now, and is extremely good at solving specific problems or dealing with very specific tasks by offering a statistically valid representation based on the analysis and use of already available content, rather than being truly ‘intelligent.’
Generative AI is built for a specific task, and cannot perform outside the narrow boundaries for which it is designed.
By contrast, AGI should be able to learn and develop solutions to entirely new problems in unfamiliar situations, without being programmed to do so, and without necessarily having past experience or knowledge.
Artificial Super Intelligence (ASI) is yet another step, and describes intelligence above and beyond human; although theoretical, this is where things get really interesting, and words of caution become meaningful.
In this case, the word ‘Artificial’ may even have to be dropped as it would simply be intelligence at an incomparable level. Generative AI is not going to take over the planet – ASI might!
One thing is certain, AI will inevitably follow Amara’s Law; the initial promise may be inflated, but the long-term impact is likely to be highly underestimated.
…
Disclaimer: This post by Mark Acton was originally created for September’s issue of Inside Networks. Find the original article here.