Despite its name, the infrastructure used by the “cloud” accounts for more global greenhouse emissions than commercial flights. In 2018, for instance, the 5bn YouTube hits for the viral song Despacito used the same amount of energy it would take to heat 40,000 US homes annually.
Large language models such as ChatGPT are some of the most energy-guzzling technologies of all. Research suggests, for instance, that about 700,000 litres of water could have been used to cool the machines that trained ChatGPT-3 at Microsoft’s data facilities.
Additionally, as these companies aim to reduce their reliance on fossil fuels, they may opt to base their datacentres in regions with cheaper electricity, such as the southern US, potentially exacerbating water consumption issues in drier parts of the world.
Furthermore, while minerals such as lithium and cobalt are most commonly associated with batteries in the motor sector, they are also crucial for the batteries used in datacentres. The extraction process often involves significant water usage and can lead to pollution, undermining water security. The extraction of these minerals are also often linked to human rights violations and poor labour standards. Trying to achieve one climate goal of limiting our dependence on fossil fuels can compromise another goal, of ensuring everyone has a safe and accessible water supply.
Moreover, when significant energy resources are allocated to tech-related endeavours, it can lead to energy shortages for essential needs such as residential power supply. Recent data from the UK shows that the country’s outdated electricity network is holding back affordable housing projects.
In other words, policy needs to be designed not to pick sectors or technologies as “winners”, but to pick the willing by providing support that is conditional on companies moving in the right direction. Making disclosure of environmental practices and impacts a condition for government support could ensure greater transparency and accountability.
Environment doesn't stop at electricity costs, it's also about manufacturing.
A simple terminal is more efficient to produce and has way longer lifespan, removing the need to update it for many, many years.
And then you can tie it either to your existing PC (which you need anyway) or cloud (which is used by other players when you're not playing, again reducing the need for components).
That's what I meant there. Generally, from an energy standpoint, gaming can absolutely be made more energy-efficient if hardware would put it as a priority. You can make a gaming machine that needs 15W or 1500W, depending on how you set it up.
Yes and manufacturing an Xbox for every single household, boxing it and shipping it to them, and then having it sit unused for 90% of the time, has a much bigger carbon cost than manufacturing a fraction of the number of Xboxes, shipping them all in bulk to the same data center, and then having them run almost 24/7 and be shared amongst everyone.
And the same thing about optimizing gaming hardware is true for AI. The new NPUs in the surface laptops can run AI models on 30W of power that my 300W GPU from 2 years ago cannot.
I feel like we went onto two very different planes here.
Sure, data centers are more efficient than a decentralized system, but the question is, to what point the limitless hogging of power and resources makes sense?
Sure, a lot of computing power goes into, say, console gaming, but that's not what I originally talked about. I talked about data centers training AI models and requiring ever more power and hardware as compared to what we expend on gaming, first of all.
And while in gaming the requirements are more or less shaped by the improvements to the hardware, for AI training this isn't enough, so the growth is horizontal, with more and more computing power and electricity spent.
And besides, we should ideally curb the consumption of both industries anyway.
But they don't. Right now the GPU powering every console, gaming PC, developer PC, graphic artist, twitch streamer, YouTube recap, etc. consumer far far more power than LLM training.
And LLM training is still largely being done on GPUs which aren't designed for it, as opposed to NPUs that can do so more efficiently at the chip level.
I understand the idea that AI training will always inherently consumer power because you can always train a model on bigger or more data, or train more parameters, but most uses of AI are not training, they're just users using an existing trained model. Google's base search infrastructure also took a lot more carbon to build initially than is accounted for when they calculate the carbon cost of an individual search.
Fair enough - I just hope the advancements in AI do not outpace our capabilities in producing a better hardware for the job, and that what's left after finds a good use in other tasks.
Because otherwise it will grow more and more into a huge ecological problem.