In my masters degree I always ran many computations as did all my peers
The reality is that more of us are than not are using huge HPC clusters / cloud computing for many hours on each project
The industry is just GPUs going BRRR
I’m wondering if this has potential implications for ML in society as AI/ML becomes more mainstream
I could see this narrative being easily played in legacy media
Ps - yeah while there are researchers trying to make things more efficient, the general trend is that we are using more GPU hours per year in order to continue innovation at the forefront of artificial inference
If you compare it to cars it is environment friendly
Human extinction from AI is a bigger problem than the carbon emissions.
Depends on if it solves the environment problem.
You are bad for the environment bro
Just large models, small models have been used for ages
Sasha Luccioni at HuggingFace is doing a lot of great work on exactly this. https://www.sashaluccioni.com/
Every single form of compute is carbon neutral.
We don’t have CO2 or methane spewing out of gpus when they run.
After training AI actually emits less CO2 emissions than humans working on the same thing. So the initial power draw is just a sacrifice for longer term eco friendly. This isn’t dumping GPU into ponds like manufacturing that actually decimates ecosystems sources of water.
Training models uses lots of energy, but so does every other human activity. Even for large companies like Google.
Google used 15,439 GW hrs in 2020. The average per capita US energy consumption is 311 GJ. That comes out to the energy equivalent of 178,715 people. Google had 135,300 employees in 2020. Barely above average energy usage and probably below average for the income google makes. Those 135,300 employees probably easily exceed the company’s energy usage with their normal household spending.
I think server clusters cost around 2-3 % of energy already. That is about airlines levels of energy usage.
Analogue chips are prob. going to help this not accelerate more than absolutely needed the next decade or two.
I mean it’s all relative, and in many cases where ML-based systems are saving use of the alternative solutions that cost a lot more energy, no ML is helping the environment. As it stands though, yes, many of the areas are incredibly energy-consuming.
Pretty sure GPUs aren’t nowhere near as harmful to the environment as e.g. air flights. The main harm comes from throwing away GPUs which are still perfectly functional.
I’m excited to see if IBM Hermes could affect this huge demand with more efficient power usage.
No reason to assume so. The largest players in AI/cloud, Google and Microsoft, are firmly on track to become carbon neutral and make significant investments in renewable energy.
Using energy isn’t the same thing as creating emissions – it depends on your source.
Machine learning also has the ability to streamline many energy intensive operations. One recent example is DeepMind generating an accurate 10-day weather forecast in under a minute which used to take hours of computation.
Or significantly speeding up drug discovery and materials research cutting out lengthy rounds of experimentation.
Unlikely … as long as the price of electricity isn’t kept artificially low, and possibly even then. The use of AI has purpose, and is going to give something in return for that electricity. As long as the AI is used for a productive purpose, it will be a net positive.
Now replace “AI” with “bitcoin”, and the answer would change.
Why do you not see the value with Bitcoin? It is a decentralised currency - that is valuable to a lot of people
Not going to get into a big debate on this one … but the market cap of bitcoin is $1.44 Trillion at the moment. Where did this “wealth” come from? Well from nothing, and it can’t be converted into anything physical or otherwise useful, so it is the currency aspect only. I don’t have recent numbers, but around 2022, the network was using 131.26 terawatt-hours of electricity annually. No idea what the cost is in terms of hardware and labour misallocation. That is an insanely inefficient decenstralised currency, so extremely unlikely to be a net positive.