AI’s voracious need for computing power is threatening to overwhelm energy sources, requiring the industry to change its approach to the technology, according to Arm Holdings Plc Chief Executive Officer Rene Haas.
I wonder if they made an error as simple as this in their projections. There’s no guarantee that AI interest continues to grow.
Sounds like some sensationalized bullshit. They don’t give a single number or meaningful statement and they are paywalled.
I don’t disagree that they should back up their claim, but it does intuitively make sense. AI - GPT LLMs in particular - are typically designed to push the limits of what modern hardware can provide - essentially eating whatever power you can throw at it.
Pair this with a huge AI boom and corporate hype cycle, and it wouldn’t surprise me if it was consuming an incredible amount of power. It’s reminiscent of Bitcoin, from a resource perspective.
No, it makes no sense. India has over a billion people. There’s no way that amount of computing power could just magically have poofed into existence over the past few years, nor the power plants necessary to run all of that.
This is a future prediction, not a current observation.
I’m not saying it’s correct as a prediction, but “where are the extra power plants” is not good counter-argument.
A couple of months ago the average temperature where I live was well below freezing. Now it’s around twenty degrees C.
By this time next year it’ll be thousands of degrees!
The current LLM’S kinda suck, but companies have fired huge swaths of their staff and plan in putting LLMs in their place. Either those companies hire back all those workers, or they get the programs to not suck. And making LLMs actually capable of working unsupervised will take more and more energy.
My take is that LLMs are absolutely incredible… for personal use and hobby projects. I can’t think of a single task I would trust an LLM to perform entirely unsupervised in a business context.
Of course, that’s just where LLMs are at today, though. They’ll improve.
Sure, but it’s simply not physically possible for AI to be consuming that much power. Not enough computers exist, and not enough ability to manufacture new ones fast enough. There hasn’t been a giant surge of new power plants built in just the past few years, so if something was suddenly drawing an India’s worth of power then somewhere an India’s worth of consumers just went dark.
This just isn’t plausible.
If only there had been another widespread, wasteful prior use of expensive and power hungry compute equipment that suddenly became less valuable/effective and could quickly be repurposed to run LLMs…
Pretty sure the big AI corps aren’t depending on obsolete second-hand half-burned-out Ethereum mining rigs for their AI training.
Yeah, don’t AI everything, please.
I wonder why countries let them.
Using up more electric power than there’s available is NOT a simple matter of demand and supply.
If they actually pull too much from the grid, they are going to cause damage to others, and maybe even to the grid itself.
Because they’re not actually pulling too much from the grid to cause damage to others or even the grid itself.
Any musings about curtailing AI due to power consumption is just bullshit for clicks. We’ll improve efficiency and increase productivity, but we won’t reduce usage.
Improving the models doesn’t seem to work: https://arxiv.org/abs/2404.04125?
We comprehensively investigate this question across 34 models and five standard pretraining datasets (CC-3M, CC-12M, YFCC-15M, LAION-400M, LAION-Aesthetics), generating over 300GB of data artifacts. We consistently find that, far from exhibiting “zero-shot” generalization, multimodal models require exponentially more data to achieve linear improvements in downstream “zero-shot” performance, following a sample inefficient log-linear scaling trend.
It’s taking exponentially more data to get better results, and therefore, exponentially more energy. Even if something like analog training chips reduce energy usage ten fold, the exponential curve will just catch up again, and very quickly with results only marginally improved. Not only that, but you have to gather that much more data, and while the Internet is a vast datastore, the AI models have already absorbed much of it.
The implication is that the models are about as good as they will be without more fundamental breakthroughs. The thing about breakthroughs like that is that they could happen tomorrow, they could happen in 10 years, they could happen in 1000 years, or they could happen never.
Fermat’s Last Theorem remained an open problem for 358 years. Squaring the Circle remained open for over 2000 years. The Riemann Hypothesis has remained unsolved after more than 150 years. These things sometimes sit there for a long, long time, and not for lack of smart people trying to solve them.
Soon they’ll need to make Duracells out of humans
main use cases: government surveillance and chatbot girl friends
Weird metric, but ok.
It won’t be needed because nobody will have a job to pay for it. I forsee kurt vonnegut’s book “Player Piano” on steroids.
This focus on individual applications shifts blame onto consumers, when we should be demanding that energy prices include the external cost of production. It’s like guilt tripping over the “carbon footprint” (invented by big oil) of your car.
Take that, India! 😎
The ENIAC drew 174 kilowatts and weighed 30 tons. ENIAC drew this 174 kilowatts to achieve a few hundred-few thousand operations per second, while an iPhone 4 can handle 2 billion operations a second and draws maybe 1.5w under heavy load.
Like, yeah, obviously, the tech is inefficient right now, it’s just getting off the ground.
ML is not an ENIAC situation. Computers got more efficient not by doing fewer operations, but by making what they were already doing much more efficient.
The basic operations underlying ML (e.g. matrix multiplication) are already some of the most heavily optimized things around. ML is inefficient because it needs to do a lot of that. The problem is very different.
There’s an entire resurgence of research into alternative computing architectures right now, being led by some of the biggest names in computing, because of the limits we’ve hit with the von Neumann architecture as regards ML. I don’t see any reason to assume all of that research is guaranteed to fail.
I’m not assuming it’s going to fail, I’m just saying that the exponential gains seen in early computing are going to be much harder to come by because we’re not starting from the same grossly inefficient place.
As an FYI, most modern computers are modified Harvard architectures, not Von Neumann machines. There are other architectures being explored that are even more exotic, but I’m not aware of any that are massively better on the power side (vs simply being faster). The acceleration approaches that I’m aware of that are more (e.g. analog or optical accelerators) are also totally compatible with traditional Harvard/Von Neumann architectures.
And I don’t know that by comparing it to ENIAC I intended to suggest the exponential gains would be identical, but we are currently in a period of exponential gains in AI and it’s not exactly slowing down. It just seems unthoughtful and not very critical to measure the overall efficiency of a technology by its very earliest iterations, when the field it’s based on is moving as fast as AI is.
The ENIAC drew 174 kilowatts and weighed 30 tons.
it’s just getting off the ground
That’s what we’re afraid of, yes.
Yeah, uh huh, efficiency isn’t really a measure of absolute power use, it’s a measure of how much you get done with the power. Nobody calls you efficient if you do nothing and use no power to do that nothing. Google, Amazon, Microsoft, and Meta all together could not get anything done as companies if they all had to split an ENIAC (vastly less powerful than an older model iPhone) between them. This is a completely meaningless comparison.
Absolute power consumption does matter, but global power consumption is approximately 160,000 TWh, so the doubling means all the largest cloud providers all together are now using less than 0.05% of all the energy used across the world. And a chunk of that extra 36 TWh is going to their daily operations, not just their AI stuff.
The more context I add in to the picture, the less I’m worried about AI in particular. The overall growth model of our society is the problem, which is going to need to have political/economic solutions. Fixating on a new technology as the culprit is literally just Luddism all over again, and will have exactly as much impact in the long run.
Google, Amazon, Microsoft, and Meta all together could not get anything done as companies
Google’s biggest revenue stream is advertisement
Amazon’s biggest revenue stream is data hosting for national militaries and police forces.
Microsoft’s biggest revenue stream is subscriptions to software that was functionally complete 20 years ago
Meta’s biggest revenue stream is ads again
So 72-TWh of energy spent on Ads, Surveillance, Subscriptions, and Ads.
Absolute power consumption does matter, but global power consumption is approximately 160,000 TWh
If these firms were operating steel foundries or airlines at 72-TWh, I would applaud them for their efficiency. Shame they’re not producing anything of material value.
The more context I add in to the picture, the less I’m worried about AI in particular.
Its not for you to worry about. The decision to rapidly consume cheap energy and potable water is entirely beyond your control. Might as well find a silver lining in the next hurricane.
So 72-TWh of energy spent on Ads, Surveillance, Subscriptions, and Ads.
Capitalism truly does end up with the most efficient distribution of resources
I don’t like these companies for their cooperation/friendly attitude towards nation-states either, but your comments are insipid. AWS has like 2 million businesses as customers. They have 30% marketshare in the cloud space, of course they provide cloud services to cops and militaries. They’re cheap, and one of the biggest providers, period. I can’t find any numbers showing their state contracts outweigh their business contracts.
And, sure, plenty of those business contracts are for businesses that don’t do anything useful, but what you don’t seem to understand is that telecoms is vital to industry and literally always has been. It’s not like there’s a bunch of virtuous factories over here producing tons of steel and airplanes, and a bunch of computers stealing money over there. Those factories and airlines you laud are owned by businesses, who use computers and services like AWS to organize and streamline their operations. Computers are a key part of why any industry is as productive as it is today.
AI, and I don’t so much mean LLM’s and stable diffusion here, even if they are fun and eye-catching algorithms, will also contribute to streamlining operations of those virtuous steel foundries and airlines you approve so heartily of. They’re not counterposed to each other. Researchers are already making use of ML in the sciences to speed up research. That research will be applied in real-world industry. It’s all connected.
Its not for you to worry about. The decision to rapidly consume cheap energy and potable water is entirely beyond your control. Might as well find a silver lining in the next hurricane.
By the same token, you shouldn’t worry about it either? So insipid.
AWS has like 2 million businesses as customers.
None of them hold a candle to the Wild and Stormy Cloud Computing contact issued by the NSA.
I don’t like defending Amazon, but your arguments are shockingly ignorant. Stop making things up on the spot and do a shred of research. The cost of the Wild and Stormy contract is ~half a billion, while AWS’s annual revenues are projected to top $100 billion this year.
So, less than half a percent of AWS’s annual revenues. Stop just making shit up off the cuff.
The cost of the Wild and Stormy contract is ~half a billion
It’s ten billion.
