Catch up on stories from the past week (and beyond) at the Slashdot story archive

 


Forgot your password?
Close
typodupeerror
×
Power Hardware

AI Needs So Much Electricity That Tech Companies Are Getting Into Energy Business (sherwood.news) 50

Posted by msmash from the adjacent-race-begins dept.
An anonymous reader shares a report: To accommodate tech companies' pivots to artificial intelligence, tech companies are increasingly investing in ways to power AI's immense electricity needs. Most recently, OpenAI CEO Sam Altman invested in Exowatt, a company using solar power to feed data centers, according to the Wall Street Journal. That's on the heals of OpenAI partner, Microsoft, working on getting approval for nuclear energy to help power its AI operations. Last year Amazon, which is a major investor in AI company Anthropic, said it invested in more than 100 renewable energy projects, making it the "world's largest corporate purchaser of renewable energy for the fourth year in a row."

AI Needs So Much Electricity That Tech Companies Are Getting Into Energy Business More | Reply

AI Needs So Much Electricity That Tech Companies Are Getting Into Energy Business

Comments Filter:

  • "On the heals"? (Score:4, Insightful)

    by 0xG ( 712423 ) on Monday April 22, 2024 @12:13PM (#64414454)

    I won't bother with sherwood.news. No need to read any further.

    • /. editors had ONE job. /s

      Guess no one cares about spelling anymore (original author or the editors.) :-/

    • Man if typos or grammatical mistakes caused you to not bother with a site you'd have closed your Slashdot account down back in the 90s.

      • Typos and grammatical mistakes are fine when the general public does this. But it used to be part and parcel of the journalism trade to actually be careful with stories, including with the spelling and grammar, you couldn't be a good journalistic publisher and have bad spelling, even if the original authors didn't spell well you were expected to do the proper work of editing and fixing articles. Or at the very least be smart enough to insert "[sic]" where apropriate. But today, it's just slam the stories

  • These companies have a lot of cash right now, but likely won't get many returns investing it in more new tech. Might as well vertically integrate and invest in something that's otherwise an expense anyway.

  • Once AI starts going on an unfettered rampage, we can scorch the sky so there's no sunlight for them to power up.

    A perfect plan indeed.

    • Re: (Score:2)

      by guruevi ( 827432 )

      Most of them are actually relying on coal, gas and nuclear reactors. The selling of carbon credits to China for solar energy is a scam, it does not actually make your energy come from solar panels, otherwise the AI datacenters would be shutting down at night.

    • Humans will make good power sources.

  • That's the real reason the only data center certification anyone cares about is actually just a financial audit.

  • Time to move to China, where cheap coal power is plentiful.

  • Refocus on hardware (Score:3)

    by Baron_Yam ( 643147 ) on Monday April 22, 2024 @12:52PM (#64414598)

    The human brain runs on about 29 watts.

    Let's focus on hardware improvements rather than massive power generation increases.

    • The high cost of AI isn't due to hardware it's due to the *large* in large language model. The LLMs achieve what they do with massive data sets. They don't work analogously to the human brain which uses only a tiny fraction of the training data (no human could ingest that amount of data)

      • Re: (Score:3)

        by znrt ( 2424692 )

        i think software improvements are more critical here. specially at scale, energy costs will for sure spur optimization of algorithms and toolchains. maybe even promote more resource-aware coding which hasn't been a priority for a long time, precisely because of hardware improvements and the fact that it tends to be cheaper and easier to throw more metal or energy at a problem than to figure out a more efficient system.

        • I don't think there's any conflict between our statements. The electricity costs for LLMs are huge. Other AI techniques are much less computationally intensive. And there will be a lot of work trying to get more out of other AI/machine learning techniques. Current LLMs are only possible as a solution because hardware has gotten so fast. Efficiency improvements will certainly come.

        • Really? History teaches us that software developers just wait for faster machines and more memory rather than optimize. It's even a religious mantra: thou shalt not prematurely optimize; and the corollary mantra: It is always premature. Just take a look at Microsoft Word - a basic word processor that barely fulfills basic needs, yet creeps along at a slow pace while consuming all resources.

    • The human brain runs on about 29 watts.

      Today, I found a text file on my computer. It was a response to a ZDNet article back in 2016 or some such. My computer remembered every word I wrote; I'd forgotten I made the post at all.

      Today, I entered my time sheet at work for the different on-site appointments I made last week. I'd forgotten one of them already. My phone kept a GPS log that knew exactly where I was and was able to ensure completeness.

      The human brain is incredible in may ways...but computers do things human brains cannot...and unless we

      • The human brain runs on about 29 watts.

        Today, I found a text file on my computer. It was a response to a ZDNet article back in 2016 or some such. My computer remembered every word I wrote; I'd forgotten I made the post at all.

        Today, I entered my time sheet at work for the different on-site appointments I made last week. I'd forgotten one of them already. My phone kept a GPS log that knew exactly where I was and was able to ensure completeness.

        The human brain is incredible in may ways...but computers do things human brains cannot...and unless we're willing to put up with the shortcomings of the human brain in our computing equipment, we'll likely need more than 27 watts to make it happen.

        Ok... so another 10 watts? After the minor miracle of replicating what a human brain does with 30 watts, you can afford to buy it a cellphone, dad.

      • The human brain is incredible in may ways...but computers do things human brains cannot...and unless we're willing to put up with the shortcomings of the human brain in our computing equipment, we'll likely need more than 27 watts to make it happen.

        NVRAM has existed for a while, is much better than human memory, and uses no power when not being accessed. I get the point you were trying to make, but the limiting things that human brains are bad at and computers are great at is raw computational speed and d

    • Task: Recite as many digits of pi as you can.

      Me: 3.14159. I know there are ways to get more.

      AI: Arbitrary number of digits subject only to some hard-coded constraint, as well as able to tell you about all the algorithms for approximating pi along with their strengths and weaknesses, then run the algorithms for you.

      Let's see 29 Watts do that.

      • Task: Recite as many digits of pi as you can.

        Me: 3.14159. I know there are ways to get more.

        AI: Arbitrary number of digits subject only to some hard-coded constraint, as well as able to tell you about all the algorithms for approximating pi along with their strengths and weaknesses, then run the algorithms for you.

        Let's see 29 Watts do that.

        My old TI calc could do that on AAA's. Not seeing how bragging about how much power you can waste calculating pi is going to win any arguments, because digital storage and processing is in your pocket, on your wrist, any AI will have access to it, they're supplemental.

        In the future, when a human-comparable AI can run on 30 watts, the 10 watt calculator it has built-in will be fucking bananas.

    • The average brain. Mine is being frugal on energy at the moment.

  • Wasteful (Score:5, Insightful)

    by TJHook3r ( 4699685 ) on Monday April 22, 2024 @12:54PM (#64414610)
    The exhortations to save trees by avoiding printing need to be updated: save coral reefs and polar bears by avoiding stupid doodles on Dall-E!
  • This seems like the ideal load for wind or solar. Just suspend processing when the power dries up and resume later
  • OK...cheap shot as I know Python is not the reason why AI projects use so much electricity, but Python and JavaScript are extremely wasteful languages....waste is fine for a quick script, but when you write major backends in either language and they require twice as many servers as a JVM, GO, or CLR alternative, perhaps we should consider how much of an ecological impact programming decisions have...especially since young people seem to have this "when all you have is a hammer, everything looks like a nail"
    • Numba and SciPy notwithstanding, "AI engineers" who don't have a clue about how things work down the stack prefer to consume resources like they were infinite and prioritize their happiness rather than consider the ethics and costs of what they're doing are the problem.

  • They're "getting into" power generation? That makes it sound like this is something brand new. I remember when Apple put in its first natural gas cogeneration plant to take its build infrastructure off the grid, back around 2002 or 2003, I think. Google has massive generators around a bunch of its buildings, presumably for the same reason. Big tech has been in the energy business quite literally for decades at this point.

  • This can all feel like a bit of spin, as these tech companies move the narrative toward their use of green energy rather than questioning whether they truly need to be consuming so much energy in the first place.

    It's a goddam opinion piece! /. can do better.

  • That's right billionaire assholes, waste energy like it has zero externalities. Google can greenwash all they want, but they're just extracting profits while socializing the costs.
  • This could be really good for the planet.

    If the hype-cycle lasts long enough, the nuclear plants might be so far along that it makes more sense to complete them than just abandon them once the fad is abandoned -- which means we can close down some coal or other fossil fuel plants.

    However, if people realize to soon that this is just another promise-don't-equal-the-hype-scam from SillyCon Valley, then the plants will be abandoned before enough investment has been made to justify continuing forward.

    Th

Slashdot Top Deals

The system was down for backups from 5am to 10am last Saturday.

Close