Jensen Huang: AI Has To Do '100 Times More' Computation Now Than When ChatGPT Was Released 32
In an interview with CNBC's Jon Fortt on Wednesday, Nvidia CEO Jensen Huang said next-gen AI will need 100 times more compute than older models as a result of new reasoning approaches that think "about how best to answer" questions step by step. From a report: "The amount of computation necessary to do that reasoning process is 100 times more than what we used to do," Huang told CNBC's Jon Fortt in an interview on Wednesday following the chipmaker's fourth-quarter earnings report. He cited models including DeepSeek's R1, OpenAI's GPT-4 and xAI's Grok 3 as models that use a reasoning process.
Huang pushed back on that idea in the interview on Wednesday, saying DeepSeek popularized reasoning models that will need more chips. "DeepSeek was fantastic," Huang said. "It was fantastic because it open sourced a reasoning model that's absolutely world class." Huang said that company's percentage of revenue in China has fallen by about half due to the export restrictions, adding that there are other competitive pressures in the country, including from Huawei.
Developers will likely search for ways around export controls through software, whether it be for a supercomputer, a personal computer, a phone or a game console, Huang said. "Ultimately, software finds a way," he said. "You ultimately make that software work on whatever system that you're targeting, and you create great software." Huang said that Nvidia's GB200, which is sold in the United States, can generate AI content 60 times faster than the versions of the company's chips that it sells to China under export controls.
Huang pushed back on that idea in the interview on Wednesday, saying DeepSeek popularized reasoning models that will need more chips. "DeepSeek was fantastic," Huang said. "It was fantastic because it open sourced a reasoning model that's absolutely world class." Huang said that company's percentage of revenue in China has fallen by about half due to the export restrictions, adding that there are other competitive pressures in the country, including from Huawei.
Developers will likely search for ways around export controls through software, whether it be for a supercomputer, a personal computer, a phone or a game console, Huang said. "Ultimately, software finds a way," he said. "You ultimately make that software work on whatever system that you're targeting, and you create great software." Huang said that Nvidia's GB200, which is sold in the United States, can generate AI content 60 times faster than the versions of the company's chips that it sells to China under export controls.
Why does it? (Score:3, Interesting)
I've ceased to give a damn about climate change when AI's carbon footprint dwarfs any personal action I might take to consume less, fly less or recycle more.
Meanwhile, I'm earning less in real terms than 15 years ago and thankful for any work I can get before the oligarchs in the white house deliberately crash the global economy sending me broke and unemployable in my 50s.
So why should anyone care about our AI overlords?
Re: Why does it? (Score:3)
Re: (Score:1)
I've ceased to give a damn about climate change when AI's carbon footprint dwarfs any personal action I might take to consume less, fly less or recycle more.
Meanwhile, I'm earning less in real terms than 15 years ago and thankful for any work I can get before the oligarchs in the white house deliberately crash the global economy sending me broke and unemployable in my 50s.
So why should anyone care about our AI overlords?
Why does it?
What is climate change, what might it mean to someone that has stopped caring?
What are common concerns about climate change? Which concerns might a person stop caring about?
What is AI's carbon footprint? How is that calculated, what is the range of estimates?
What are personal actions you can take to reduce carbon footprint? Is flying less or recycling more an acceptable characterization?
Does real terms mean inflation adjusted?
Is it worthwhile to search for three decades of inflation adjusted ind
Can we have a "No AI promotion day"? (Score:2)
Like, every Thursday /. doesn't post any crap from used car salesman types about "AI"?
Please?
Re: Can we have a "No AI promotion day"? (Score:2)
What would you do on Thursday ?!
I've actually wonder if the editors have been replaced by AI... like, how would we know?
Re: (Score:2)
No way to know, Turing was just joking with that test.
Re: (Score:2)
I've actually wonder if the editors have been replaced by AI... like, how would we know?
If the editors were replaced by AI, there would be no dupes.
More hardware required... (Score:5, Interesting)
... says man who sells hardware.
Yeah, ok Jensen, whatever you say.
The whole fuss around DeepSeek is it used far LESS hardware resources than current models so god knows where this guy got the 100x extra from other than out of his backside.
Re: More hardware required... (Score:1)
Re: (Score:2)
Re: (Score:3)
But a lot less, when the entire purpose of the comment was to allege how much more would be needed.
And NO AI requires the use of NVidia hardware. Funny how the world's best intelligence uses no GPU acceleration or Nvidia hardware at all.
Re: (Score:2)
yeah and that means everyone will run them in parallel, or think they can build them themselves for custom purposes. They probably can't, because their idea is fundamentally unworkable magic... but they can still spend the money trying.
There's currently a 1 year lead on H200s, incidentally.
Re: (Score:3)
And I think that still his point holds. When chatGPT released, it used a GPT3 architecture, that's 175 billions parameter. DeepSeek R1 is 671 billion parameters, which is about 4 times more.
But also the (so-called) reasoning models are running the query through the model multiple times in order to create that reasoning ability. It is not uncommon to see it run the query 50 times.
So even though DeepSeek has half the parameters of GPT4, using it still probably uses 100 times more flops than GPT3 did.
Re: More hardware required... (Score:2)
Re: (Score:1)
I think he's referring to the CPU usage when running the DeepSeek R1 model. It uses a ton more CPU (not GPU) than other models. So if you have a crappy CPU, on older models it didn't matter. With DeepSeekR1 it does. I don't know if that means Nvidia's going to try to offload that CPU usage or what, but that's how I read his statement.
Re: (Score:2)
... says man who sells hardware.
Yeah, ok Jensen, whatever you say.
The whole fuss around DeepSeek is it used far LESS hardware resources than current models so god knows where this guy got the 100x extra from other than out of his backside.
You're talking about V3, and Jensen's talking about R1.
The more you buy the more you save (Score:2)
AI (Score:5, Insightful)
"Throw more processing at it, that'll make it intelligent!"
The cry of every AI researcher since the 60's.
And now we're training it on the entire human Internet, now we're using specifically designed stupendously parallel hardware in datacentre-volumes spread across the globe, now we're pulling huge amounts of global energy, now we're using so much RAM, GPUs and storage just to do so that it's creating hardware shortages, now we're training it for years at a time on the new input of the human race live and in real-time...
Is it intelligent? No.
It's almost like there's something MORE to intelligence than brute-force and complexity, eh? Maybe start researching that. Because as far as I'm concerned we've hit the absolute limit of current technology on a global scale that can/should be deployed to this one task and what does it do? A poor Google combine with some waffle, while lying often, making stuff up and being inherently unreliable, and the only way to "control" it is to ask it nicely because you feed it the user's question.
It's utterly ridiculous and we need to lay this to bed before we burn yet-more-gigawatts trying to make it somehow magically pop into intelligence, and go back and THINK. There's a concept. Think. Of something new to try. Because this... ain't intelligent and it ain't working. We've made a very, very expensive Google-Alexa hybrid that shows no sign of intelligence at all. With the entirety of the world's technological resources for the past... decade?
Give it up and actually sit and think and do research.
Re: (Score:2)
"And now we're training it on the entire human Internet", this is the worrying part, aside from the energy requirement and what they will do to the climate.
The "humans" have not exactly excelled at running Earth lately. So AI is likely to draw all the wrong conclusions and feed them back to humans giving us a Moronic Feedback Loop.
Re: (Score:3)
"Throw more processing at it, that'll make it intelligent!"
The cry of every AI researcher since the 60's.
I'm not sure it is accurate to say that is what the ML (sure it is being called AI, but I feel like the field is still more accurately Machine Learning, or statistical filtering) researchers are saying, but maybe the few trying to publicize their efforts.
Also I don't think LLMs are so popular because they are the most accurate ML models, but because they are the easiest to train. Most of the best ML models are as accurate as the labelled training data they are provided. Sure there are some ML models that
Re: (Score:2)
Is it intelligent? No.
Is it intelligent? Depending on the task and use case, absolutely yes.
Is it AGI? No.
Re: (Score:2)
Re: (Score:2)
If you're looking for an absurdly reductionist example of intelligence I'd go with a thermostat, because clocks don't respond to external stimulus.
Re: AI (Score:2)
ONLY 100 TIMES!! (Score:1)
Two orders of magnitude based on less than half a decade of study or experiene or science or stats.
Sure, HEAD OF NVIDIA, the DEMAND for YOUR PRODUCT is going to go up not 3 fold... not 5 fold... not 10 fold... but 100 FOLD.
Disclaimer: I don't have stok in Nvidia and never will.
E
not the flex he thinks it is (Score:5, Insightful)
"The amount of computation necessary to do that reasoning process is 100 times more than what we used to do..."
He intends this as a brag when in reality it is an admission that AI developers don't know what they're doing.
And it's not a reasoning process, it's a "reasoning" process. It's just a name, there's no reasoning being done.
Jensen Huang may know plenty about GPUs and graphics acceleration, but every time he opens his mouth it is clearly apparent he doesn't know anything about AI. He's a right-place, right-time poser, not an AI expert.
Re: (Score:2)
I guess we will find out just how big a relational database can get before it collapses - at that point AI will have to actually start over to be more like - actual AI.
Orly? (Score:2)
AI Competing with humans for natural resources (Score:2)