Intel Says New Gaudi 3 AI Chips Top Nvidia H100s in Speed and Cost 32
Intel on Tuesday unveiled its new "Gaudi 3" AI chip that the company claims is over twice as power-efficient and can run AI models one-and-a-half times faster than Nvidia's H100 GPU. "It also comes in different configurations like a bundle of eight Gaudi 3 chips on one motherboard or a card that can slot into existing systems," adds CNBC. From the report: Intel tested the chip on models like Meta's open-source Llama and the Abu Dhabi-backed Falcon. It said Gaudi 3 can help train or deploy models, including Stable Diffusion or OpenAI's Whisper model for speech recognition. Intel says its chips use less power than Nvidia's. Intel said that the new Gaudi 3 chips would be available to customers in the third quarter, and companies including Dell, Hewlett Packard Enterprise, and Supermicro will build systems with the chips. Intel didn't provide a price range for Gaudi 3.
Gaudi 3 is built on a five nanometer process, a relatively recent manufacturing technique, suggesting that the company is using an outside foundry to manufacture the chips. In addition to designing Gaudi 3, Intel also plans to manufacture AI chips, potentially for outside companies, at a new Ohio factory expected to open in 2027 or 2028, CEO Patrick Gelsinger told reporters last month. "We do expect it to be highly competitive" with Nvidia's latest chips, said Das Kamhout, vice president of Xeon software at Intel, on a call with reporters. "From our competitive pricing, our distinctive open integrated network on chip, we're using industry-standard Ethernet. We believe it's a strong offering."
Gaudi 3 is built on a five nanometer process, a relatively recent manufacturing technique, suggesting that the company is using an outside foundry to manufacture the chips. In addition to designing Gaudi 3, Intel also plans to manufacture AI chips, potentially for outside companies, at a new Ohio factory expected to open in 2027 or 2028, CEO Patrick Gelsinger told reporters last month. "We do expect it to be highly competitive" with Nvidia's latest chips, said Das Kamhout, vice president of Xeon software at Intel, on a call with reporters. "From our competitive pricing, our distinctive open integrated network on chip, we're using industry-standard Ethernet. We believe it's a strong offering."
Marketing might want to rethink that name ... (Score:1)
... because Gaudi sounds WAY too close to gaudy. /s
Re:Marketing might want to rethink that name ... (Score:4, Interesting)
Re: (Score:1, Redundant)
Gaudi *was* an artsy type... https://barcelonanavigator.com... [barcelonanavigator.com]
Re: (Score:2)
>> close to gaudy
It is somewhat memorable though, so maybe that's intentional.
Re: (Score:2)
It's a reference to the paintor Antonio Gaudí [wikipedia.org]
Re: (Score:2)
Crap, not painter, architect and designer. I was thinking of Antonio Caba.
1.5x faster, not available for another 4 years (Score:1)
Basically they got a decent prototype but they are still 3 years behind nVIDIA. It will be great to see some competition but like their previous attempts it is always too late and doesnâ(TM)t scale. At least they are better at providing support than AMD but the nVIDIA ecosystem is all around better at this point. Hopefully people learn that open source is where it is at and nVIDIA seems to understand that.
Re: (Score:3)
Re: (Score:2)
What isn't in the summary is how many Gaudi 3 products you can obtain in 2024. If current product is only available as part of a limited run, then it may be several months or years before Gaudi 3 could be sold in significant volume.
Also the question still remains, why would anyone be interested in a 2024 product that competed with aging h100 in 2027? Why would they still be producing Gaudi 3 by that point?
Re: 1.5x faster, not available for another 4 years (Score:2)
Re: (Score:1)
It will be made available to partners for integration by next quarter. Basically you'll get some engineering samples next quarter. You need to set up a fab for general availability.
Re:1.5x faster, not available for another 4 years (Score:4, Insightful)
Correction: Intel's tech is launching later this year. 5-6 months, not 4 years. 4 years is the time it will take for their new, dedicated facility to be constructed, but they are using 3rd party fabs to manufacture Gaudi 3 this year.
As for "Nvidia has won everything and there's no point in trying" - this is a nascent market where a 3-year lead isn't much. A real software lead is Windows. Excel. Photoshop. Android. These are real standards. Nvidia's tech isn't a standard yet, because everyone is trying to compete with it. You know you're the standard when everyone else adopts you (see Tesla EV charging port as an example; everyone else adopted it). Right now you have at least Amazon, Microsoft, Meta, and Alphabet all working on their own AI chips, in addition to AMD and Intel. So Nvidia just has a head start, but the race isn't over and Nvidia can't even supply all the demand. All you have to do right now is be good-enough and you get deals because you'll be far cheaper than the best. That's dangerous for Nvidia because it means their capacity limits are allowing the existence of competition. This is only the start of the race, not the end.
Re: 1.5x faster, not available for another 4 years (Score:2)
Re: (Score:2)
Re: (Score:2)
The home-grown stupidity is right up there. Jimmy Kimmel had someone out in the street asking random people what was the recent eclipse. The answers were so bizarre that I wonder how these people haven't taken themselves out of the gene pool from sheer stupidity.
Re: (Score:2)
What, not Cthulhu paying a visit? I am shocked! Shocked, I tell you!
Well, maybe one reason so many people mistakenly believe AI is competent at things is just because they are utterly incompetent at most things, obviously including estimating the capability of AI.
I smell a catch, (Score:2)
...Intel is a known spinner, so where's the catch, it's in the haystack there somewhere...
5nm (Score:1, Interesting)
Re: (Score:3)
If they're lying in the press release I'm not going to bother looking into it. Nothing humans manufacture in bulk is 5nm.
Ignorance is bliss I guess, you could have looked up and found multiple foundries use 5nm processes and in fact some are now using 3nm. Note this doesn't mean the product produced is 5nm, but by the sounds of your ignorance I doubt you will care to learn anyway.
Re: (Score:1)
Re: (Score:3)
Re: (Score:2)
He doesn't appear to be confused to me. He appears to be pointing out that the industry standard nomenclature is marketing bullshit.
Re: (Score:3)
Nothing humans manufacture in bulk is 5nm.
Correct. If you exclude every iPhone for the past 3 years, the Ryzen 7000 platform from AMD, the 14th gen Intel chips, NVIDIA's RTX40XX series GPUs, and AMD RX7000 series GPUs you would be perfectly spot on the money.
Re: (Score:1)
Re: (Score:2)
No one has ever claimed they have feature sizes below 25nm. They said they are produced on a 5nm node. I actually think you fundamentally don't understand what a process node is. Hint: It hasn't had anything to do with feature size for well over a decade. If you're claiming that all production of chips with identical feature sizes are the same well that's just as absurd as calling an EV the same as an ICE car simply because they have the same overall length.
Yes, but how fast can it (Score:1)
Make bitcoins?
I need to fund my NFT collection of dark quantum energy apes.
Congrats Intel (Score:1)
You're only 2 years behind NVIDIA (H100 is from 2022). This year's GB200 is a beast compared to the H100 and GH100 of yore. As architecture is more than just one chip, the fabric you used to connect them matters. An example of that is in the GB200 NVL72 system, which has 36 GB200's (each is effectively 2 B200's on a chip), connected with NVlink as the fabric. (NVlink has significantly lower latency than PCIe)
Re: (Score:3)
Facts. Nvidia at this point isn't just selling chips. They are selling full feature solutions that you put directly onto your data center floor. I'm not seeing such vertical integration offerings from Intel or AMD. Heck, Nvidia even have the software solution that will help you plan your data center build-out with omniverse that you can use and plan just how their solution will fit in your environment, or, if you want to build one from scratch.
Nvidia is playing on another level.
Re: (Score:2)
You should probably save the congratulations for the eventual market winner. Being 2 years ahead, but unable to produce enough hardware to supply the market, leaving the market and enough revenue open for at least 1 other competitor, is not a position to brag about. And then the competitors include at least 4 of the Mag 7, plus AMD, plus Intel. Plus Qualcomm. To say they are well capitalized is to say Apple has a few nickels of cash in the bank. Nvidia was 1st into the AI market but now finds itself with un