Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/intels-latest-ai-chip-is-a-direct-shot-at-nvidias-moneymaker.20000/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Intel's latest AI chip is a direct shot at Nvidia's moneymaker

Daniel Nenni

Admin
Staff member
Intel on Tuesday debuted a new AI chip designed to take on one of Nvidia’s most popular AI processors as part of its Intel Vision event.

The company's Gaudi 3 chip matches and exceeds Nvidia’s H100 AI processor when it comes to training and deploying generative AI models. The H100 is one of Nvidia's most popular chips, powering artificial intelligence applications for some of the world's biggest tech companies, from Microsoft to Google.

According to Intel, the Gaudi 3 is 40% more power efficient with 50% faster inferencing than Nvidia’s H100. Training refers to the time it takes to train an AI foundation model like OpenAI’s GPT-3, while inferencing is the process of actually using the models for apps.

Intel says the Gaudi 3 is up to 1.7 times faster when training common large language models than the H100. Intel even says its chip is an average of 1.3 times faster than Nvidia’s beefier Nvidia H200 when it comes to inferencing certain language models.

In addition to the Gaudi line of chips and Xeon CPUs, Intel is also marketing so-called AI PCs, computers equipped with neural processing units that can handle onboard AI tasks.

These AI maneuvers come at a critical time for Intel.

The company is working to reinvent itself as the leader in sophisticated chip technology while building out its chip manufacturing capabilities in the US and abroad. This includes a foundry business that will build Intel’s own chips as well as those for third-party customers. The company has already confirmed it will build chips for Microsoft.

But this reinvention is coming at a steep cost, with Intel announcing earlier this month a $7 billion operating loss for 2023 as part of its new financial reporting structure. Intel stock is down 25% so far this year. Rivals Nvidia and AMD, conversely, are up 15% and 74% so far this year, respectively.

Intel has been teasing the Gaudi 3 chip for some time, with CEO Pat Gelsinger showing off a version of the platform during a December press event. It’s seen as Intel’s best chance yet to compete with Nvidia in the AI market, a space the graphics chip giant dominates thanks to its powerful hardware and software.

Intel CEO Pat Gelsinger discusses his company's plans to capitalize on the booming demand for the chips needed to power artificial intelligence, Wednesday, Feb. 21, 2024 during a conference in San Jose, Calif. (AP Photo/Michael Liedtke)

Intel CEO Pat Gelsinger discusses his company's plans to capitalize on the booming demand for the chips needed to power artificial intelligence during a conference in San Jose, Calif., Feb. 21, 2024. (Michael Liedtke/AP Photo) (ASSOCIATED PRESS)

The company's new chips also come just a month after Nvidia debuted its own next-generation AI platform called Blackwell. That system includes Nvidia’s B200 AI chip, which is made up of two Blackwell chips mated together to act as a single processor.

Intel is putting its AI systems forward as an alternative to Nvidia’s offerings, claiming an open-source approach that will allow customers to use the services and software that they want.

And while Nvidia is the leader in overall AI technology, Intel is betting enterprises will remain wary of sticking with a single source for high-value needs like AI hardware.

For instance, hyperscalers like Microsoft, Google, Amazon, Meta, and others already either offer their own AI chips or are developing them in addition to using Nvidia’s chips.

Intel’s Gaudi 3 will give those hyperscalers another option in the chip market.

Intel also announced on Tuesday a new form of Ethernet connectivity for Gaudi 3 nodes, competing with Nvidia’s InfiniBand connectivity technology to ensure uninterrupted connections and processing. The company also debuted its Xeon 6 processors for use in AI systems.

Taking a page out of Nvidia’s book, Intel is offering the entire setup as a reference design that partners like Dell, HP, Lenovo, and Super Micro can use to build their own server cabinets.

Intel says customers will be able to order the Gaudi 3-powered systems in sizes ranging from individual nodes and small boxes to enormous megaclusters that can fill entire data centers.

 
Even Intel's chances of success are slim, it would be very stupid for Nvidia to place their order with IFS.
IFS division has to be pulling their hair out over these info graphics that specifically call out NVIDIA 😵‍💫. Product division marketing and sales didnt get the memo that IFS is supposed to save the entire company. I've always known Jensen to take slights (perceived or otherwise personally). Between this and Pat explicitly sh*t talking by saying cuda's dominance was sheer luck. I don't know why nvidia would use IFS? It doesn't make any sense.
 
If NVidia was thinking about using IFS before, they are likely rethinking it now.

Intel seems to take frequent pot shots at companies that they need to be building relationships with, like TSMC and now Nvidia.

Hopefully the marketing people who were involved in this were part of the 10,000 let go.
 
Jensen is shrewd -- he's going to take the best fab deal with the most advanced technology that he can get.

He fell out with TSMC but had no problem going back to them when Samsung failed to keep up with his needs.
 
Jensen is shrewd -- he's going to take the best fab deal with the most advanced technology that he can get.
He fell out with TSMC but had no problem going back to them when Samsung failed to keep up with his needs.

Nvidia is a mutli fab sourcing company, they have been for a while. Samsung was not able to deliver enough good die at 5/4nm so Nvidia is exclusive to TSMC now but I doubt it will stay that way. Unless packaging makes a difference. Engineers I know at Nvidia did not like Samsung but love TSMC, they did not always get what they wanted. From what I have heard lately the TSMC N2 PDK is already better than Intel 18A. It is all about the PDK. The TSMC Symposium is two weeks away. Should be a good one.
 
Nvidia is a mutli fab sourcing company, they have been for a while. Samsung was not able to deliver enough good die at 5/4nm so Nvidia is exclusive to TSMC now but I doubt it will stay that way. Unless packaging makes a difference. Engineers I know at Nvidia did not like Samsung but love TSMC, they did not always get what they wanted. From what I have heard lately the TSMC N2 PDK is already better than Intel 18A. It is all about the PDK. The TSMC Symposium is two weeks away. Should be a good one.

An interesting thing about the Intel Gaudi 3 is that it is manufactured by TSMC using N5 process.

Without considering other factors, it provides a great opportunity to compare it with Nvidia H100 using TSMC's N4 process.
 
Last edited:
An interesting thing about the Intel Gaudi 3 is that it is manufactured by TSMC using N5 process.

Without considering other factors, it provides a great opportunity to compare it with Nvidia H100 using TSMC's N4 process.
Out of curiosity, what would you be comparing? They are two completely different microarchitectures. The Gaudi 3 is not a GPU.
 
They both use TSMC N5/N4 class process and both are targeting AI market. We can compare them in things such as:

1. Design and strategy
2. Features and functionality
3. Price and performance
4. Availability
5. Totally cost of ownership
6. Software, hardware, and ecosystem support and collaboration
7. Target markets, users, and applications
And more...

It will be very interesting to see how Intel and Nvidia utililze TSMC's capabilities to grab the huge market opportunity of AI.

Another thought about it is that "if" Intel Gaudi 3 can't compete against Nvidia's even both are manufactured with a similar TSMC manufacturing environment, we have to say Intel's problem is less on manufacturing and more on product design, timing, price and performance or something else.
 
Last edited:
Ya'll may flame me for this, and I have no inside info, but to me strategy has become clear, especially baiting Nvidia and going after their customers aggressively.

1) Divide intel product biz from foundry to create safe space for product competitors to engage IFS (despite well-known intel product competition in this area)
2) Create competitive product that begins to win some biz AND pisses off Jensen (it's product biz, not foundry).
3) Kill Falcon Shores and axe the segment in exchange for Nvidia commitment to multi-gen supply agreement. (and/or license the IP exclusively to them to do what they want). All is possible as long as intel still owns both segments.

Brilliant! Pat is no idiot! Everything they are doing is expanding their options to win business in the long game.
 
Here is the 1 hour 37 minutes keynote for the Intel Vision 2024 conference:


Or, if you prefer a shorter 21 minutes version:



Or, if you really don't have time to watch either one, here is the one-minute edition:

 
Last edited:
My mistake, the press release I looked at where 10,000 was cited was from Intel's 2023 layoff round. Seems like the layoffs at Intel are more or less constant.
Intel needs to lean up, IMO. As of the end of 2023, Intel had nearly as many employees as TSMC, AMD, and Nvidia combined. That's ridiculous.
 
Even Intel's chances of success are slim, it would be very stupid for Nvidia to place their order with IFS.
But the truth is it's better to make Intel Products so strong so it can attrach a wave of customers (not necessarily Nvidia, but Google, Amazon, or Microsoft) away from TSMC. I would rather hurt Nvidia so badly, so they count me on and know what I can do without them.
 
IFS division has to be pulling their hair out over these info graphics that specifically call out NVIDIA 😵‍💫. Product division marketing and sales didnt get the memo that IFS is supposed to save the entire company. I've always known Jensen to take slights (perceived or otherwise personally). Between this and Pat explicitly sh*t talking by saying cuda's dominance was sheer luck. I don't know why nvidia would use IFS? It doesn't make any sense.
Even if Nvidia want to, Intel don't have much capacity they can offer that compete with scale of TSMC.
 
If NVidia was thinking about using IFS before, they are likely rethinking it now.

Intel seems to take frequent pot shots at companies that they need to be building relationships with, like TSMC and now Nvidia.

Hopefully the marketing people who were involved in this were part of the 10,000 let go.
This mentality is so sick. It's like turn a blind eye to Samsung to Qualcomm, Qualcomm to Apple, Google to Apple, Microsoft to Apple, Arm to Intel, Arm to Qualcomm, and many others.

They can be supplier, collaborator, or frenemy. But there's no rule in this world that stated if you are an enemy of mine, then I'm not going to be friend of you.
 
Nvidia is a mutli fab sourcing company, they have been for a while. Samsung was not able to deliver enough good die at 5/4nm so Nvidia is exclusive to TSMC now but I doubt it will stay that way. Unless packaging makes a difference. Engineers I know at Nvidia did not like Samsung but love TSMC, they did not always get what they wanted. From what I have heard lately the TSMC N2 PDK is already better than Intel 18A. It is all about the PDK. The TSMC Symposium is two weeks away. Should be a good one.
It should be, it's going to be surprising if TSMC, the kind of PDK, lose to Intel, who just get started with foundry
 
Here is the 1 hour 37 minutes keynote for the Intel Vision 2024 conference:


Or, if you prefer a shorter 21 minutes version:



Or, if you really don't have time to watch either one, here is the one-minute edition:


We were briefed on it. Pat went off script a bit:

 
3) Kill Falcon Shores and axe the segment in exchange for Nvidia commitment to multi-gen supply agreement. (and/or license the IP exclusively to them to do what they want). All is possible as long as intel still owns both segments.

My take on this is they probably looked at their limited engineering pool and decided to use it for getting the next gen product after Falcon Shores out in as good of shape as possible.

They’ve made a lot of similar decisions lately - Extending Raptor Lake an extra year because it’s “good enough” while they focus on 4, 3, 20A processes and related products. Or delaying Battlemage to work on software or use other TSMC capacity for other products, etc.

Despite what the market is showing, these are still early days in the AI/ML era.
 
Back
Top