Semiwiki on LinkedIn Semiwiki on Facebook Semiwiki on Twitter Register / Log in 
Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/intels-guadi-3-flop.20017/
        )

    [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 Guadi 3 Flop

Getting semiconductor technology and market advice from a financial journalist with no technical background is never a good idea. An analysis of Intel's several AI programs and their chances for success is a very in-depth and complicated task. And since actual customer reception information is very hard to come by, it is nearly impossible at this point, IMO, to predict Intel's chance of success with any confidence. Certainly the prediction of less than 1% market share is just clickbait.

BTW, Gaudi (not Guadi) is apparently named after a catholic church architect named Gaudi. The next generation naming has been announced to be Falcon Shores, which follows Intel's corporate policy to code-name projects after North American geographic locations.

Falcon Shores will apparently become a merger of GPU and Gaudi architectures, whatever that means. I'm looking forward to the technical explanation.

I recommend these high-level semi-technical articles on what Intel is doing:

spectrum.ieee.org

Intel’s Gaudi 3 Goes After Nvidia

The company predicts victory over H100 in LLMs
spectrum.ieee.org spectrum.ieee.org

www.hpcwire.com

Intel Plans Falcon Shores 2 GPU Supercomputing Chip for 2026 

Intel is planning to onboard a new version of the Falcon Shores chip in 2026, which is code-named Falcon Shores 2. The new product was announced by CEO Pat Gelsinger during […]
www.hpcwire.com www.hpcwire.com

It is likely that Intel's most significant AI program so far, business-wise, is the NPU in their newest client CPUs, especially since they're partnering with Microsoft for the applications in Windows.
 
Gaudi is not a "catholic church architect", even if he has built a cathedral in Barcelona. He was an influent designer of the 1900's, building houses you would dream to live, designing crazy parks your kids will love to play. If you have a chance to go to Barcelona, visit the "Sagrada Familia" (the cathedral) and walk in certain streets to discover the buildings (designed by Gaudi)...
 
Eric Esteve said:
Gaudi is not a "catholic church architect", even if he has built a cathedral in Barcelona. He was an influent designer of the 1900's, building houses you would dream to live, designing crazy parks your kids will love to play. If you have a chance to go to Barcelona, visit the "Sagrada Familia" (the cathedral) and walk in certain streets to discover the buildings (designed by Gaudi)...
Click to expand...
I don't know much about Gaudi or care. Shortly after Habana Labs was acquired by Intel, I messaged a friend of mine in Intel and asked the origin of the "Gaudi" name. It was a subtle joke between us, because Intel's Israeli teams sometimes chose code names which annoyed customers or partners in some other countries, which drove the rigid code-naming rule about places in North America. I was surprised the name wasn't changed after the acquisition. The church architect explanation was what I got from him. I don't know how accurate the answer was, which is why I said "apparently". I did a simple search on the name after I got his answer, and it included the nickname "God's architect", which made me chuckle, and that was enough for me.
 
blueone said:
It was a subtle joke between us, because Intel's Israeli teams sometimes chose code names which annoyed customers or partners in some other countries, which drove the rigid code-naming rule about places in North America.
Click to expand...
Could you give some examples? My understanding is that everything was just serial numbers for a while (4004, 8088, 286, 486, etc) before moving to things in Oregon (Tualatin, Willamette, etc), and finally I assume the current lakes and rapids aren't all in Oregon.
 
nghanayem said:
Could you give some examples? My understanding is that everything was just serial numbers for a while (4004, 8088, 286, 486, etc) before moving to things in Oregon (Tualatin, Willamette, etc), and finally I assume the current lakes and rapids aren't all in Oregon.
Click to expand...
I would rather not give specific examples of the controversial names, because I like to keep away from politics on this site. The examples you listed of microprocessors which had numeric designations were product names in the marketplace, not project code-names. The initial use of place names in Oregon was a coincidence; the rule was (and I think still is) places in North America, but many Intel people were and are in Oregon sites (like Jones Farm) and chose familiar place names in that state. I chose the code-names for a couple of projects, and as I remember the process included a submission to Intel Legal, which approved or denied the names.
 
Lets focus on Gaudi and not codenames (I can give a history of code names and "problematic" ones (not all Israeli) over a beer.

Why is Intel struggling to get any traction in AI hardware? Intel has been working on this for years. What is holding them back
 
MKWVentures said:
Why is Intel struggling to get any traction in AI hardware? Intel has been working on this for years. What is holding them back
Click to expand...
Integrated graphics in client CPUs debased the business case for discrete client GPUs, and Nvidia and Radeon had customer traction, name recognition, and software support. The likely unit volumes for Intel discrete client GPUs at the time (15-20 years ago) was almost certainly not interesting to the fabs either, even on the N-1 processes.

And then there was Larrabee, which tried to extend CPU architecture to GPU workloads. I diverge from Gelsinger's opinion on this project; it was a bad idea. Trying to compete with intrinsically parallel state machine logic with CPU cores, software, and hardware assists (like wide vector units) puts you a few steps behind the starting line of the performance race. Yeah, software could enable new stuff, like ray-tracing support, faster, but then once the competition has it in hardware you're not competitive.

I think that once Intel got so far behind the Nvidia and Radeon product pipelines it would have taken a technical miracle to attain leadership. No miracles were apparently forthcoming.
 
MKWVentures said:
Why is Intel struggling to get any traction in AI hardware? Intel has been working on this for years. What is holding them back
Click to expand...
My view is that Intel has been struggling with 4 concurrent challenges:
1) Historically, the traditional Innovators Dilemma. The same force that enabled Intel (and to some degree AMD) to replace IBM mainframes and Sun servers in the server space with x86, thwarted their success in smartphones and discrete GPU/ GPU-based AI processors. Intel and AMD did remove the low end of the discrete GPU market with integrated GPUs, but because of the competition, they weren't able to monetize very well, and likely accelerated both NVIDIA's and AMD's push into higher-end GPU applications and eventually AI apps.
2) Product focus - Intel seems to be trying to win in the GPU visualization, AI and HPC markets with a broad array of organic and acquired solutions (x86 + SIMD math engines, FPGAs, GPUs, Gaudi, and Habana). Major undertaking just to bring those together under OneAPI, but that's just the beginning of providing a sufficient platform compared to NVIDIA and AMD.
3) Market structure - Lots of competitors in every segment with lots of different strategies. Pat made a lot of noise a couple days ago about pushing IT departments to deploy AI PCs, yet there are many sources for AI client computers - Apple (both Mac and iOS), AMD, Android phones via Qualcomm, etc. Apple, Google and Microsoft are the ones most likely to deliver killer-apps on the client side that drive businesses and consumers to upgrade. And on the GPU => AI Processor path for the data center, Intel is battling NVIDIA, AMD, the big CSPs like Amazon, Microsoft and Google that are building some of their own chips, plus new entrants like Groq and Cerebras. And foolishly, Intel has also chose to battle it out in the gaming and visualization space where AMD and NVIDIA are highly entrenched - Intel is till battling up the driver curve with the ARC line and being forced to live off thin or no margins.
4) Embattled cash cow products - Intel is battling to grow against entrenched players, while protecting its embattled cash cows from AMD and TSMC growth, and shrinking data center $$ for general purpose servers. More challenging for them to spend the way they need to on new products and product ecosystems while they are also investing in foundry capacity and protecting existing money-making products.
 
Last edited:
KevinK said:
2) Product focus - Intel seems to be trying to win in the GPU visualization, AI and HPC markets with a broad array of organic and acquired solutions (x86 + SIMD math engines, FPGAs, GPUs, Gaudi, and Habana). Major undertaking just to bring those together under OneAPI, but that's just the beginning of providing a sufficient platform compared to NVIDIA and AMD.
Click to expand...
In one area, OpenVINO, which is part of the visualization and inference strategy, I'm confused. If OpenVINO is "open", why does it only work with Intel hardware? Well, OpenVINO now includes an Arm CPU plug-in, but it's not available in Intel's distribution. I'm especially fascinated by this quote from OpenVINO web pages:

ARM® CPU plugin is a community-level add-on to OpenVINO™. Intel® welcomes community participation in the OpenVINO™ ecosystem, technical questions and code contributions on community forums. However, this component has not undergone full release validation or qualification from Intel®, hence no official support is offered.
Click to expand...

I admit, I don't understand where Intel is going with their AI software enabling strategy. Since the AI accelerators in Xeons are proprietary too, I'm not understanding the "Open" part of VINO at all.
 
blueone said:
In one area, OpenVINO, which is part of the visualization and inference strategy, I'm confused. If OpenVINO is "open", why does it only work with Intel hardware? Well, OpenVINO now includes an Arm CPU plug-in, but it's not available in Intel's distribution. I'm especially fascinated by this quote from OpenVINO web pages:
Click to expand...

The OpenVino “support” chart below kind of highlights the challenge #2 I mentioned, as well as raising the question you highlight, “is it really a standard outside of Intel ?” It brings together 8 1/2 different hardware architectures, 8 of which are all Intel inference architectures. And where is Gaudi ? I guess it’s not part of the party because Gaudi is focused on training and inference ? training.
 

Attachments

  • IMG_0233.jpeg
    IMG_0233.jpeg
    431 KB · Views: 28
KevinK said:
And where is Gaudi ? I guess it’s not part of the party because Gaudi is focused on training and inference ? training.
Click to expand...
Ha! Good point. I missed that Gaudi wasn't in the supported list.

One of the references I was reading yesterday did mention Gaudi for training on medical imaging apps, and then OpenVINO post-training for image inferencing, and I just made the (incorrect) leap that OpenVINO supported Gaudi. Silly me!

https://blog.openvino.ai/blog-posts/ai-workflow-habana-training-openvino-quantization-inference
 
Last edited:
One other interesting dynamic I heard yesterday is that 90% of the training CPU cycles and hardware for generative AI is concentrated in 9 companies. Seems like it would be very hard to get a foothold in these 9 at any kind of scale, especially with several already developing their own chips as well. At the same time, there might be strategic openings. It is rumored that FaceBook bought up a huge chunk of NVIDIA’s supply to speed delivery of Llama 3 while denying OpenAI hardware to delay GPT 5. So OpenAI might be a great opportunity for Intel if Gaudi can really deliver the goods. BTW - just heard an interview with Dario Amodei of Anthropic, that gave some relative costs. In his mind the current generation of generative AI models in use cost $100M each to train, within 2-3x, and the cost to train the next generation models in training today, has moved to $1B.

 
You must log in or register to reply here.
Back
Top