You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Are we moving towards a future where Big tech design their own customized chips to suit their particular requirements, outsourcing manufacturing to foundries and bypassing Intel and AMD?
A vertical integration between Annapurna Labs and Amazon Web Services is paying off
spectrum.ieee.org
Are we moving towards a future where Big tech design their own customized chips to suit their particular requirements, outsourcing manufacturing to foundries and bypassing Intel and AMD?
All the hyperscalers will be here in the next few years. Supporting several design teams and the full developing including design, TO, validation in the hundreds of millions of dollars more than justifies the outlay and they all can afford it for top to bottom integration similar to what Apple has done. TSMc is sitting pretty here and IFS dreams of also becoming a supplier with 18A and 14A. We will see in 2025 if Intel delivers
You're late too. Microsoft is already previewing the Cobalt100 CPU. Google is about to deploy the Axion CPU in Google Cloud. Meta is currently only designing AI chips, but they're looking for a partner to help them design a custom Meta CPU chip. My baseless guess is Broadcom becomes the partner.
You're late too. Microsoft is already previewing the Cobalt100 CPU. Google is about to deploy the Axion CPU in Google Cloud. Meta is currently only designing AI chips, but they're looking for a partner to help them design a custom Meta CPU chip. My baseless guess is Broadcom becomes the partner.
I wonder what really happened to the AI acquisitions Intel made a few years ago. Intel spent $2B on Habana Labs in 2019? They also acquired Nirvana in 2016 for $400M. These companies were packed with very intelligent people! The expectation was for Intel to be an AI powerhouse. So many missed opportunities. How is the Habana Labs Gaudi chip doing? It was supposed to be the Nvidia killer. I don't think the Nirvana chip ever made it out?
Intel to Produce Custom AI Fabric Chip on Intel 18A and Custom Xeon 6 Chip on Intel 3 for AWS; Multi-Year, Multi-Billion-Dollar Collaboration Accelerates Development of Chip Manufacturing in Ohio
Intel to Produce Custom Al Fabric Chip on Intel 18A and Custom Xeon 6 Chip on Intel 3 for AWS; Multi-Year, Multi-Billion-Dollar Collaboration Accelerates Development of Chip Manufacturing in Ohio
I wonder what really happened to the AI acquisitions Intel made a few years ago. Intel spent $2B on Habana Labs in 2019? They also acquired Nirvana in 2016 for $400M. These companies were packed with very intelligent people! The expectation was for Intel to be an AI powerhouse. So many missed opportunities. How is the Habana Labs Gaudi chip doing? It was supposed to be the Nvidia killer. I don't think the Nirvana chip ever made it out?
The rumors and conjecture in the press were that customers who sampled the Nirvana chips (there were two, one for training and one for inference) were unhappy with the performance they got, and Intel just killed the projects.
Habana's Gaudi 3 looks much more sophisticated than Nirvana's chips, perhaps the most sophisticated AI processor internally I've seen (though this is not my primary field), SIMD VLIW, and Gaudi 3 is claimed to be good for training and inference, but Intel's revenue outlook for Gaudi 3 is only $500M in 2024. For whatever reason, that's obviously disappointing.
Are we moving towards a future where Big tech design their own customized chips to suit their particular requirements, outsourcing manufacturing to foundries and bypassing Intel and AMD?
I think the issue for AWS, Google Cloud, and Oracle is that they still need x86 CPUs for apps running in the cloud. The cloud vendors' own software stacks can run on any CPU they care to port or write the software to, but many business apps which have roots in enterprise computing are still compiled and tested on x86. Enterprise x86 CPUs, which include most Intel and AMD CPUs, are probably not what the cloud vendors exactly have in mind.
Yup, and even for AI, we get so wrapped up in the accelerated model (which can use specialized chips), but the all data management, dataset cleaning, RAG, have to run on x86 stacks.
I wonder what really happened to the AI acquisitions Intel made a few years ago. Intel spent $2B on Habana Labs in 2019? They also acquired Nirvana in 2016 for $400M. These companies were packed with very intelligent people! The expectation was for Intel to be an AI powerhouse. So many missed opportunities. How is the Habana Labs Gaudi chip doing? It was supposed to be the Nvidia killer. I don't think the Nirvana chip ever made it out?
the same thing that happens with Intel. Intel is very slow to deliver on chips. If they control the architecture (cpus/bus), that is fine because they determine everything. But if changes and adaptation are required, Intel takes a year to do this. The end result is that Intel cannot adjust in fast paced markets. In Habana/Nirvan case (and intel was working on AI long before this). they needed to make changes and the teams are too slow and Intel management is too slow. eventually they have a meeting where the presentation is "this product doesnt deliver what customers want, it would take 3-4 years to do a new product" and the product or division is correctly killed.
When Intel guesses right, or when the roadmap is long and steady, Intel delivers well. Chevy, Ferrari, and mercedes are different companies with different culture and markets
I can review detailed examples on AI, Mobile, comms, and other acquisitions that highlight why this doesnt work for Intel but it works great for nvidia.
Are we moving towards a future where Big tech design their own customized chips to suit their particular requirements, outsourcing manufacturing to foundries and bypassing Intel and AMD?
No, they are too low volume to reap economic benefit. Even smartphone makers who ventured into basic chips have found out that their volumes are uneconomical.
Their entire tapeout to date is likely 1 week of Intel Xeon production.
I think the issue for AWS, Google Cloud, and Oracle is that they still need x86 CPUs for apps running in the cloud. The cloud vendors' own software stacks can run on any CPU they care to port or write the software to, but many business apps which have roots in enterprise computing are still compiled and tested on x86. Enterprise x86 CPUs, which include most Intel and AMD CPUs, are probably not what the cloud vendors exactly have in mind.
I think IFS with Intel design will be a sweet way for Hyperscaler to reduce cost with x86 they can get quite a deal vs the traditional deal if IFS can deliver cause Intel's product team can deliver for sure
spectrum.ieee.org
Microsoft is already previewing the Cobalt100 CPU. Google is about to deploy the Axion CPU in Google Cloud. Meta is currently only designing AI chips, but they're looking for a partner to help them design a custom Meta CPU chip. My baseless guess is Broadcom becomes the partner.
