"The long and short of the story there is that Intel has decided that they mistimed the market for their first XPU, and that Falcon Shores as an XPU would have wound up being premature."
That actually makes some sense.
Any number of companies have been talking up revolutionary architectures for years... And outside of HPC, the ones that didn’t emulate Nvidia boxes pretty much failed, even ones that match the workload really well (like an XPU would match a LLM).
Yeah normally I eyeroll at the corporate speak to excuse technology failures, but I have to agree this does make sense. They may also be waiting for a more mature tile packaging so they can meet those dynamic needs which would make them the only game in town.
Even Nvidia is coupling the CPU tighter with Grace and it's obvious that they'll do more of that before Falcon Shores in 2025. This is just another pathetic Intel execution failure excuse.
If Intel really thinks the market won't be sure by 2025 they could develop both(with and without) to be 100% sure. Instead they are just waiting it out??? No, they are not able to integrate the CPU.
This AI opportunity will be bigger than the rest of the DC market. They wouldn't half-ass it, they see that it's hopeless for them to develop good XPUs by 25.
I dunno... I think most AI trainers/runners will stick to their x86 Nvidia boxes rather than trying to get the grace superchips to work, at least for awhile. Even basic optimizations are often passed up because its not turnkey in the popular repos/solutions.
Hence Intel's pivot here. There's a chance users will pick an Intel (or AMD) GPU at a discount if it "just works" with the right docker image and a python import or whatever. They are not going to spend weeks optimizing for a unified memory XPU, if thats even possible on their framework of choice.
What are you talking about? The vast majority of AI revenues come from internal use at huge companies like Meta, Amazon, Microsoft and Google. These companies are not too lazy to optimize anything and all of them even develop their own libraries.
BTW: XPUs won't need any optimization from small end-users. The optimization needs to come from Intel for the most important libraries. It's not like end-users use anything but standard applications.
I think they figured out that merging the Habana Gaudi architecture ROCE and tripling the in-package HBM would increase performance more than putting a Xeon in the package. They do have 800Gbe, CXL 3.0 networking, in-package optical interconnects, IPUs and other IP to throw at future designs.
I read a Marvell interview that was promoting CXL with optical switching for AI processing.
Intel typically compared their next gen CPU with last gen AMD (AMD current gen typically 1.5x to 2.3x performance of last gen). Independent reviews when available won't look this pretty. YMMV.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
14 Comments
Back to Article
brucethemoose - Monday, May 22, 2023 - link
"The long and short of the story there is that Intel has decided that they mistimed the market for their first XPU, and that Falcon Shores as an XPU would have wound up being premature."That actually makes some sense.
Any number of companies have been talking up revolutionary architectures for years... And outside of HPC, the ones that didn’t emulate Nvidia boxes pretty much failed, even ones that match the workload really well (like an XPU would match a LLM).
thestryker - Monday, May 22, 2023 - link
Yeah normally I eyeroll at the corporate speak to excuse technology failures, but I have to agree this does make sense. They may also be waiting for a more mature tile packaging so they can meet those dynamic needs which would make them the only game in town.EuropeOverlords - Monday, May 22, 2023 - link
Only thing what makes me not sell my NVIDIA stock is that they acquired Mellanox few years back.So if you think about that,...
brakdoo - Tuesday, May 23, 2023 - link
Even Nvidia is coupling the CPU tighter with Grace and it's obvious that they'll do more of that before Falcon Shores in 2025. This is just another pathetic Intel execution failure excuse.If Intel really thinks the market won't be sure by 2025 they could develop both(with and without) to be 100% sure. Instead they are just waiting it out??? No, they are not able to integrate the CPU.
This AI opportunity will be bigger than the rest of the DC market. They wouldn't half-ass it, they see that it's hopeless for them to develop good XPUs by 25.
A5 - Tuesday, May 23, 2023 - link
Yeah. 2025 they'll be looking at Grace 2 + Blackwell, and they're already waaaay behind Hopper in perf/W.name99 - Saturday, May 27, 2023 - link
Falcon Shores' successor, codenamed Pauly Shores, will CRUSH Grace 2 + Blackwell!brucethemoose - Tuesday, May 23, 2023 - link
I dunno... I think most AI trainers/runners will stick to their x86 Nvidia boxes rather than trying to get the grace superchips to work, at least for awhile. Even basic optimizations are often passed up because its not turnkey in the popular repos/solutions.Hence Intel's pivot here. There's a chance users will pick an Intel (or AMD) GPU at a discount if it "just works" with the right docker image and a python import or whatever. They are not going to spend weeks optimizing for a unified memory XPU, if thats even possible on their framework of choice.
brakdoo - Tuesday, May 23, 2023 - link
What are you talking about? The vast majority of AI revenues come from internal use at huge companies like Meta, Amazon, Microsoft and Google. These companies are not too lazy to optimize anything and all of them even develop their own libraries.BTW: XPUs won't need any optimization from small end-users. The optimization needs to come from Intel for the most important libraries. It's not like end-users use anything but standard applications.
JayNor - Wednesday, May 24, 2023 - link
I think they figured out that merging the Habana Gaudi architecture ROCE and tripling the in-package HBM would increase performance more than putting a Xeon in the package. They do have 800Gbe, CXL 3.0 networking, in-package optical interconnects, IPUs and other IP to throw at future designs.I read a Marvell interview that was promoting CXL with optical switching for AI processing.
EuropeOverlords - Monday, May 22, 2023 - link
Aurora -1,359 TB of CPU HBM RAM
8,159 TB of GPU HBM RAM
21,248 TB of DDR5 RAM " cache " XD
do you even need SSDs ? XD
JayNor - Wednesday, May 24, 2023 - link
Intel delivered a new generation of Optane and their DAOS storage in Aurora also.tygrus - Thursday, May 25, 2023 - link
Intel typically compared their next gen CPU with last gen AMD (AMD current gen typically 1.5x to 2.3x performance of last gen). Independent reviews when available won't look this pretty. YMMV.CHADBOGA - Friday, May 26, 2023 - link
If the forum mods could unban me that would be good. I have served my time.nandnandnand - Friday, May 26, 2023 - link
You're spamming on this on multiple articles, so you should serve a life sentence.