News NVIDIA and Intel to Develop AI Infrastructure and Personal Computing Products

[Joe NYC]

This seems to be Intel trying to stop the bleeding vs AMD in DC by adding nvlink integration

Basically Intel is the new mediatek (with some additional packaging capabilities)

Looking for an edge in asymmetric fashion is a good out of the box thinking. Only reason Intel is doing this is because it is kind of dark in the box. Without NVidia keeping Intel afloat (even before this deal), Intel's prospects in the datacenter CPU were not great. Especially with hyperscalers.

I have not heard what percentage of Intel's sale to hyperscalers are due to being in NVidia reference designs. Probably high enough number, meaning Intel has no choice but to do what NVidia wants.

 

[adroc_thurston]

Actually I was thinking along the lines of consumer market, where nVidia could only sell NVLink enalbed parts. Mediatek would cover WoA, Intel normal Windows

WoA is dead, x86 forever.

Btw do you suppose that Vera will be their last custom ARM affair?

Well, yeah, Vera is having the struggle session of its life.

Intel's prospects in the datacenter CPU were not great. Especially with hyperscalers.

I have not heard what percentage of Intel's sale to hyperscalers are due to being in NVidia reference designs. Probably high enough number, meaning Intel has no choice but to do what NVidia wants.

CPU contents are tiny in GPGPU farms and do not drive a fat ASP mix.
This play is by Nvidia, for Nvidia. Against AMD.

 

[511]

I want to drop a bomb as well

Spoiler: leak

 

[jpiniero]

CPU contents are tiny in GPGPU farms and do not drive a fat ASP mix.

That's why the nVidia GPU x86 clusters have been using Xeons.

I don't think this really changes much for HPC. If products do come out, I think it's more aimed at Strix Halo and XboxPC than HPC.

 

[adroc_thurston]

That's why the nVidia GPU x86 clusters have been using Xeons

There are very specific reasons for that.

I don't think this really changes much for HPC

Yeah it does. Not for Intel, though.

If products do come out, I think it's more aimed at Strix Halo and XboxPC than HPC.

Client parts irrelevant.

 

[Doug S]

I think people are overselling this with talk about antitrust concerns and worry over whether Intel will end all GPU development.

Nvidia is already present in higher end Intel laptops, this allows for a more integrated solution and will be attractive to OEMs because they can design laptops that don't provide for a dGPU. So its a win for Intel for obvious reasons, its a win for Nvidia because they limit AMD's ability to place dGPUs in OEM laptop configs, its a win for OEMs because it simplifies BTO with less stuff to configure. Maybe it extends beyond laptops but if it does it will be for "laptop like" configs like SFF or similar stuff targeted at the enterprise/embedded market that's got a soldered CPU and maybe soldered RAM too. Don't see it for traditional desktops that include an x16 PCIe slot.

It remains to be seen how broadly Nvidia will offer but I'm betting it will be higher end only - stuff that would be a dGPU and is above any Intel iGPU offerings performance wise.

It would be nice if Nvidia committed to having the GPU chiplets built at Intel to give them some business but even if they decided that today it would be at least a couple years away as they'd have to port one of their designs, tape it out etc.

 

[adroc_thurston]

I think people are overselling this with talk about antitrust concerns and worry over whether Intel will end all GPU development.

Nvidia is already present in higher end Intel laptops, this allows for a more integrated solution and will be attractive to OEMs because they can design laptops that don't provide for a dGPU. So its a win for Intel for obvious reasons, its a win for Nvidia because they limit AMD's ability to place dGPUs in OEM laptop configs, its a win for OEMs because it simplifies BTO with less stuff to configure. Maybe it extends beyond laptops but if it does it will be for "laptop like" configs like SFF or similar stuff targeted at the enterprise/embedded market that's got a soldered CPU and maybe soldered RAM too. Don't see it for traditional desktops that include an x16 PCIe slot.

It remains to be seen how broadly Nvidia will offer but I'm betting it will be higher end only - stuff that would be a dGPU and is above any Intel iGPU offerings performance wise.

It would be nice if Nvidia committed to having the GPU chiplets built at Intel to give them some business but even if they decided that today it would be at least a couple years away as they'd have to port one of their designs, tape it out etc.

Client stuff is whatever, you don't want Intel inside laptops anyway.

It's the DC part that is really, really interesting.

 

[511]

The thread is certainly trending xDd

 

[Doug S]

Client stuff is whatever, you don't want Intel inside laptops anyway.

It's the DC part that is really, really interesting.

I don't know that the DC part is all that interesting. It isn't like the CPU ISA in an AI cluster has much importance. This does drive some volume Intel's way but it is tough to see what Nvidia gains out of this part of the deal, unless they're getting more out of the client part than we're seeing.

 

[inquiss]

Client stuff is whatever, you don't want Intel inside laptops anyway.

It's the DC part that is really, really interesting.

In what ways and how badly does this affect AMD in DC? Both CPU and GPU?

 

[adroc_thurston]

I don't know that the DC part is all that interesting

Maybe not for you.

It isn't like the CPU ISA in an AI cluster has much importance.

Oh it is, NV spent years moving the CUDA stack in preparation for inhouse aa64 SoCs, only to make x86 a first class citizen again.

what Nvidia gains out of this part of the deal

They get x86.

In what ways and how badly does this affect AMD in DC? Both CPU and GPU?

Not being chained to butthorrible Tegra roadmap is one.

 

[511]

Not being chained to butthorrible Tegra roadmap is one.

They hired Intel Griffin Cove Chief Architect for this btw

 

[Josh128]

I want to drop a bomb as well

Spoiler: leak

View attachment 130513

20% more perf at same power? What the heck is WCL? Water Canon Lake? Water Cooler Lake? West Covina Lake??

 

[adroc_thurston]

They hired Intel Griffin Cove Chief Architect for this btw

I know, my linkedin-fu isn't so rusty as to not track senior staffer movement.

 

[511]

20% more perf at same power? What the heck is WCL? Water Canon Lake? Water Cooler Lake? West Covina Lake??

Wild Cat Lake

 

[511]

I want to drop a bomb as well

Spoiler: leak

View attachment 130513

@igor_kavinski

 

[jpiniero]

There are very specific reasons for that.

Probably just offered it for dirt cheap.

 

[igor_kavinski]

@igor_kavinski

What is that and why in this thread if it's Panther Lake?

 

[511]

What is that and why in this thread if it's Panther Lake?

i wanted to drop a bomb when another bomb has been dropped 🤣 🤣

 

[Josh128]

@igor_kavinski

What a strange graph... so WCL is about 22% better perf/w than Lunar Lake. You would think they would have used LL at 1.0 and place WCL accordingly, but from my best estimate this is saying .99 WCL /.81 LL perf at ISO power, or +22%.

So +22% perf/w going from Intel 4 to 18A at presumably Wild Cats sweet spot. Passable, but certainly not a bombshell.

 

[511]

What a strange graph... so WCL is about 22% better perf/w than Lunar Lake. You would think they would have used LL at 1.0 and place WCL accordingly, but from my best estimate this is saying .99 WCL /.81 LL perf at ISO power, or +22%.

also it's not strange xDd it's normal for any CPU Company to do this

So +22% perf/w going from Intel 4 to 18A at presumably Wild Cats sweet spot. Passable, but certainly not a bombshell.

Just for reference

 

[Josh128]

View attachment 130517
Just for reference

Is this power draw or score?

**EDIT, must be score.

 

[511]

Is this power draw or score?

this is 258V Score so a baseline of 1 for the graph

SPEC CPU 2017 | David Huang's Blog

edit: thanks i choose the wrong mission while playing COD MW alt tabbing.

 

[MS_AT]

Just for reference

Is the reference form the same source or just random score from the net that is irrelevant for the comparison at hand due to unknown compiler settings?

 

[511]

Is the reference form the same source or just random score from the net that is irrelevant for the comparison at hand due to unknown compiler settings?

Oh F I forgot about the compiler Intel uses ICX for their testing but it's nearly the same arch it's like comparing GLC with RPC. I don't know about the exact flag just that it's ICX with the same flags.