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

[Covfefe]

So I listened to the call. My takeaways (in no particular order).

• AI! AI! AI!
• They focused more on server CPUs than client GPUs.
• They said Intel is selling server CPUs to Nvidia who will sell them to customers.
• The deal has supposedly been in the works for nearly a year
• If I heard him right, Lip-Bu Tan mentioned unified memory for the client SOCs which would imply it's a true iGPU.
• Huang said this deal has no impact on Nvidia's ARM roadmap.
• No announcement of Nvidia products being fabbed by Intel Foundry
• No mention of what happens to Intel's graphics IP or datacenter AI accelerators.
• When asked why the cash infusion into Intel, Huang said "It's a good investment".

 

[igor_kavinski]

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

Interesting, though two things:

1) If it's better than expectations, it will be priced too high.

2) I would rather have the cheaper WCL as that looks to be the real hero of performance/$.

 

[jpiniero]

If I heard him right, Lip-Bu Tan mentioned unified memory for the client SOCs which would imply it's a true iGPU

I think I mentioned this in one of the other threads but I believe that's how the Mediatek product works, using NVLink to connect the two dies.

 

[Josh128]

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.

So if I am looking at this correctly, 13900K/7950X3D vcache die single thread integer perf at Lunar Lake power draw. Not bad.

 

[gdansk]

So if I am looking at this correctly, 13900K/7950X3D vcache die single thread integer perf at Lunar Lake power draw. Not bad.

Nope, you're extrapolating from 5W to LNL at not 5W.
David's spec test is ~ peak frequency. What 511 posted is not.
Even the curve of the part of the line you can see will suggest this is a bad extrapolation.

 

[Josh128]

Nope, you're extrapolating from 5W to LNL at not 5W.
David's spec test is ~ peak frequency. What 511 posted is not.
Even the curve of the part of the line you can see will suggest this is a bad extrapolation.

Does a 258V pull more than 5W during an ST run?

 

[gdansk]

Does a 258V pull more than 5W during an ST run?

I don't have one. But 226V, yes very much so if it's platform power.

 

[511]

So if I am looking at this correctly, 13900K/7950X3D vcache die single thread integer perf at Lunar Lake power draw. Not bad.

Nope that is not true it is likey M2 PRO ST Perf at 7-8W

Nope, you're extrapolating from 5W to LNL at not 5W.
David's spec test is at peak frequency. What 511 posted is not.
Even the curve of the part of the line you can see will suggest this is a bad extrapolation.

this is relative graph not absolute though

 

[gdansk]

this is relative graph not absolute though

And post the rest of the graph if you want to predict how it will perform in David's tests. It is unrelated to the side of the chart shown.
Frequency at low voltage increasing is not a surprise. Peak frequency not increasing? Also... not a surprise.

 

[511]

Does a 258V pull more than 5W during an ST run?

It pulls about 12W

 

[511]

And post the rest of the graph if you want to predict how it will perform in David's tests. It is unrelated to the side of the chart shown.

I agree with you but we can extrapolate based on the data of Huang's Graph.
It's ICX vs GCC though.

 

[Josh128]

yeah it is, x86 server SOCs with NVL-C2C means a whole lot of things and none of them are good in context of the server Tegra roadmap.

Huang said this deal has no impact on Nvidia's ARM roadmap.

Told you man.

 

[Josh128]

So I listened to the call. My takeaways (in no particular order).

• AI! AI! AI!
• They focused more on server CPUs than client GPUs.
• They said Intel is selling server CPUs to Nvidia who will sell them to customers.
• The deal has supposedly been in the works for nearly a year
• If I heard him right, Lip-Bu Tan mentioned unified memory for the client SOCs which would imply it's a true iGPU.
• Huang said this deal has no impact on Nvidia's ARM roadmap.
• No announcement of Nvidia products being fabbed by Intel Foundry
• No mention of what happens to Intel's graphics IP or datacenter AI accelerators.
• When asked why the cash infusion into Intel, Huang said "It's a good investment".

1. Strix Halo competitor, I knew it. Huang is that egotistical, it just pisses him off not having one of his own.

2. So NV ARM not ded.

3. Dagger in the heart. This is a far cry from Nvidia buying Intel that 'tards were screaming about this morning. No wonder INTC went from +32% pre-market morning down to +22% currently.

 

[NTMBK]

Such an underrated movie....

It's definitely one of the movies of all time

 

[desrever]

The first product from this won't be until likely 2028? From what it looks like, Intel will have to develop something with nvlink and then buy Nvidia GPU chiplets to glue onto their own stuff. Also I don't see DC being a factor, DC will be using Nvidia's solutions so won't need to involve anything intel. Theres no reason for an x86 "APU" in the DC space.

I am pretty pessimistic its going to be a game changer. Likely Arc is dead but not sure what even is the point of this thing outside of just politics.

 

[NTMBK]

• If I heard him right, Lip-Bu Tan mentioned unified memory for the client SOCs which would imply it's a true iGPU.

I've heard this one before

 

[LightningZ71]

NVidia still had a hole in their portfolio in the sub xx60 series dGPU market in notebooks. Oh, they offer an xx50, but it's decidedly unpopular. Making a tile for Intel's SoCs gives them a product to sell at volume in that space.

Desktop is still, effectively, tiny. Mobile is far more important. AMD is still getting volume in consoles, mobile APUs and still has some dGPU presence however tiny. NVidia needs shipped volume with their IP to maintain developer interest.

This does run squarely into Nova Lake Premium big iGPU though...

 

[adroc_thurston]

Huang said this deal has no impact on Nvidia's ARM roadmap.

Told you man.

He would never, ever admit surrender in public.
What did you expect?

 

[Tup3x]

It seems like NVIDIA just wants to make sure that what ever little Intel is going to sell in the future, it will come with NVIDIA GPU.

 

[Josh128]

It seems like NVIDIA just wants to make sure that what ever little Intel is going to sell in the future, it will come with NVIDIA GPU.

Yeah, its a high profile nothingburger, really.

 

[Khato]

It's easy to believe that some portion of this deal has been in the works for a year. Intel courting NVIDIA with a custom processor for server AI makes sense on both sides. Sure NVIDIA would like to displace x86, but they aren't going to make much more money by doing so. Getting nvlink on Intel server CPUs meanwhile is a direct upgrade to what they've been selling for years. And in the current state Intel is happy to have the business.

The $5B investment is easily understood from the political perspective. And is the same reason why I'll not be the least bit surprised to see some reason for Apple investing in Intel at some point. Unlike NVIDIA they don't have other justification to point to other than an intention to use Intel Foundry at some point. (NVIDIA can fairly claim no current plans to use Intel Foundry... but they can change those plans whenever they want.)

The NVIDIA graphics chiplet offering is the real unknown here. Is it just going to be proper high-end similar to what they were intending with their ARM offering? Or are they going to be displacing Intel's integrated offerings? I can easily see either given Intel's tendency toward shortsighted strategies.

 

[511]

What I see is this displacing the high end gaming laptops Rtx 5090 laptops will be large APUs and the low tier would be Xe iGPU.

 

[CouncilorIrissa]

Does a 258V pull more than 5W during an ST run?

Hell yeah it does (yes, cinebench, yadda-yadda).

 

[jpiniero]

The NVIDIA graphics chiplet offering is the real unknown here. Is it just going to be proper high-end similar to what they were intending with their ARM offering? Or are they going to be displacing Intel's integrated offerings? I can easily see either given Intel's tendency toward shortsighted strategies.

Killing Arc is something Intel should have done when the second 10 nm delay happened. They don't have to kill the IP entirely, only slash the spending to keep it adequate for non-gaming purposes.

As someone eluded to before, I could easily see regulators in some other country killing this deal.

 

[Khato]

Killing Arc is something Intel should have done when the second 10 nm delay happened. They don't have to kill the IP entirely, only slash the spending to keep it adequate for non-gaming purposes.

What market opportunities remain for Intel if they kill Arc? Best case they regain CPU market share, but then what? All the high margin opportunities are on graphics/AI, not CPU.