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

[Joe NYC]

Another thing I picked up from the conference call: Intel / NVidia client PC is to use NVLink to connect CPU and GPU.

From what we have seen from AMD Strix Halo and Medusa Helo leaks, these processors will have memory controllers on the iGPU chiplet. Which, presumably, NVidia / Intel will also implement.

So, this will then mean that for Intel CPU chiplet to get access to memory, it has to go over NVLink to the NVidia GPU. Intel will have to go over foreign IP just to get a memory access for its CPU.

This to me, this looks like a surrender on part of Intel.

 

[fastandfurious6]

Every + Intel added after 10nm+++ over the last decade was foreshadowing unconditional surrender

 

[Joe NYC]

Every + Intel added after 10nm+++ over the last decade was foreshadowing unconditional surrender

A lot of people lived under the illusion that Intel's problems are centered around fab missteps.

But with this deal, Intel is basically licensing 2 sets of IP from NVidia. Which means, can no longer create its own IP.

 

[adroc_thurston]

From what we have seen from AMD Strix Halo and Medusa Helo leaks, these processors will have memory controllers on the iGPU chiplet. Which, presumably, NVidia / Intel will also implement.

you're making a dumb assumption. don't.
GB10 itself has NV tile for basically just graphics. And it'll be the same here, too.

But with this deal, Intel is basically licensing 2 sets of IP from NVidia. Which means, can no longer create its own IP.

This is nonsense.

 

[Joe NYC]

you're making a dumb assumption. don't.
GB10 itself has NV tile for basically just graphics. And it'll be the same here, too.

It is an assumption, but why do you think it is dumb?

@Kepler_L2 posted that he thinks Medusa level GPU chiplets will be the ones with the memory controllers, not CPU chiplets. It makes sense when it gets to > 2 memory channels.

Why would NVidia do it differently, and not have the memory controllers on the GPU chiplet?

 

[jpiniero]

Why would NVidia do it differently, and not have the memory controllers on the GPU chiplet?

So the CPU can be sold without a GPU?

 

[Joe NYC]

So the CPU can be sold without a GPU?

Why would Intel be selling a CPU with NVLink with nothing attached to it? Rather then on of its other CPUs without the NVLink?

 

[adroc_thurston]

It is an assumption, but why do you think it is dumb?

You're applying AMD things to companies not AMD?

@Kepler_L2 posted that he thinks Medusa level GPU chiplets will be the ones with the memory controllers, not CPU chiplets. It makes sense when it gets to > 2 memory channels.

I know what gfx13 is jfc.

Why would NVidia do it differently, and not have the memory controllers on the GPU chiplet?

more flexibility and you get to offload validation costs to someone else.
see gb10/n1x.

 

[adroc_thurston]

Why would Intel be selling a CPU with NVLink with nothing attached to it? Rather then on of its other CPUs without the NVLink?

you do understand that the d2d PHY will probably be flexIO with NVL-C2C being one of the protocols supported?

 

[Joe NYC]

you do understand that the d2d PHY will probably be flexIO with NVL-C2C being one of the protocols supported?

I suspected it may just be the a protocol.

But if memory controllers are indeed on the GPU die, that will still mean Intel CPU will have to access its memory via someone else's protocol.

It is by no means given that the NVidia's iGPU chiplet for Intel will follow GB10. AMD will be going to 2nd generation of Halo product with memory controllers on the same die as iGPU. Which may suggest that it is a more optimal solution.

 

[adroc_thurston]

But if memory controllers are indeed on the GPU die, that will still mean Intel CPU will have to access its memory via someone else's protocol.

and?
RISC-V shithouse talks almost universally in AMBA CHI and this hasn't been a problem even once.

AMD will be going to 2nd generation of Halo product with memory controllers on the same die as iGPU. Which may suggest that it is a more optimal solution.

AMD always does things their own way, stop applying their crackpipe worship to everyone else.

 

[Joe NYC]

and?
RISC-V shithouse talks almost universally in AMBA CHI and this hasn't been a problem even once.

It's just ironic how far the mighty have fallen if Intel's CPU cores have to use someone else's protocol to access memory.

AMD always does things their own way, stop applying their crackpipe worship to everyone else.

There have been multiple approaches by AMD to chiplets and they never had one where the memory controllers / PHY were on the same die as CPU, and multiple on where memory is on the same die as memory controller.

But, on Intel side, it appears that Intel is going back to having the memory controller on the same die as CPU (with Panther Lake).

BTW, this gives NVidia opportunity to copy the entire AMD GPU roadmap, and have this iGPU chiplet to be the bases of low end GPUs using LPDDR5

 

[adroc_thurston]

It's just ironic how far the mighty have fallen if Intel's CPU cores have to use someone else's protocol to access memory.

No?

There have been multiple approaches by AMD to chiplets and they never had one where the memory controllers / PHY were on the same die as CPU, and multiple on where memory is on the same die as memory controller.

again please stop applying permament AMD 24/7 cracksmoke to anyone else.

But, on Intel side, it appears that Intel is going back to having the memory controller on the same die as CPU (with Panther Lake).

NVL has discrete SOC tiles again.

BTW, this gives NVidia opportunity to copy the entire AMD GPU roadmap, and have this iGPU chiplet to be the bases of low end GPUs using LPDDR5

They don't have to.

 

[Doug S]

But with this deal, Intel is basically licensing 2 sets of IP from NVidia. Which means, can no longer create its own IP.

So Apple could not create its own IP when it licensed GPUs from Imagination? Qualcomm could not create its own IP when it licensed CPU cores from ARM?

 

[Joe NYC]

So Apple could not create its own IP when it licensed GPUs from Imagination? Qualcomm could not create its own IP when it licensed CPU cores from ARM?

Intel is now licensing IP for what is Intel's core business, replacing own IP with licensed IP.

Not the case in your examples. In fact, Qualcomm is going the opposite direction and developing its own Arm cores.

 

[jpiniero]

Intel is now licensing IP for what is Intel's core business, replacing own IP with licensed IP.

Intel is not licensing anything other than NVLink.

I was under the impression that nVidia will be the one selling the products, if any end up coming out.

 

[adroc_thurston]

So Apple could not create its own IP when it licensed GPUs from Imagination?

Genuinely wonder what happened there. Apple managed to piecemeal rebuild imgtec IP bit by bit starting with A8.

Intel is now licensing IP for what is Intel's core business, replacing own IP with licensed IP.

omaigotto Xscale existed and after that Intel used imgtec GPU IP for Atom. And modems. They used external modems before buying out Infineon's radio business.

Your poasts are straight up embarrassing.

I was under the impression that nVidia will be the one selling the products, if any end up coming out.

yes for DC, no for client.

 

[jpiniero]

yes for DC, no for client.

That does make sense.

 

[adroc_thurston]

That does make sense.

Yeap, DC is just putting x86 back where it belongs (top-end GPGPU rackscale things), client is a new product.

 

[poke01]

WoA will become irrelevant as soon as Nova Lake comes out. This Nvidia deal with Intel, killed it even more

 

[adroc_thurston]

WoA will become irrelevant as soon as Nova Lake comes out

To become irrelevant it needs to become relevant first.
NVL isn't even good or noteworthy or interesting anyway.

This Nvidia deal with Intel, killed it even more

Well yeah duh that's pretty obvious.

 

[Darkmont]

NVL isn't even good or noteworthy or interesting anyway.

It’s a kitchen sink product and that’s entertaining enough for me

 

[adroc_thurston]

It’s a kitchen sink product and that’s entertaining enough for me

18A-P parts will be interesting for sure. Otherwise, meh.

 

[poke01]

Hopefully Zen6 is interesting

NVL isn't even good or noteworthy or interesting anyway.

 

[adroc_thurston]

Hopefully Zen6 is interesting

It's a normal AMD tick + N2p is nice.
But the lineup is very curious.