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

[Josh128]

I think building a set of products around Nvidia GPUs that compete with it could cause considerable harm to Intel. If Nvidia prioritizes their own set of products, for example, it will be hard for Intel to compete when customers simply can't get their hands on the latest GPGPUs unless sold as part of an Nvidia rack.

It could also work fine. But risk remains. If the situation for Nvidia demands it why wouldn't they throw them under the bus.

Intel is currently dead in the GPGPU / datacenter GPU space. Therefore in that regard I really think they have

 

[marees]

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)

 

[StinkyPinky]

Intel is currently dead in the GPGPU / datacenter GPU space. Therefore in that regard I really think they have

View attachment 130503

Such an underrated movie....

 

[Josh128]

Such an underrated movie....

One of the funniest of all time, IMO.

 

[inquiss]

eh good enough.
MI500 comp here we gooo.

low.

high.

You think they can get past TSMC? When?

 

[adroc_thurston]

You think they can get past TSMC? When?

idk if past, but very competitive? oh yes.
And it all hinges on A14 (and A10) execution.

18A-P will be an aight node when they manage to make it yield, but that's N3p best-case perf/power.
They need to close the gap further.

 

[gdansk]

Intel is currently dead in the GPGPU / datacenter GPU space. Therefore in that regard I really think they have

I'm not talking about Intel GPGPUs. If Intel builds Xeons for NVLink but if it turns out Xeon customers cannot get anything to NVLink together then it's a lost investment.
If Nvidia did this because they felt pressured they may feel no compunction when the pressure dies of fat. And then do what they rather preferred all along.

 

[adroc_thurston]

If Intel builds Xeons for NVLink but if it turns out Xepn customers cannot get anything to NVLink together then it's a lost investment.

You do understand that those will be kustom Xeons with kustom IODs?
Like you're misreading what's happening because nvidia is EVILLLLLLLLLLLLL and so on. boring.

 

[Josh128]

eh good enough.
MI500 comp here we gooo.

low.

high.

Even if they manage to catch up with TSMC, they have to be able to do it at similar cost. That is going to be the real challenge.

For CPUs, they wont have a problem other than the above.

Thats the conundrum, if 2 is dependent on 1, and 1 is also quite dependent on 2, that makes it doubly hard and is basically why they are in the situation they are in, currently.

 

[adroc_thurston]

they have to be able to do it at similar cost

That's the easy part. Perf/power/yields are Very Hard.

 

[Josh128]

That's the easy part. Perf/power/yields are Very Hard.

Disagree completely on that being the easy part. The two go hand in hand. Taiwan labor costs and working environment expectations are not the same as those in the United States. The circa 2017 days of selling monolithic 8 or 10 core prosumer CPUs for $2000 ($4000 in todays dollars) are long gone.

 

[adroc_thurston]

Taiwan labor costs and working environment expectations are not the same as those in the United States.

Labor costs or 996 slavery do not impact raw wafer prices much.
Node itself, foundry capex intensity etc do.

 

[Josh128]

Labor costs or 996 slavery do not impact raw wafer prices much.
Node itself, foundry capex intensity etc do.

Engineering=labor.

 

[gdansk]

You do understand that those will be kustom Xeons with kustom IODs?
Like you're misreading what's happening because nvidia is EVILLLLLLLLLLLLL and so on. boring.

You're misreading it because muh Tegra sucks. Nvidia isn't stopping anything.

Intel benefits if they can develop better custom chips on time and competitive with what Nvidia is doing. If they miss, tough luck. No reason for it to exist if they mess it up again.

 

[adroc_thurston]

Engineering=labor.

Again, wafer costs are driven not in per-headcount opex, but general node costs and foundry capex intensity.
That's why Intel is e-begging for 14A customers. 14A will be a very capex-intensive node and they need to recoup the investment somehow.

 

[adroc_thurston]

Nvidia isn't stopping anything.

They obviously are, it's a de-facto surrender.

Intel benefits if they can develop better custom chips on time and competitive with what Nvidia is doing

Well, it's x86 and it works, which is 2/2 going to Xeon.

 

[511]

18A-P will be an aight node when they manage to make it yield, but that's N3p best-case perf/power.

i disagree the perf/power will be near base N2. 18A is already close to N3P in CPU IP

 

[adroc_thurston]

i disagree the perf/power will be near base N2

nope.

18A is already close to N3P in CPU IP

5%-ish behind N3e give or take.

 

[Joe NYC]

I fully expect Nvidia to start investing heavily into Intel's labs and start making some of their products there in the 2030s. This is a win for Nvidia, they need to be diverse in their manufacturing

Why do you think NVidia would be OK on inferior process node competing against AMD on superior node?

It will happen with Rubin vs. Mi400, and NVidia wants to correct it in generation after that, and wants to jump to A16

 

[511]

5%-ish behind N3e give or take.

Definitely nope Cougar is like 10-15% better perf-watt vs LNL LNC N3B

 

[adroc_thurston]

Cougar is like 10-15% better perf-watt vs LNL LNC N3B

That's because LNC physdes was a catastrophe.

 

[marees]

Why do you think NVidia would be OK on inferior process node competing against AMD on superior node?

It will happen with Rubin vs. Mi400, and NVidia wants to correct in generation after that, and wants to jump to A16

In the history of GPUs, anything made outside TSMC fabs has suffered an efficiency penalty

That has not & will not change. Nvidia will make all its chiplets at TSMC. Maybe Intel can compete with CoWoS

 

[511]

That's because LNC physdes was a catastrophe.

Not on LNL

 

[adroc_thurston]

Not on LNL

Still there too, just less scuffed.

 

[MS_AT]

This is more relevant in context of server Tegra.

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. Since gamers buy nVidia because its nVidia it could sell. Still probably in practice they may not want to do that as every sell counts.

Btw do you suppose that Vera will be their last custom ARM affair? Or Intel plot is only to increase TAM for their "super chips"?