Question Nvidia to enter the server CPU market

[NTMBK]

NVIDIA Unveils Grace: A High-Performance Arm Server CPU For Use In Big AI Systems

www.anandtech.com

Based on Neoverse cores, not a Denver derivative.

This makes sense with the ARM acquisition attempt. They want to fully control the server stack.

 

[beginner99]

This makes sense with the ARM acquisition attempt. They want to fully control the server stack.

Which means the worst that was expected from this acquisition will probably happen. Typical nv property locked-gardens.

 

[NTMBK]

Which means the worst that was expected from this acquisition will probably happen. Typical nv property locked-gardens.

I'm not so sure. This feels very special purpose to their GPU servers- NVLink, soldered on LPDDR5x memory. I think they will still be happy to sell IP to competitors making general purpose server CPUs.

 

[Atari2600]

I'm not so sure. This feels very special purpose to their GPU servers- NVLink, soldered on LPDDR5x memory. I think they will still be happy to sell IP to competitors making general purpose server CPUs.

Who would trust them?

As soon as the ecosystem develops (i.e. your company has done the hard yards), you know Nvidia will pull it all inhouse.

 

[moinmoin]

I think they will still be happy to sell IP to competitors

I really want to believe that, Nvidia's track record so far is just too one sided though.

 

[andermans]

Looking at the image the entire thing looks like an OAM module to me? AKA this is likely replacing their A100 based offerings instead of going straight for the host CPU in servers.

 

[Thala]

Which means the worst that was expected from this acquisition will probably happen. Typical nv property locked-gardens.

Because Nvidia dares to release a product based on ARM IP? I mean you did expect NVidia buying ARM and not using its IP?

This is a move to make NVidias server offering independent on AMD and to some extend on PCIE in the first place - which in the past was a weak point.

 

[itsmydamnation]

Because Nvidia dares to release a product based on ARM IP? I mean you did expect NVidia buying ARM and not using its IP?

This is a move to make NVidias server offering independent on AMD and to some extend on PCIE in the first place - which in the past was a weak point.

No he means look at their history,

xbox
PS3
vendor lock-in api after vendor locking api.
the apple saga

They have a reputation, and what people are saying is that based on past performance, come critical mass they would expect NV to marginalise other ARM server entrants.

 

[beginner99]

Because Nvidia dares to release a product based on ARM IP? I mean you did expect NVidia buying ARM and not using its IP?

This is a move to make NVidias server offering independent on AMD and to some extend on PCIE in the first place - which in the past was a weak point.

Of course them making ARM products was to be expected but they don't need to buy it for that in fact what's the point of taking on that burden when you can just get a license and be done with it?

Only reason to buy ARM is to control what new IP gets released, when it gets releases and what proprietary secret sauce vendors must include like nvlink so that you get an edge over your competitors.*

* EDIT:

To elaborate if all ARM server cpus are forced to contain some thing that greatly favors NV GPUs no one will put a Intel or AMD card in there. That way NV can get the whole ARM GPU server market for themselves. So it's a move to defend their compute cards from intel mostly and AMD.

 

[Thala]

No he means look at their history,

xbox
PS3
vendor lock-in api after vendor locking api.
the apple saga

They have a reputation, and what people are saying is that based on past performance, come critical mass they would expect NV to marginalise other ARM server entrants.

The comment from beginner99 was not related to NVidias history - this was know before - but related to the topic in question (e.g. the press release). The press release does give no indication in either direction, they just announced a product, which they could have as well without the acquisition.

 

[Thala]

Of course them making ARM products was to be expected but they don't need to buy it for that in fact what's the point of taking on that burden when you can just get a license and be done with it?

Only reason to buy ARM is to control what new IP gets released, when it gets releases and what proprietary secret sauce vendors must include like nvlink so that you get an edge over your competitors.*

As i said above. You might like conspiracy theories but above press release does not give any indication. They just replaced the typical PCIE link with nvlink - if i were NVidia i would have made precisely the same.
And no-one is stopping AMD from designing an ARM server architecture with Infinity-fabric either - they just appear to have no interest in such effort.

That way NV can get the whole ARM GPU server market for themselves. So it's a move to defend their compute cards from intel mostly and AMD.

And rightfully so. Neither AMD nor Intel supports the ARM ecosystem - so they are the last who should start crying if they were shut out of ARM servers as far their GPU offerings are concerned. And still an Integrator could plug-in their GPUs in the available PCIE slots - they would just be not particularly competitive compared to the nvlink offerings. In addition AMD/Intel would have provide the ARM64 drivers for their GPUs - as far they did not show any particular enthusiasm for this to say the least.

 

[coercitiv]

Neither AMD nor Intel supports the ARM ecosystem - so they are the last who should start crying if they were shut out of ARM servers as far their GPU offerings are concerned.

Sounds to me like you're also entertaining the idea of Nvidia GPUs being shut down in x86 space.

 

[Thala]

Sounds to me like you're also entertaining the idea of Nvidia GPUs being shut down in x86 space.

Read carefully. I am not even saying they are shut out technically, as PCIE is still existing - but the advantage is clearly on NVidia's side. I could entertain the Idea, that AMD is offering Infinity Fabric links to their GPUs - pretty similar to what NVidia is doing here.

 

[coercitiv]

Read carefully.

Write carefully.

they are the last who should start crying if they were shut out of ARM servers as far their GPU offerings are concerned

 

[moinmoin]

Neither AMD nor Intel supports the ARM ecosystem

AMD implements and supports ARM's TrustZone in all its Zen based processors, so you are technically wrong on that count already.

 

[Thala]

Write carefully.

I said:
'they are the last who should start crying if they were shut out of ARM servers as far their GPU offerings are concerned'

You probably have missed, that i used the subjunctive form - so again read carefully.

 

[DrMrLordX]

I'm not so sure. This feels very special purpose to their GPU servers- NVLink, soldered on LPDDR5x memory. I think they will still be happy to sell IP to competitors making general purpose server CPUs.

My expectation from the start is that nVidia wanted to do exactly as you indicated - control the server stack from top to bottom. Current Neoverse iterations will not be affected per existing licenses, yada yada ya. N1, V1, and possibly N2 will be untouched. But what do you think will happen to future Neoverse platform iterations? Beyond a certain point, when you license Neoverse, you will be required to support the nVidia stack in its entirety, or you won't license it. That's my prediction, anyway.

So when you get Neoverse, you get NVLink, and you get pushed into using NV's GPGPU accelerators. Or nothing else. No AMD cards, no Intel cards.

 

[Thala]

AMD implements and supports ARM's TrustZone in all its Zen based processors, so you are technically wrong on that count already.

Do we really dive into the difference of "supporting" and "using (parts of)" an ecosystem? If this is not understood i cannot help either.

 

[moinmoin]

Do we really dive into the difference of "supporting" and "using (parts of)" an ecosystem? If this is not understood i cannot help either.

Since you were stating that Intel and AMD are not in the position to complain as being not directly affected by Nvidia controlling ARM IPs, apparently yes. Your blanket statement is only true for Intel if at all.

 

[EVI]

Which means the worst that was expected from this acquisition will probably happen. Typical nv property locked-gardens.

I think its very presumptuous to assume that the deal will get approved. NVidia proposed to acquire ARM Holdings and they still need regulatory approval in the US, UK, EU and most importantly, China. Sure NVidia proposed remedies, but with customers in the US complaining to the DoJ and UK regulators already looking into the deal, geographies that use codified reasons on whether to approve or not could very well block the deal. To add onto that hurdle, China needs to approve, and as we've seen, they've blocked deals just because; see the Qualcomm/NXP deal. However, they approved Marvell/Inphi, pretty quickly especially since that combo will have an impact on 5G. So I think the road is very long for NVidia to get ARM done.

Creating a locked garden like x86 or App Store is typically legal when its built internally but creating one via acquisition is something that most regulators will block.

 

[coercitiv]

You probably have missed, that i used the subjunctive form - so again read carefully.

I did not miss that at all, I only applied your line of thinking into the x86 ecosystem as well. You have yet to answer, even in subjunctive form, whether Nvidia should start crying for being shut out of x86 as retaliatory measure.

Funny how a proponent of the ARM ecosystem on this forum is so easily charmed into the same vendor lock-in thinking that plagued x86 for decades.

 

[moinmoin]

I think its very presumptuous to assume that the deal will get approved.

I think the strong reaction against Nvidia's traditional handling of business in this thread so far is more reflective of an expectation/hope that the deal doesn't get approved than an assumption/fear that it does. At least for me.

 

[Thala]

I did not miss that at all, I only applied your line of thinking into the x86 ecosystem as well. You have yet to answer, even in subjunctive form, whether Nvidia should start crying for being shut out of x86 as retaliatory measure.
Funny how a proponent of the ARM ecosystem on this forum is so easily charmed into the same vendor lock-in thinking that plagued x86 for decades.

I never proposed that any vendor lock-in is good thing, independent of ARM or x86. I merely stated that IF we would come into a situation, where the GPU solutions from AMD and Intel were sort of incompatible with the direction the ARM ecosystem is moving to, then they should be the last to complain, because they could have participated - which they clearly have no interest in. I then followed up with the observation, that NVidia offering an nvlink solution does not lock-out anyone.

 

[moinmoin]

I never proposed that any vendor lock-in is good thing, independent of ARM or x86. I merely stated that IF we would come into a situation, where the GPU solutions from AMD and Intel were sort of incompatible with the direction the ARM ecosystem is moving to, then they should be the last to complain, because they could have participated - which they clearly have no interest in. I then followed up with the observation, that NVidia offering an nvlink solution does not lock-out anyone.

I think it's pretty clear to everybody now that you're really just looking at this announcement as is, while most others look at the implications this has for Nvidia's potentially impending ARM acquisition. Combining ARM IP with proprietary Nvidia IP dedicated to sell Nvidia accelerators, I'm sure most people don't mind that as is (and for me it's actually odd Nvidia didn't do that much earlier already, taking Denver some more productive way or jumping on Neoverse with N1 already). But that combined with the control over the whole ARM ecosystem Nvidia would get should the acquisition go through is worrying. Because exacting control over its products, IPs and ecosystem is something Nvidia traditionally likes to do.

 

[Thala]

I think it's pretty clear to everybody now that you're really just looking at this announcement as is, while most others look at the implications this has for Nvidia's potentially impending ARM acquisition.

Of course i am doing this, because this is the topic of this very thread and i am happily discussing implications of such a product in this context. However i totally object implications like

"Which means the worst that was expected from this acquisition will probably happen. "