AMD stealing Intel Server Marketshare?

[Pip313]

AMD just got a contract to supply one of Europe’s largest high performance computing (HPC) clusters. They will supply 4,600 epyc 7601 cpu’s.

 

“Düsseldorf and Tokyo, June 18, 2019 - NEC Corporation (NEC; TSE: 6701) today announced that Deutsches Zentrum für Luft- und Raumfahrt (DLR) - the German Aerospace Center - has awarded NEC with the delivery and installation of a new High Performance Computing (HPC) cluster solution, which constitutes one of the largest recent HPC tenders in Europe”

 

 

https://www.nec.com/en/press/201906/global_20190618_02.html

[FloRolf]

That is soooo cool! 

 

I wonder what the price difference for an comparable Intel system would be. 

[GoldenLag]

2 minutes ago, FloRolf said:

 

I wonder what the price difference for an comparable Intel system would be.

Price of a 28 core is like 30% more isnt it over a 7601?

 

It would cost about 48% more based on numbers i havent bothered looking up. You need 14% more CPUs to get to the same corecount

[porina]

37 minutes ago, GoldenLag said:

Price of a 28 core is like 30% more isnt it over a 7601?

 

It would cost about 48% more based on numbers i havent bothered looking up. You need 14% more CPUs to get to the same corecount

You can be sure they don't pay the same kind of pricing we see.

 

Also HPC is traditionally rated in FP64 performance, something AMD CPUs still lag behind Intel. AVX-512 offers about 4x the peak (IPC) performance of Zen, and Zen2 only halves that gap. AMD's counter would be to offer FP64 capability in GPU, but the link doesn't seem to mention any GPUs for compute nodes. It may be for their intended workload, this doesn't matter.

[Humbug]

Intel has resigned itself to losing the majority of high end desktop/workstation market share to AMD for now.

 

However they will defend their laptop and server market share very hard. In the laptop market they still have competitive products and AMD is not as aggressive, they are taking their sweet time in launching good APUs with 7nm Zen2 cores + Navi graphics.

In the server world I think AMD will get a few good percentage points of market share, which is big money for AMD.

[porina]

8 minutes ago, Humbug said:

Intel has resigned itself to losing the majority of high end desktop/workstation market share to AMD for now.

Intel may have lost the desktop crown, or will do in about 2.5 weeks, but I doubt they'll lose the "majority" of their share. For sure, significant future sales will go to AMD, but it isn't like everyone will stop buying Intel CPUs once 7/7 ticks over.

 

For HEDT/workstation, it is harder to say since neither AMD nor Intel have revealed their hands yet for the next generation. Especially as they already differ in their existing offerings, they're not a direct comparison against each other, and I suspect that will remain the case. It is easier to predict on AMD side they will just pack even more chiplets into their, while some have speculated Intel may switch to some variation of Cascade Lake for next HEDT. It's about all they have until 10nm server next year.

 

8 minutes ago, Humbug said:

However they will defend their laptop and server market share very hard. In the laptop market they still have competitive products and AMD is not as aggressive, they are taking their sweet time in launching good APUs with 7nm Zen2 cores + Navi graphics.

The APUs have lagged quite a while behind the desktop CPUs so don't hold your breath for this one. I haven't followed the rumours closely, but a question is if AMD will implement APU with chiplets, or go monolithic. 

[FlappyBoobs]

I don't know why people are talking about this like it's something new...the CPUs that are being used in this application were released in 2017! 

 

[leadeater]

36 minutes ago, porina said:

Also HPC is traditionally rated in FP64 performance, something AMD CPUs still lag behind Intel. AVX-512 offers about 4x the peak (IPC) performance of Zen, and Zen2 only halves that gap. AMD's counter would be to offer FP64 capability in GPU, but the link doesn't seem to mention any GPUs for compute nodes. It may be for their intended workload, this doesn't matter.

AMD is making up the difference through core count and memory bandwidth. It's also cheaper to deploy larger memory capacity nodes on AMD than it is on Intel, there's multiple ways where costing is saved which can be put in to increasing node count further increasing performance.

 

You still have to combat the fact that everything is so highly Intel optimized and uses Intel compilers so there is the safety factor in going with an Intel deployment. You would also go with Intel if you are targeting frequency bound applications because Intel just cannot be matched here, 400Mhz to 1000Mhz all core lead.

 

I can't say the specific configuration but an Intel Xeon 6154 based system with 'a lot of memory' is ~20k USD at the prices we pay. For the same number of cores as this EPYC deployment you would need 4089 nodes, 4089 * 20000 = 81780000 (81.78m) which is roughly 72 million Euro, just a tad higher than the 20+ million being talked about.

 

So how many can you get within the same budget? About 1000 nodes or 36000 cores versus 147000 from AMD which works out to be exactly (4.08) 4 times as many.

 

Edit: EPYC2 is even more in favor of an AMD deployment here too.

[leadeater]

5 minutes ago, FlappyBoobs said:

I don't know why people are talking about this like it's something new...the CPUs that are being used in this application were released in 2017! 

 

Because server markets move about as fast a glaciers lol

[Trixanity]

40 minutes ago, Humbug said:

Intel has resigned itself to losing the majority of high end desktop/workstation market share to AMD for now.

 

However they will defend their laptop and server market share very hard. In the laptop market they still have competitive products and AMD is not as aggressive, they are taking their sweet time in launching good APUs with 7nm Zen2 cores + Navi graphics.

In the server world I think AMD will get a few good percentage points of market share, which is big money for AMD.

There are rumors that APUs launch in 4-5 months. Said rumors also calls it a 7nm refresh though with implications that it might still be Vega for whatever reason. Grain of salt and all that.

[williamcll]

56 minutes ago, Trixanity said:

There are rumors that APUs launch in 4-5 months. Said rumors also calls it a 7nm refresh though with implications that it might still be Vega for whatever reason. Grain of salt and all that.

I thought it would remain as 12nm.

[Trixanity]

42 minutes ago, williamcll said:

I thought it would remain as 12nm.

Those are already out. So it wouldn't make sense to launch them again in 4 months.

[Humbug]

2 hours ago, Trixanity said:

There are rumors that APUs launch in 4-5 months. Said rumors also calls it a 7nm refresh though with implications that it might still be Vega for whatever reason. Grain of salt and all that.

Seems almost criminal to use Vega when you have RDNA/navi which needs far less CUs and far less die space to achieve the same performance, especially in an APU.

[Trixanity]

10 minutes ago, Humbug said:

Seems almost criminal to use Vega when you have RDNA/navi which needs far less CUs and far less die space to achieve the same performance, especially in an APU.

I'd be very disappointed if Vega would be coming along for 7nm. Keep in mind it's still a rumor and I haven't seen any evidence for it.

[porina]

24 minutes ago, Humbug said:

Seems almost criminal to use Vega when you have RDNA/navi which needs far less CUs and far less die space to achieve the same performance, especially in an APU.

Is navi actually more much compact than vega, if build on the same process?

[Humbug]

13 minutes ago, porina said:

Is navi actually more much compact than vega, if build on the same process?

For a given performance point it is a lot more compact. For example the RX 5700 has the same number of CUs as Polaris. But it is performing like a Vega 64. Even if we equalize the clockspeeds it doesn't come close to making up the difference of a 36 CU part performing like a 64 CU part... That's why AMD was promoting 25% greater performance per clock at the RDNA launch. It's a big architectural leap.

 

On an APU with limited die space which fits only 10-20 compute units you would much rather those be RDNA based rather than GCN based in order to get maximum work done.

[porina]

Just now, Humbug said:

For a given performance point it is a lot more compact. For example the RX 5700 has the same number of CUs as Polaris. But it is performing like a Vega 64. Even if we equalize the clockspeeds it would still be way ahead... That's why AMD was promoting 25% greater performance per clock at the RDNA launch. It's a big architectural leap.

 

On an APU with limited die space which fits only 10-20 compute units you would much rather those be Navi based in order to get maximum work done.

Since die area was bought up, that's the point of interest. If you made Vega on 7nm, how would it compare to navi at 7nm? That's the question. We can't even directly compare counts of units inside since the rebalanced it. It is easy to show an improvement if they're comparing it to 12/14nm parts so that needs to be taken out. Even the Radeon VII isn't a good comparison point since that has a lot more FP64 performance in it.

[Humbug]

1 minute ago, porina said:

Since die area was bought up, that's the point of interest. If you made Vega on 7nm, how would it compare to navi at 7nm? That's the question. 

I honestly don't know, could be the same or not much difference... 

 

2 minutes ago, porina said:

We can't even directly compare counts of units inside since the rebalanced it. It is easy to show an improvement if they're comparing it to 12/14nm parts so that needs to be taken out. Even the Radeon VII isn't a good comparison point since that has a lot more FP64 performance in it.

My point was that for gaming performance you need far less RDNA cores to achieve the same result as GCN cores. So on a given manufacturing node RDNA is going to be lower die space for a particular gaming performance point. Unless the RDNA compute unit itself is a lot bigger...

[porina]

11 minutes ago, Humbug said:

Unless the RDNA compute unit itself is a lot bigger...

I don't know how a navi core compares in content to a vega core. They did change the number of something inside them, which I'd have to look up separately for detail. We agree a core is not the same between them (in functionality) so whatever that difference is would likely manifest also in effective area if built to same process.

 

I suspect the navi cores are doing more of the things that matter to gaming, but I don't know if they cut any of the things that don't, as they probably want to keep it somewhat general.