AMD Pledges to Take On NVIDIA’s High-End Turing With 7nm Radeon GPUs in 2019

[LePhuronn]

I think the 590 is the result of global's attempt at Polaris 30 pre 7nm Navi.

Sorry, I don't what you mean.

Are you suggesting GloFo wanted to try their hand at a GPU on 12nm given they've sacked off their own 7nm developments? I can't think why because all GPUs from this point on will be 7nm. GloFo know their 12nm is good because that's what Zen+ is built on.

So in the absence of a full RX 600 line, and the specs of the RX 3060 in the leak, I can't see Polaris 30 now being anything other than padding out a contractual agreement.

 

[LePhuronn]

The hype train for Zen 2 is so up in the sky its not even funny anymore. I for one don't want coming down crashing when real numbers are revealed

Well we have seen cherry-picked numbers for EPYC Rome, so we do have a tiny inkling of what Zen 2 can do. It's not like this particular hype train is 100% in the dark...it's about 95%

 

[D.P.]

The only hype train is on forums with people believing every piece of "info" thats posted. Then if the end product is anything under the hype its somehow a disappointment.

What about when that information comes directly from AMD?

 

[muziqaz]

Well we have seen cherry-picked numbers for EPYC Rome, so we do have a tiny inkling of what Zen 2 can do. It's not like this particular hype train is 100% in the dark...it's about 95%

I somehow doubt that little demo was cherry picked I would say it was underperforming big time, as it showed barely outperforming their own dual socket same core count overall Naples system. They said they widened their AVX resources and done some bits and bobs, yet in their demo they did not showcase any of that, they just showed that core for core Rome is just a bit faster than Naples
On the other hand when compared to Intels chip, anything not using AVX512 (or any AVX workloads) can be called as cherry picked

 

[D.P.]

Or, If they do get the perf per watt back it will suddenly have little meaning

This seems unliekly given what Vega20 tells us. It pulls 300w and is slower than the 300w NVLINK TV100 Volta Tesla, which is built on 12nm (aka 16nm).

Despite all the 7nm hype the node is not actually that amazing. TSMC suggests about a 25% performance increase for the same power, or a 50% power reduction for the same performance, This is what we see with Vega 20.

We know Navi is still GCN but the last GCN iyeration, so it is not liekly to offer any significant architectural changes that would alleviate the efficiency issues. Looks hat AMD achieved with Fiji and Vega 10.


To get the performance close to a 1080ti the power will likely go up if anything, and if AMD want a more realistic power profile and die size for profitability, then performance will likely be nearer 1080.

As AMD themselves said, this will bring Navi close to Turing performance but using a whole new node. As soon as Nvidia jumps to 7nm then AMD will once again be that 1.5 to 2 generations behind.

 

[Gerard]

What about when that information comes directly from AMD?

I still think you're better off waiting for reviews.

 

[muziqaz]

I still think you're better off waiting for reviews.

Oh, yeah, for sure. I'm just saying it did not make much sense with their demoed performance. I do believe Rome was clocked much lower than Naples. Naples was dual socket, which adds performance penalty a well. Though on the other hand, if Rome was drawing similar amount of power than single chip Naples, but has double the cores, then the demo is a bit more impressive, since now you can downsize your server to single socket, and halve the power usage while remaining at same core count and similar performance as dual socket Naples.
Just to be clear my interest is in architectural improvements, more than in overall package

 

[pete910]

*snip.

It was a joke response to Jedi.
I never said they would either, I said IF

 

[LePhuronn]

I somehow doubt that little demo was cherry picked

OK, cherry picked as negative connotations. What I meant was the demos and numbers would have been carefully selected to present EPYC Rome in the best light.

 

[LePhuronn]

As AMD themselves said, this will bring Navi close to Turing performance but using a whole new node. As soon as Nvidia jumps to 7nm then AMD will once again be that 1.5 to 2 generations behind.

Which is why all eyes will be on Arcturus a mere 12 months later. And because Navi won't really trouble the RTX line too much, Nvidia will hold off on their own 7nm until 2020, by which time Arcturus is supposed to be on 7nm+

AMD's return to high-end competition is with Arcturus, not Navi. We'll see end of 2020 if they can pull it off, unless this rumoured Navi 20 is a thing, in which we might see it a bit sooner.

 

[chaparral]

It won't be happening on VGA, AMD cannot afford to your also forgetting a major factor, a big cost of graphics cards is the memory and a GPU cost a lot more money to make than a CPU, a lot more.

Someone needs to tell intel that as there CPU prices can easily match even NVidia sky high GPU prices...

 

[Brazo]

AMD's return to high-end competition is with Arcturus, not Navi. We'll see end of 2020 if they can pull it off, unless this rumoured Navi 20 is a thing, in which we might see it a bit sooner.

I heard the same old stuff when Vega was released. Navi was meant to be its saviour. I love my Vega card but I’m not under the impression that AMD are going to be able to compete at the high end probably ever again.

 

[LePhuronn]

...but I’m not under the impression that AMD are going to be able to compete at the high end probably ever again.

lol OK. I remember similar being said about Bulldozer, Piledriver, etc, and now Intel are about to get utterly spanked for a few years. I didn't say Arcturus will return AMD to the top end, I said Arcturus is their gateway to that competition and we'll see if they can actually pull it off in 2020. And bear in mind that Arcturus is a brand new architecture, so we have no idea what it'll do. Much like Ryzen was brand new.

 

[4K8KW10]

lol OK. I remember similar being said about Bulldozer, Piledriver, etc, and now Intel are about to get utterly spanked for a few years. I didn't say Arcturus will return AMD to the top end, I said Arcturus is their gateway to that competition and we'll see if they can actually pull it off in 2020. And bear in mind that Arcturus is a brand new architecture, so we have no idea what it'll do. Much like Ryzen was brand new.

With the help of chiplets, it will be easy for AMD to build a ultra-high performance graphics card. They also need to find a way to introduce again CrossFire.

 

[Gregster]

It would be great to see AMD catching up to NVidia. We can all see what happens when NVidia has it all it's own way.