** The AMD Navi Thread **

[CAT-THE-FIFTH]

I would wait and see how it performs and how it is priced. If its about RTX2070 level,or a bit better,and they price it say the same,I don't think it will sell well TBH,as RTX will be a value added feature. If they can push it closer to RTX2060 pricing then it has a fighting chance. If it is a Polaris replacement in launch price,it would be nice,but with GPUs nowadays both companies will only do the minimum they can get away IMHO.

Edit!!

Also,there were rumours of water cooled overclocked cards too. It makes me wonder whether the thermal dissipation of the die is reasonably high. Say if they get close to Turing level efficiency or a bit higher,and have smaller chips,it could be a bit hot running,like the HD4870/HD4890.

 

[bru]

Well there were rumours that Raja had fixed GCN or at least some of the issues with it and here we have it.

Introducing RDNA the new AMD graphics architecture that will power their GPU's going forward.

AMD High-Performance Gaming Updates
AMD unveiled RDNA, the next foundational gaming architecture that was designed to drive the future of PC gaming, console, and cloud for years to come. With a new compute unit10 design, RDNA is expected to deliver incredible performance, power and memory efficiency in a smaller package compared to the previous generation Graphics Core Next (GCN) architecture. It is projected to provide up to 1.25X higher performance-per-clock11 and up to 1.5X higher performance-per-watt over GCN12, enabling better gaming performance at lower power and reduced latency.

RDNA will power the upcoming 7nm AMD Radeon RX 5700-series graphics cards which feature high-speed GDDR6 memory and support for the PCIe 4.0 interface.

During the keynote, Dr. Su showcased the power of RDNA and one of the new AMD Radeon RX 5700-series graphics cards in a head-to-head comparison with a RTX 2070 card running a Strange Brigade gameplay demo, beating the competition delivering incredible ~100 FPS gaming.13

AMD Radeon RX 5700-series graphics cards are expected to be available in July 2019. Learn more at the AMD E3 livestream event on June 10, 2019 at 3 pm PT.

Foot notes

10: AMD APUs and GPUs based on the Graphics Core Next and RDNA architectures contain GPU Cores comprised of compute units, which are defined as 64 shaders (or stream processors) working together. GD-142
11: Testing done by AMD performance labs 5/23/19, showing a geomean of 1.25x per/clock across 30 different games @ 4K Ultra, 4xAA settings. Performance may vary based on use of latest drivers. RX-327
12: Testing done by AMD performance labs 5/23/19, using the Division 2 @ 25x14 Ultra settings. Performance may vary based on use of latest drivers. RX-325
13: Testing done by AMD performance labs 5/23/19, using the Strange Brigade @ 25x14 Ultra settings. Performance may vary based on use of latest drivers. RX-328

So still 64 shader, compute units, but improved to give more performance than before, as shown in certain circumstances.
the 1.25X boost across 30 games, even if tested at 4k with 4xAA, in my opinion is most telling and definitely a good thing.

Just what it will mean in general terms for us the enthusiast, we will have to wait and see.

Info taken from here.

Roll on July 7th.

 

[Participant]

Oh right so it is a new architecture after all ? Unless we're arguing that a new architecture means not using silicone this time or something.

I'll take 40FPS at UHQ 4K for £350.

 

[Poneros]

I'm keeping my eyes peeled, I want reasonable price upgrade for 4k 60 or minimum 4k40 consistently. Think new GPU not sure about any need to upgrade the ryzen 1600 at 3.7

I think v64 might still take it at 4k due to higher bandwidth, unless Navi really improves on that front.

I suspect the pricing on this will be very competitive. That die looked pretty damn small. GTX 2070 / 2060 are a 445mm2 die. 12nm may be an older, larger node than 7nm, but I highly doubt 12nm is cheap - it was a custom SHP shrink of 16nm designed purely for NVIDIA.

I'd guess the die that the 5700 product stack is using is less than half that size. I doubt the wafers are anywhere near double the cost, now that TSMC's 7nm is a year old, and of course yields will be much better for a small die than a big die.

Yes I think so too, on Anandtech they clarified that the 1.5x efficiency is both architecture and process, i.e. 7nm. Thus its competitiveness is not looking so great. I just can't imagine they will release what's a higher priced vega 64 with some efficiency improvements.

 

[bru]

Well at least we now know why NVidia are supposedly readying a RTX 2070Ti. It does kind of show that they knew where Navi was going to position its self in the market all along.

 

[Participant]

Well at least we now know why NVidia are supposedly readying a RTX 2070Ti. It does kind of show that they knew where Navi was going to position its self in the market all along.

Presume any 2070Ti will be ~£500? In which case Navi at £400 could still make sense.

 

[bru]

Presume any 2070Ti will be ~£500? In which case Navi at £400 could still make sense.

True, pricing will be key, but wasn't there rumours of a 2070Ti and price cuts as well.

 

[D.P.]

Oh right so it is a new architecture after all ? Unless we're arguing that a new architecture means not using silicone this time or something.

I'll take 40FPS at UHQ 4K for £350.

If you believe whatever AMD's marketing department says. In which car printing a different box is a new architecture as AMD claimed when rebranding the 290 and 390.


For a factual answer I prefer to look at the actual architecture, which is clearly GCN with double the shader engines

 

[Participant]

If you believe whatever AMD's marketing department says. In which car printing a different box is a new architecture as AMD claimed when rebranding the 290 and 390.


For a factual answer I prefer to look at the actual architecture, which is clearly GCN with double the shader engines

But didn't they have to do something to GCN in order to be able to double the shader engines, in which case at point does it constitute a new architecture in your eyes?

You might be right and they're just scared of the PR around using GCN again so are desperate to tout it as something else when it isn't really, but I genuinely don't understand what constitutes a new architecture if this doesn't.

 

[D.P.]

But didn't they have to do something to GCN in order to be able to double the shader engines, in which case at point does it constitute a new architecture in your eyes?

You might be right and they're just scared of the PR around using GCN again so are desperate to tout it as something else when it isn't really, but I genuinely don't understand what constitutes a new architecture if this doesn't.

They made changes, but They also made changes in VEGA, Polaris, Fiji, Hawaii...etc.

These changes are bigger than the last few for sure, and more than I expected.

The changes seem on par with Pascal with Turing, which I would have called an architectural evolution. But turing introduced 2 completely new compute units (tensor and RTX cores).

If you want to be generous you could call it a new architecture, but nevertheless it is still GCN through and through. But I really don't think that matters at all. Am evolution of EGCN that resolved any critical shortcomings is much more preferable than a new architecture that fails to live up to the hype like the 2900x and Bulldozer.

As I said in my very first post on the matter, the only thing consumers should care about is price and performance

 

[Poneros]

My worry is that the 1.25x perf adv is overstated. Testing with 4x AA is dubious, because what's 4x AA - most likely an msaa option, which is where AMD struggled but it's where the major improvements from RDNA might come from. Ok, so what's the problem? The problem is msaa has stopped mattering & isn't used much at all anymore. So is it still going to be 1.25x with AA off or some form of TAA?

 

[pmc25]

The cache hierarchy was already changed in Vega AFAIK and the CU to SP change with double the shader engines I already commented on. That was the natural extension of GCN, just the same way as nvidia changes the ratio of compute cores and CUDA units between generation to maintain efficiency with scaling.


You are othe one that seems to be upset at the fact that this is an evolution of GCN and not a whole architecture. I clear;ly stated it is irrelevant, and only the price and performance is what matters.

1) Minor changes. Much more major ones here, and much more of the cache. That's not a natural extension of GCN, it's a fundamental change to it. It's not just increasing the upper limit, but completely changing how the units work.

2) More trolling. I posted exactly the same thing 3 days ago, saying that the argument about whether Navi is 'GCN' is totally irrelevant as there were major changes in the architecture, so the name was moot. I also said I didn't expect to see much of an architectural change going to 'Next Gen' / Arcturus, but potentially either chiplets, or moving the design closer to being suitable for it, because it would totally ruin the synergy between PS5 / XBN / Stadia and PC if there was an all new architecture for 'Next Gen'.

 

[Stephanie Peterson]

Interesting that the first, and only, navi gpu we know of is the 5700... leaving quite the space above for a 3800 part...???? That could be 1080ti levels which if priced right would be nice.

That said AMD seem to have taking a page out of intel and nvidias book on pricing to some extent, so hopefully they could get out a 5800 part out for decent money.

 

[fs123]

Interesting that the first, and only, navi gpu we know of is the 5700... leaving quite the space above for a 3800 part...???? That could be 1080ti levels which if priced right would be nice.

That said AMD seem to have taking a page out of intel and nvidias book on pricing to some extent, so hopefully they could get out a 5800 part out for decent money.

the x700 series has always been a low-mid tier so there is a good chance of 5800 etc.

 

[Stephanie Peterson]

Yeah thats what i was thinking.
If this is a new arch, and not just GCN with bells on, then it could be a whole different ballgame.
Interesting that whotsherface said GCN would still live on in vega for professional applications, so AMD move to a two architecture setup for there gaming and pro cards at last..
This has been the problem with there chips for the last few years, GCN is/was great for compute and other professional stuff. But not for gaming where it was hobbling other aspects.

 

[Grim5]

My worry is that the 1.25x perf adv is overstated. Testing with 4x AA is dubious, because what's 4x AA - most likely an msaa option, which is where AMD struggled but it's where the major improvements from RDNA might come from. Ok, so what's the problem? The problem is msaa has stopped mattering & isn't used much at all anymore. So is it still going to be 1.25x with AA off or some form of TAA?

Just about everything these days uses TAA + Motion Blur

 

[Grim5]

But didn't they have to do something to GCN in order to be able to double the shader engines, in which case at point does it constitute a new architecture in your eyes?

You might be right and they're just scared of the PR around using GCN again so are desperate to tout it as something else when it isn't really, but I genuinely don't understand what constitutes a new architecture if this doesn't.

That’s a modified architecture not “brand new” like AMD claimed then.

It’s a good thing so that AMD doesn’t lose it’s compute advantage

It’s basically like Turing is a modified Pascal and not brand new

 

[pmc25]

There are clearly going to be one or more larger chips, and as I said before, rumours say soonish.

The big question, for me, is whether they'll be HBM2(.5?) or GDDR6. With the architectural changes I don't think it'll be nearly as bad a bandwidth hog as Vega.

They could keep memory config the same as the RX 5700, but double it to 16GB, and double the bandwidth, for the bigger chip. That would certainly be cheaper for both the memory itself, and packaging. I guess it depends how much bigger the big boy is.

I think it's unlikely there will be 2 larger chips (high end and super high end), as it looks like AMD want a new CPU and GPU product stack with new architecture every year now, or that's the aim. If that's the case they'd be better off iterating annually and getting closer to or taking the super high end crown that way, and keep each year's number of new dies to an absolute minimum. It's worked brilliantly for them with Zen.

Multi GPU is going to make a return in some form, sooner or later, though. Small chance it's with Navi? RX 5990?

 

[Dicehunter]

I'm curious to see if the Radeon 7 will already be replaced by AMD with this years Navi.

 

[pmc25]

I'm curious to see if the Radeon 7 will already be replaced by AMD with this years Navi.

If big(ger) Navi is coming at GamesCom, possibly. Will Vega 20 be phased out? No. But what do they do with the lower bins, then? Maybe process is sufficiently tuned now so as not to have too many lower bins not suitable for Instinct / Pro cards. In that case, definitely EOL.