Is it likely that high end Navi will be a redesign of Vega 64?

[g67575]

The 5700 XT has 2560 Shader Processors, the 14nm Vega 64 has SP 4096 SPs, the 7nm Radeon VII oddly has slightly less at 3840 SPs and some expensive HBM2 memory.

Does it seem likely that the process shrink to 7nm+ combined with RDNA redesign (v2) will allow high end navi to include upto 4096 SPs (plus more compute units), presumably with a lower TDP than Vega 64 and corresponding lower GPU core temps?

Also the Vega 64 has a chip die size of 495 mm², which clearly shows it's possible for AMD to design chips upto this size if needed, compare this to current gen. Navi with die of just 251 mm².

If all this is true, it suggests perf. improvement of between 33.3-40 % over the 5700 XT, resulting in similar performance to the RTX 2080 Super for high end Navi.

Speculation - HBM memory will be reserved for exclusive super expensive GPUs, like a Radeon VIII, because AMD can increase memory bandwidth with memory bus size increases instead, Vega also has a larger memory bus size than current Navi.

Unrelated fact - Vega was released in August 2017, maybe a clue to the release of high end Navi?

 

[Panos]

The 5700 XT has 2560 Shader Processors, the 14nm Vega 64 has SP 4096 SPs, the 7nm Radeon VII oddly has slightly less at 3840 SPs and some expensive HBM2 memory.

Does it seem likely that the process shrink to 7nm+ combined with RDNA redesign (v2) will allow high end navi to include upto 4096 SPs (plus more compute units), presumably with a lower TDP than Vega 64 and corresponding lower GPU core temps?

Also the Vega 64 has a chip die size of 495 mm², which clearly shows it's possible for AMD to design chips upto this size if needed, compare this to current gen. Navi with die of just 251 mm².

If all this is true, it suggests perf. improvement of about 33.3-40 % over the 5700 XT, resulting in similar performance to the RTX 2080 Super for high end Navi.

Speculation - HBM memory will be reserved for exclusive super expensive GPUs, like a Radeon VIII, because AMD can increase memory bandwidth with memory bus size increases instead, Vega also has a larger memory bus size than current Navi.

Unrelated fact - Vega was released in August 2017, maybe a clue to the release of high end Navi?

Vega & Navi are different architectures. That 495mm2 while feasible to manufacture, was too expensive and AMD was losing money on each one it sold. However would have been pointless to throw down the bin all failed MI50s with their substrate and HBM2 ram.

Around 429mm2 is the maximum TSMC can do at 7nm+ (EUV). Thats within the 15-20% higher density and physical size shrink over the 7nm process (max size around 500mm2). However such big chip would be expensive. Don't expect AMD been a charity and sell cards at loss like it did with RVII if they make such big chip.

RDNA2 is more efficient than RDNA1 so even a ~380mm2 chip will be over 50% faster than the 5700XT, which is well within the profitable yield zone.
Saying that I wouldn't be surprised if AMD has gone already the MCM route from this generation, not waiting to be forced by the 5nm process.

What we need to look at as good indicator of where things going is the Arcturus based MI100 launch isn't that far away. (Arcturus is Vega's evolution)

 

[g67575]

Your right, Vega GPUs did sell at loss...

However, one of the biggest costs of Vega was HBM memory and probably, the 2048-bit memory bus wasn't cheap either. Also, 14nm would've cost more to produce (than 7nm/7nm+).

Navi had none of these problems however.

Also, Radeon VII is 7nm and has nearly as many SPs (3840) as Vega 64, and that was achieved on die size of just 331 mm². Both Navi and Vega GPUs were produced at TSMC foundaries.

So, I think at least that number of SPs is possible to produce for next Navi GPUs.

 

[Finners]

I'll see if I can find a source but I thought that only the CPU's were going to be made on the 7nm+ process and the GPU's on 7nm like the 5700XT is currently.

Reason being that TSMC only have a certain amount of capacity and it's more profitable to be making EPYC with it than GPU's. I believe that's why we are still waiting now for that performance bracket GPU from AMD because it's better for business to be making CPU's that sell like hot cakes

 

[Rroff]

Saying that I wouldn't be surprised if AMD has gone already the MCM route from this generation, not waiting to be forced by the 5nm process.

At least as far as consumer cards go I can't see it - to do anything useful in games versus compute use there are some big architectural changes needed that almost certainly we'd see an evolution of over a generation first in terms of hardware tweaks and software development on non-MCM cards.

 

[Panos]

At least as far as consumer cards go I can't see it - to do anything useful in games versus compute use there are some big architectural changes needed that almost certainly we'd see an evolution of over a generation first in terms of hardware tweaks and software development on non-MCM cards.

We didn't need much changes to go to MCM CPUs though. The system sees the Ryzen 3000 as 1 processor not 2. Same applies even to the TR 2000 series.

 

[Panos]

I'll see if I can find a source but I thought that only the CPU's were going to be made on the 7nm+ process and the GPU's on 7nm like the 5700XT is currently.

Reason being that TSMC only have a certain amount of capacity and it's more profitable to be making EPYC with it than GPU's. I believe that's why we are still waiting now for that performance bracket GPU from AMD because it's better for business to be making CPU's that sell like hot cakes

GPUs also
https://www.techpowerup.com/263384/amd-to-debut-2nd-gen-rdna-architecture-in-2020

And the new Xbox at least is confirmed going to be 7nm EUV with RDNA2

 

[Rroff]

We didn't need much changes to go to MCM CPUs though. The system sees the Ryzen 3000 as 1 processor not 2. Same applies even to the TR 2000 series.

GPUs are a whole different kettle of fish - we've been able to do core to core with minimal penalty on CPUs since almost the start. With game rendering there are a whole load of considerations even if you can persuade developers to adopt explicit multi-adaptor techniques.

Being able to develop a GPU that can be seen as 1 by the system using MCM requires either completely changing how resources are addressed or spreading out a conventional monolithic design into its separate sub-systems and essentially having one big GPU just multiple chips that all handle a different bit of the process (kind of like the original Voodoo cards but with the chips on the same substrate) - you can't just magically have a chiplet design like Zen with GPUs.

 

[BigBANGtheory]

No it’s a different architecture, zero chance imho

 

[g67575]

The main thing that matters I think in performance terms is how many Shader Processors high end navi will have. I'm curious to know if it will have 4096 SPs like the Vega 64, or 3840 SPs like the Radeon VII? Or perhaps even less than this?

 

[greatash]

Your guess is as good as mine. Could have 56/3584 like the V56, 64/4096 like the V64 or a monster 80/5120 chip.

Any would be good depending on the price, the 56/3584 could perform in and around a 2080Super, the 64/4096 like a 2080ti and the 80/5120 to compete with 3080/3080ti. Sooner the better too, as if they wait too long I will just wait for the cards that come out after the next gen consoles as those usually always have a big performance boost to put PC well ahead of consoles again, those cards will usually last the whole gen too just like the gtx 970 did.

 

[D.P.]

We didn't need much changes to go to MCM CPUs though. The system sees the Ryzen 3000 as 1 processor not 2. Same applies even to the TR 2000 series.

So much ignorance in 1 post. And I am blunt about this because you repeatedly make the same ignorant comment and are repeatedly corrected yet you carry on in ignorance posting rubbish.

A Ryzen 3600 for example is seen as 6 physical core as 12 logical cores. Not 1. To do anything with all 6/12 cores/thread requires explicit programming from the developer. The developer does not see 1 CPU, they see 12.

You can do the exact same thing with Xfire and Dx12.


Conversely, the 570XT has 2560 cores and the developer only sees 1 GPU.

 

[D.P.]

The main thing that matters I think in performance terms is how many Shader Processors high end navi will have. I'm curious to know if it will have 4096 SPs like the Vega 64, or 3840 SPs like the Radeon VII? Or perhaps even less than this?

It is pretty irrelevant though. More cores that can do less work and runa t a lower clock with bottleneck issue due in inability balance workloads, vs few cores that are more efficient at higher clocks with better load balancing. The latter will probide much better performance in games at a higher efficiency.

AMD went down the brute force approach of shoving mroe and more SP in to each GPU but the architecture didn't scale as AMD required developers to support their architecture rather than Nvidia's approach of designing GPUS around what game developers actually do in reality.

 

[gthlm]

So much ignorance in 1 post. And I am blunt about this because you repeatedly make the same ignorant comment and are repeatedly corrected yet you carry on in ignorance posting rubbish.

A Ryzen 3600 for example is seen as 6 physical core as 12 logical cores. Not 1. To do anything with all 6/12 cores/thread requires explicit programming from the developer. The developer does not see 1 CPU, they see 12.

Obviously he is not talking about cores. He's talking about discrete processors. A Zen chiplet is really a discrete processor, and most Ryzen/TR have four of them. The IO die joins them all together and the system reports one CPU node.

I'm sure you know all this, so I have to assume that you're deliberately misrepresenting his argument in an attempt to make him look stupid...

 

[g67575]

One point that hasn't been said yet is - There is no reason why the high end Navi cards won't be just as efficient per shader processor as the current Navi / rdna v1. if anything, it is likely to be more efficient due to it using an improved RDNA design (an obvious attempt at marketing / jargon tbh, kind of reminds me of phrases such as 'business DNA' ) and the improved 7nm EUV process.

So, if the new gpu(/s) performance scales with the current gen (per SP) why wouldn't performance be at least equivalent to an RTX 2080 Super, if SP count can be increased by at least a third vs the 5700 XT? Especially if GPU core and memory clocks clocks are the same or higher.

I suppose the worst case would be similar performance to the Radeon VIIs but this seems unlikely (not to mention, a waste of AMD's time and effort) because the R7 is already very close to the 5700 XT in performance. This case would also assume lower core / mem clocks than current Navi, similar to the R7 or Vega 64.

Also, is load balancing of SPs really such an issue for current GPUs? Can't that be improved with new drivers?

 

[g67575]

According to the link below, the next gen Xbox will have a Navi GPU (presumably, RDNA 2 architecture) with 56 compute units, which apparently implies 3584 Shaders, which is a 40% increase in the number of shaders vs the 5700 XT. It also has a similar core frequency to current Navi GPUs at 1700MHz (The XT has a 'game clock' of 1755). And potentially, GDDR6 with a decent memory bandwidth of 560GB/s.

So, I think this would mean a 33.3% to 40% boost in performance if we see a similar PC GPU, perhaps called a 5800. Sorry for repeating this theory once again!!

Maybe we will get a 5800 XT as well thats ~10% faster than this released at the same time?

Next Xbox GPU info:
https://www.guru3d.com/news-story/amd-based-xbox-series-x-could-have-3584-shader-processors.html

Note - the PS5 GPU doesn't look that promising - if the info is correct, it has the same numer of compute units as the Radeon RX 5700, but with about 400Mhz higher GPU Core clock.

 

[greatash]

I cant see the Xbox having a full 56 CU Navi chip at 1700mhz, no console will want to draw 300w+ overall. Saying that it depends on the improvements to perf/watt with RDNA 2.0.

56 CUs at something like 1400-1500mhz would be my guess. 5700XT is incredibly power efficient at 1500-1600mhz drawing only around 100w.
Also remember big Navi is so late it wont be competing with Turing but Ampere, so if they match a 2080S that means it will only be competing with a 3060 or 3070.

 

[g67575]

But AMD never tries to compete with Nvidia's top end (they struggle enough in the mid-high end), people with that kind of money tend to buy Nvidia graphics cards. AMD might release a higher end card later and charge a premium for it, like they did with the Radeon VII.

Why would next gen. consoles be designed for 300w max, considering the top end models will be nearly as powerful as current top end PCs with higher power requirements?

I think you are right about the clock speed though, the 1700mhz is apparently for the boost clock, so the 'game clock' / real spec is probably 100/200 mhz less than this.

 

[greatash]

But AMD never tries to compete with Nvidia's top end, people with that kind of money tend to buy Nvidia graphics cards.

This has only been true for the past couple of years, AMD always had high end cards until they flopped with Fury vs 980ti and again with Vega vs 1080ti (and then they didnt even bother with Turing). R7 was a stop gap due to Navi being delayed again and to sell failed MI60 cards instead of chucking them in the bin.

I would hope that RDNA 2.0 will compete at the high end, in fact im absolutely sure of it. 5700XT is only 20% slower than a 2080S and 35% slower than a 2080ti, thats not a big gap to make up with all the time they've had.Everything points to RDNA 2.0 beating the 2080ti and competing with Ampere, just depends on the pricing they go for.

 

[gthlm]

I would hope that RDNA 2.0 will compete at the high end, in fact im absolutely sure of it.....Everything points to RDNA 2.0 beating the 2080ti and competing with Ampere...

Nobody would love that to be true more than me, but given what we actually know right now that's extremely speculative.

All we really have right now to base that assumption on are some unsubstantiated leaks of optimism within the RTG from some niche youtubers.

I mean, I kind of agree with you, but I'm trying not to get my hopes up