[Rumor, Tweaktown] AMD to launch next-gen Navi graphics cards at E3

[crisium]

BTW, Navi could definitely match or even beat GTX 2080 Ti for 300W – after all, 5700 proves Navi, when clocked right, can match Turing PPW. It certainly will need more shaders though. There’s absolutely no reason to believe otherwise. That’s all the discussion is; that Navi needs more than 64CUs to match 2080 Ti. Not that Navi can’t do it, just that it needs more than 64CU.

I wonder how large and expensive a 72-80 CU Navi would be though.

 

[Glo.]

Eh? How are you the most combative reflexive poster on this forum, and still miss the obvious?

Look. At. The. Context. I. Talked. About. Clocks. Read. It. Again.

5700 clocks are efficient, 5700 XT are not. And regardless of efficiency at all, AMD cannot match Nvidia shader per shader right now.

Have you actually got what I was talking about?

You said shader per shader Navi is slower, despite the fact, it CLOCKS LOWER than Turing. At first glance it appears that shader per shader both architectures are exactly the same. But that would have to be confirmed by the tests. So far all we have is the clocks of RX 5700 in games which are around 1692 MHz, vs over 1.8 GHz for RTX 2060 Super. So that is pretty much significant clock speed difference which mitigates lower core count, of RTX 2060S.

 

[crisium]

Start with the post I was quoting, then read mine, then use the flow of logic. The original post was musing over whether a 300W 64 CU (4096) Navi could match a (2080 Ti) 4352 Turing.

We have a 2304 Navi at efficient clocks (matching Turing PPW) but it’s slower than a 2176 Turing. So a 4096 Navi at efficient clocks would match a 4352 Turing how? Impossible. You cannot even question this.

We have a 2560 Shader Navi clocked very high (worse than Turing PPW) and it’s still slower than a 2560 Turing. So a 4096 Navi even clocked very high beat a 4352 Turing how? It's possible if it's clocked to the moon maybe. But then how does it fit into a 300W envelope? When the 2560 Navi at those clocks is already ~220W?

Navi just needs more shaders to do it, but it is very possible as my last post said. 5700 proves it is possible, they just have to go wider. Based on AMD's love of 4096 shader flagships for the past several years, they might just go with it again though, and instead get a 2080 Super killer that fits between the 2080 Ti and 2080 Super.

 

[Glo.]

Start with the post I was quoting, then read mine, then use the flow of logic. The original post was musing over whether a 300W 64 CU (4096) Navi could match a (2080 Ti) 4352 Turing.

We have a 2304 Navi at efficient clocks (matching Turing PPW) but it’s slower than a 2176 Turing. So a 4096 Navi at efficient clocks would match a 4352 Turing how? Impossible. You cannot even question this.

We have a 2560 Shader Navi clocked very high (worse than Turing PPW) and it’s still slower than a 2560 Turing. So a 4096 Navi even clocked very high beat a 4352 Turing how? It's possible if it's clocked to the moon maybe. But then how does it fit into a 300W envelope? When the 2560 Navi at those clocks is already ~220W?

Navi just needs more shaders to do it, but it is very possible as my last post said. 5700 proves it is possible, they just have to go wider. Based on AMD's love of 4096 shader flagships for the past several years, they might just go with it again though, and instead get a 2080 Super killer that fits between the 2080 Ti and 2080 Super.

Dude, I do not care about what it needs to match RTX 2080 Ti.

You said that Navi is slower Shader per Shader versus Nvidia. I asked you about clock speeds. Then you went back about the context of what Navi needs, to match RTX 2080 Ti. Which I never talked about, because frankly, I don't care.

I'll use, for the sake of this discussion the data from TPU's reviews, that are separated by just 5 days.
RTX 2060 Super: 2176 ALUs with around 1.8 GHz core clock.
https://www.techpowerup.com/review/nvidia-geforce-rtx-2060-super/33.html
RX 5700: 2304 ALUs with around 1.692 GHz core clock in gaming.
https://www.techpowerup.com/review/amd-radeon-rx-5700/34.html

Do you still believe Navi is slower per ALU/Shader than Turing, considering it does not clock as high as Nvidia GPUs? It actually may have MORE performance per shader/ALU than Turing! But that would have to be confirmed by reviews, and tests of this, which may be very hard to say the least.

 

[crisium]

2070 Super Clock Speed is 1879MHz Average:

https://tpucdn.com/review/nvida-geforce-rtx-2070-super/images/clock-vs-voltage.jpg


5700 XT Clock Speed is 1887 Average:

https://tpucdn.com/review/amd-radeon-rx-5700-xt/images/clock-vs-voltage.jpg


They both have 2560 shaders. They have the same ROPs and memory config. They are running at virtually the same average clocks at TPU. The 2070 Super is 12% faster here though:

https://tpucdn.com/review/amd-radeon-rx-5700-xt/images/relative-performance_2560-1440.png


Turing has higher performance-per-shader.

 

[Glo.]

2070 Super Clock Speed is 1879MHz Average:

https://tpucdn.com/review/nvida-geforce-rtx-2070-super/images/clock-vs-voltage.jpg


5700 XT Clock Speed is 1887 Average:

https://tpucdn.com/review/amd-radeon-rx-5700-xt/images/clock-vs-voltage.jpg


They both have 2560 shaders. They have the same ROPs and memory config. They are running at virtually the same average clocks at TPU. The 2070 Super is 12% faster here though:

https://tpucdn.com/review/amd-radeon-rx-5700-xt/images/relative-performance_2560-1440.png


Turing has higher performance-per-shader.

And how do you explain the case of RTX 2060S, and RX 5700 and comparison between their performance per ALU, eh?

Because right now you took what you wanted to see in the world, posted here, and called it a day.

 

[Sunaiac]

Trying to buy an anniversary edition. Not yet on amd's site apparently. Anyone found one ?

OK available on AMD.com. Ordered.

 

[crisium]

And how do you explain the case of RTX 2060S, and RX 5700 and comparison between their performance per ALU, eh?

Because right now you took what you wanted to see in the world, posted here, and called it a day.

Yes, how convenient that I compared 2 cards that both have 2560 shaders at virtually the same clocks, 64 ROPs, and 448 GB/s bandwidth. I am really twisting the facts here.

Ok, let's take a look at your 2060 Super vs 5700.


Turing 2176 Shaders at 1839Mhz average:

https://tpucdn.com/review/nvidia-geforce-rtx-2060-super/images/clock-vs-voltage.jpg
https://tpucdn.com/review/amd-radeon-rx-5700/images/clock-vs-voltage.jpg

The Turing shaders have the same TFLOPs as 2304 Shaders at 1737MHz.

Navi 2304 Shaders at 1672MHz average:
https://tpucdn.com/review/amd-radeon-rx-5700/images/clock-vs-voltage.jpg

Turing (edit to fix) has 3.9% more TFLOPs in this comparison.

Yet 2060 Super is 8% faster, scaling beyond its TFLOPs:
https://tpucdn.com/review/amd-radeon-rx-5700/images/relative-performance_2560-1440.png

Turing has higher performance-per-shader.

 

[Guru]

2070 Super Clock Speed is 1879MHz Average:

https://tpucdn.com/review/nvida-geforce-rtx-2070-super/images/clock-vs-voltage.jpg


5700 XT Clock Speed is 1887 Average:

https://tpucdn.com/review/amd-radeon-rx-5700-xt/images/clock-vs-voltage.jpg


They both have 2560 shaders. They have the same ROPs and memory config. They are running at virtually the same average clocks at TPU. The 2070 Super is 12% faster here though:

https://tpucdn.com/review/amd-radeon-rx-5700-xt/images/relative-performance_2560-1440.png


Turing has higher performance-per-shader.

Navi is MUCH faster in DX12 and Vulkan in general, Nvidia still has slightly better DX11 architecture. It all depends how much DX12 is being adopted going forward, but judging from Steam hardware survey, 2 years ago win7 was at 50% rate and win10 at about 45% rate, these days win10 is at about 75% rate vs 25% rate of win7.

Navi has a much faster performance per shader in DX12 and Vulkan, though it obviously still loses on average in DX11 shader per shader performance.
I think it's inevitable that most future games (2020 and onwards) will be DX12 or Vulkan. Again with Nvidia's own push for ray tracing which requires DX12, I know that most games will be DX12 from 2020 and onward, especially as next gen consoles get released and they all use Navi architecture.

Great cards, to summarize everything up, the RX 5700 is essentially 5% faster than the RTX 2060 for the same price and consumes about the same power, while the RX 5700XT is 5% faster than the RTX 2060 super for the same price and a tad bit higher power consumption.

$350 RX5700 5% faster than RTX 2060, same power consumption. Clearly a much better buy, especially with the 8GB vs 6GB situation.
$400 RX5700XT 5% faster than RTX 2060super, 15W higher power consumption, a less clear choice, but taking into account Radeon Chill, FX fidelity, reduced latency in gaming, etc... the RX5700XT beats out the RTX2060s in value.

Temperature seems to be the only issue, but seeming as it is all over the place, it leads me to believe that some of the cards probably got slightly damaged in the transport and are not cooling the card properly. Its easy in these hot summer days and rocky rides for the thermal paste to lose a bit of contact and perform significantly worse.

Personally I'd take the RX 5700 and RX 5700XT over RTX 2060 and/or RTX 2060s because they are better value and because Nvidia is a scummy company that pushed gimpworks for decades that ruins performance in games in order to hurt the competition more, the GTX 970 3.5GB memory fiasco, the push to have partners use their "gaming" brands only with Nvidia and drop AMD(clearly an illegal move in Europe, anti competitive move) the GT 1030 with DDR4 fiasco with no way for consumers to know, essentially Nvidia committing fraud and a scam with the DDR4 version, and their endless push to increase prices to absurd levels, leading to $350 and $400 MID range cards.

 

[Glo.]

Yes, how convenient that I compared 2 cards that both have 2560 shaders at virtually the same clocks, 64 ROPs, and 448 GB/s bandwidth. I am really twisting the facts here.

Ok, let's take a look at your 2060 Super vs 5700.


Turing 2176 Shaders at 1839Mhz average:

https://tpucdn.com/review/nvidia-geforce-rtx-2060-super/images/clock-vs-voltage.jpg
https://tpucdn.com/review/amd-radeon-rx-5700/images/clock-vs-voltage.jpg

The Turing shaders have the same TFLOPs as 2304 Shaders at 1737MHz.

Navi 2304 Shaders at 1672MHz average:
https://tpucdn.com/review/amd-radeon-rx-5700/images/clock-vs-voltage.jpg

Navi has 3.9% more TFLOPs in this comparison.

2060 Super is 8% faster:
https://tpucdn.com/review/amd-radeon-rx-5700/images/relative-performance_2560-1440.png

Turing has higher performance-per-shader.

Its extremely, extremely stupid to make comparison of TFLOPs performance and how it translates to gaming performance.

Otherwise, GTX 980 would not be as fast, or faster than GTX 780 Ti in gaming, despite having almost 30% less theoretical compute performance(4 TFLOPs vs 5.7 TFLOPs).

Also, TFLOPs is a measure of theoretical COMPUTE performance. And games are not really prone to compute performance.

What did I said about you? You took what you wanted to see in the world, posted here and called a day.

So, once again. How do you REASONABLY explain the performance per ALU between RX 5700 and RTX 2060 Super. Which is exactly the same, at least at first glance?

 

[crisium]

Waiting for a shred of evidence that Navi has higher performance per shader than Turing. I've posted 2 comparisons; one of which refuted what you posted.

Are you actually admitting you're a troll now?

 

[Glo.]

Waiting for a shred of evidence that Navi has higher performance per shader than Turing.

And what do you have to show that it has higher? I claimed, every single time, that Navi has the same performance per ALU as Turing.

Every review shows that RTX 2060 Super and RX 5700 is the same in performance in most games. The outliers are skewing the overall performance metrics between those two GPUs in favor of RTX 2060 Super.

Also, how come people "dislike" poists they do not agree with on this forum? Its stupidest thing ever.

 

[crisium]

I want to be clear.

You are personally bearing witness to a 2560:64:448 (shaders, ROPs, bandwidth) Navi running at HIGHER CLOCKs than a 2560:64:448 Turing and yet being slower, and are still denying that Turing has better performance per shader?

 

[railven]

Because right now you took what you wanted to see in the world, posted here, and called it a day.

Really? Dude, you told people to stop using sources that didn't line up to the stuff you cherry picked. You are not one to make this kind of claim.

 

[Glo.]

I want to be clear.

You are personally bearing witness to a 2560:64:448 (shaders, ROPs, bandwidth) Navi running at HIGHER CLOCKs than a 2560:64:448 Turing and yet being slower, and are still denying that Turing has better performance per shader?

And I want to be clear. You are also not explaining how come RX 570 having 2304 ALUs, with 1692 MHz core clock, has the same performance as RTX 2060 S which has 2176 ALUs and 1.8 GHz average clock, despite, by your own omission, Turing having higher Per ALU performance.

Really? Dude, you told people to stop using sources that didn't line up to the stuff you cherry picked. You are not one to make this kind of claim.

And I still go by that. I do not use TPUs numbers from gaming performance, and power consumption to make any comparisons. Let alone their Performance Summary. I find their testing/calculating methodology to be misleading. Extremely.

 

[crisium]

The RX 5700 is slower than the RTX 2060 S. I have already demonstated to you that in no way does the 5700 vs 2060 S comparison show that the 5700 has better performance per shader. Would you like me to quote myself incase you forgot?

You keep mentioning it, but are not showing the numbers that explain how this comparison shows the Navi has better performance-per-shader (it does not).

 

[railven]

And I still go by that. I do not use TPUs numbers from gaming performance, and power consumption to make any comparisons. Let alone their Performance Summary. I find their testing/calculating methodology to be misleading. Extremely.

Considering all your posts in this thread, I'd be weary of where you get your information. But, that's just me.

 

[Paratus]

2070 Super Clock Speed is 1879MHz Average:

https://tpucdn.com/review/nvida-geforce-rtx-2070-super/images/clock-vs-voltage.jpg


5700 XT Clock Speed is 1887 Average:

https://tpucdn.com/review/amd-radeon-rx-5700-xt/images/clock-vs-voltage.jpg


They both have 2560 shaders. They have the same ROPs and memory config. They are running at virtually the same average clocks at TPU. The 2070 Super is 12% faster here though:

https://tpucdn.com/review/amd-radeon-rx-5700-xt/images/relative-performance_2560-1440.png


Turing has higher performance-per-shader.

AT has the 2070S at 5% faster relative performance while using 10% more power than the 5700XT. They are basically roughly equivalent.

The main issue is AMD needs 7nm to be “about equivalent”.

 

[Auer]

 

[crisium]

AT has the 2070S at 5% faster relative performance while using 10% more power than the 5700XT. They are basically roughly equivalent.

The main issue is AMD needs 7nm to be “about equivalent”.

AT's numbers are much better, yeah. Navi ahead in power efficiency which is very encouraging. I also see they recorded their clock rates.

https://www.anandtech.com/show/14586/geforce-rtx-2070-super-rtx-2060-super-review/15
https://www.anandtech.com/show/14618/the-amd-radeon-rx-5700-xt-rx-5700-review/15

2070 Super averages 1873.6 MHz and the 5700 XT averages 1837.2 MHz. So the 2070 Super is 2% higher clocked and 5% faster at AT. So AMD is only little behind in PPTFLOP, and a little ahead in PPW in that review.

Agree with the 7nm sentiment.

 

[guachi]

I'll deal with it by not buying it. I have a mini-TX setup and so I need cards that run much cooler than 90-some odd degrees otherwise my computer sounds like a hand drill.

One: it's a blower card so all the air is going out the back. A blower will actually keep your PC cooler.

Two: The temperature of the GPU is irrelevant. What matters is how many watts of energy are being dumped into your case.

 

[Glo.]

You keep mentioning it, but are not showing the numbers that explain how this comparison shows the Navi has better performance-per-shader (it does not).

https://www.computerbase.de/2019-07/radeon-rx-5700-xt-test/4/

What will you say now?

And lastly. I never claimed that it has higher per ALU performance. I said that it has the same performance per ALU as Turing. And that it MAY have higher performance per ALU, versus Turing, but that remained to be confirmed. Now it is.

 

[DrMrLordX]

Two: The temperature of the GPU is irrelevant. What matters is how many watts of energy are being dumped into your case.

Uh. Not with TSMC 7nm it isn't. I'm seeing the same thing on Matisse that all us Radeon VII owners experience: lower the temps, and your clockspeeds can go up. It's really important to keep AMD 7nm chips cool, GPU or CPU!

 

[Glo.]

Uh. Not with TSMC 7nm it isn't. I'm seeing the same thing on Matisse that all us Radeon VII owners experience: lower the temps, and your clockspeeds can go up. It's really important to keep AMD 7nm chips cool, GPU or CPU!

Two words: Heat Density.

 

[linkgoron]

The main issue is AMD needs 7nm to be “about equivalent”.

This is actually quite similar to the Polaris launch, when AMD got totally owned by 16nm Pascal (perf/watt). 14nm Polaris was "about equivalent" to 28nm Maxwell. AMD are looking decent vs the previous generation (Pascal) but also vs the latest release as Nvidia is still on a bigger node (12nm, which is essentially 16nm+ from what I understand).

IMO, as a product, it doesn't really matter why the 5700 is essentially equal in perf/watt to the 2060 and 2070S. The important thing is that the 5700 is competitive with Nvidia's latest cards which are launching at the same time.



However, unless either drivers get much better or AMD have a real breakthrough with RDNA2, AMD will get totally destroyed by Nvidia's next release.