AMD Vega (FE and RX) Benchmarks [Updated Aug 10 - RX Vega 64 Unboxing]

[Muhammed]

I don't think you realize exactly how bad the GCN launch drivers were.

These two examples you gave were the exceptions, in the rest of the games, 7970 was well ahead of GTX 580. Besides GCN was a complete new departure for AMD from VLIW, some kinks had to be worked out. Vega is NOT a new architecture, it's new iteration of GCN with some major features added in, but the baseline remains the same. Maybe in new games it will pick up some performance courtesy of RPM, better handling of geometry ..etc, but don't expect sweeping changes in already released games.

 

[Veradun]

They should really get rid of the GCN name. So that people on forums could stop arguing about architectures basing claims on their names.

 

[leoneazzurro]

These two examples you gave were the exceptions, in the rest of the games, 7970 was well ahead of GTX 580. Besides GCN was a complete new departure for AMD from VLIW, some kinks had to be worked out. Vega is NOT a new architecture, it's new iteration of GCN with some major features added in, but the baseline remains the same. Maybe in new games it will pick up some performance courtesy of RPM, better handling of geometry ..etc, but don't expect sweeping changes in already released games.

According to these words, Pascal is still Kepler with some improvements.

 

[Qwertilot]

People do actually say that. It gets really rather irritating.

Even the most 'thrown together' ones of these involve objectively huge amounts of engineering effort to make them work. Getting the sorts of results that NV seem to do consistently in recent times - and AMD have done at times - is really hard.

 

[Glo.]

I keep hearing that but you are dreaming. Primitive shader will improve geometry performance, yes, but will take resources from the shader array, thus lowering the ALU throughput. Basically you use your resources for difference purpose. No free lunch here. 1080Ti will still be much faster...

How can culling usused parts of rendered scene take out resources from other parts of pipeline? Even what Mike Mantor said in the talk with Gamers Nexus proves that your thinking is incorrect.

 

[CatMerc]

You could trace modern NVIDIA architectures all the way back to Tesla G80. Should we be saying Pascal is just tweaked Tesla with new features?

If not, then neither is Vega.

 

[HurleyBird]

These two examples you gave were the exceptions, in the rest of the games, 7970 was well ahead of GTX 580. Besides GCN was a complete new departure for AMD from VLIW, some kinks had to be worked out. Vega is NOT a new architecture, it's new iteration of GCN with some major features added in, but the baseline remains the same. Maybe in new games it will pick up some performance courtesy of RPM, better handling of geometry ..etc, but don't expect sweeping changes in already released games.

Did you actually read the GCN launch review? Those results aren't exceptions. There are games that paint a bit of a better picture, sure, but there are also many others that show both cards similarly close. And even then, what we're looking at is the improvement relative to launch.

Besides that, you're blatantly goalpost moving at this point. You're the one who asserted that a 20% gain due to drivers is unprecedented. Now that that's been shown as incorrect, you've changed your argument to something along the lines of "Vega won't see the same improvement as Tahiti because it's a smaller change." Well, no doubt. VLIW->GCN is obviously a larger architectural change. No one is saying that Vega is likely to see the same 35-40%+ gains as Tahiti even though the current drivers are obviously immature, and even though Vega is the most significant change since VLIW->GCN, no one is arguing that it is a greater change. I don't think anyone, as far as I can recall, is even going as far as to say that Vega absolutely will see a 20% improvement. It's you who is saying that Vega absolutely will not, and then insulting the more rational people who don't deal in silly absolutes on top of that.

There actually isn't any harm in admitting that you were wrong on the internet, by the way.

 

[Glo.]

@Glo. I appreciate your enthusiasm for the underlying workings of GPU's , but we all do need to accept that while AMD has been technically superior in some ways before, it doesn't often translate into increased performance in mainstream products. Primitive Shaders very well may actually have a 40% performance increase in geometry bound scenarios, but with a 20% market share and not enough engineers to send to help game developers write this code it's also entirely possible we may never see a true implementation of Primitive Shaders. This is why people are fighting with you, because a lot of them are sick of AMD having great technology that doesn't translate into performance that matters to them. I personally have no problem with your optimism in regards to AMD's uarch prowess, I just think when you're the little guy and you have to ask for people to help you out, it usually doesn't work out well for you.

Vega is base for all of Next generation GPUs from AMD. Developers will have to optimize for Vega, regardless if they want it or not.

These two examples you gave were the exceptions, in the rest of the games, 7970 was well ahead of GTX 580. Besides GCN was a complete new departure for AMD from VLIW, some kinks had to be worked out. Vega is NOT a new architecture, it's new iteration of GCN with some major features added in, but the baseline remains the same. Maybe in new games it will pick up some performance courtesy of RPM, better handling of geometry ..etc, but don't expect sweeping changes in already released games.

On low - level, no GCN is not new architecture. It still has 4 cycle latency cadence, it still is 4 16 Wide SIMD in each CU, that is working with 64 KB Wavefront and fed by 256 KB Register File Size per 64 cores.

But L2 cache is bigger. 100% bigger. All of graphics engines are now clients of L2 cache, which changes pipeline dramatically, and requires rewriting of software you use, and makes your drivers not backwards compatible, without layer of abstraction. It has new memory controller type, with new memory paging system, which gives more complexity to your hardware.

Vega is new architecture layout. It is big enough change to call it new architecture.

 

[Glo.]

Guys, do you even know how is geometry processed during pipeline stages currently? You are already using Hull,Domain and Geometry shaders processed on ALUs .

Primitive shader is in fact combination of all of these merged to single simplified shader. So it's opposite to your claims, It's actually less ALU demanding than current approach

Precisely.

 

[Glo.]

I hate to break it to you, but you don't know how game development works. Developers will include a feature if they feel that the benefits outweigh the costs, nothing more and nothing less. Getting assistance from an IHV increases the benefit. You're looking at the subject as an absolute, but nothing is further from the truth. You have to judge each situation individually.

Its the entire point of Primitive Shaders. They not only increase the Geometry Throughput of Vega.

They also simplify the process of Game Development, and allow for saving time and money on designing worlds.

 

[Muhammed]

and then insulting the more rational people who don't deal in silly absolutes on top of that.

Rational doesn't describe a statement of a "30% faster than 1080Ti".

You're the one who asserted that a 20% gain due to drivers is unprecedented.

I never asserted that, my posts were directed at Glo for his irrational and overhyped analysis of the situation. For Vega to be 30% faster than 1080Ti, it will have to pick up at least 60% more performance.

And moving goal posts is the practice of several AMD sympathizers here, first it was Vega will trash TitanXp. then Vega FE have "Fiji" drivers, and magical fps increases will occur at launch, and now Vega is going to pick up 70% more performance later on. Yeah this is indeed silly. It's a never ending cycle of hyping a product AMD themselves are not.

Did you actually read the GCN launch review?

Ok, I will admit I didn't know enough about 7970 vs 580. Now I do thanks.

On low - level, no GCN is not new architecture. It still has 4 cycle latency cadence, it still is 4 16 Wide SIMD in each CU, that is working with 64 KB Wavefront and fed by 256 KB Register File Size per 64 cores.

Really? So GCN is not a new architecture compared to VLIW, but Vega "which is GCN" is a new architecture compared to GCN! Solid logic there.

 

[Glo.]

Really? So GCN is not a new architecture compared to VLIW, but Vega "which is GCN" is a new architecture compared to GCN! Solid logic there.

Tell this to software developers at AMD . And tell this to Game Developers, as well .

P.S. What do you have a proof for your statement that with properly optimized software, Vega will not be 30% faster than GTX 1080 Ti?

Let me give it to you this way:

Putting Fury X on 1.6 GHz core clock already would tie it with GTX 1080 Ti. How much faster will be Vega with all of the architecture improvements, 2.5 times higher Geometry Throughput and load balancing?

 

[Muhammed]

Putting Fury X on 1.6 GHz core clock already would tie it with GTX 1080 Ti. How much faster will be Vega with all of the architecture improvements, 2.5 times higher Geometry Throughput and load balancing?

A 1.6GHz Fury is barely 1080 level. That's what Vega is really.

P.S. What do you have a proof for your statement that with properly optimized software, Vega will not be 30% faster than GTX 1080 Ti?

So now I am the one who has to provide a proof? You see how silly this is? You pull out numbers from your imagination without any shred of evidence or trace of knowledge and I am the one who should provide the proof your claim is not true?

This keeps getting better and better. There is no end for Vega reigns supreme in here I guess.

 

[Glo.]

A 1.6GHz Fury is barely 1080 level. That's what Vega is really.

So now I am the one who has to provide a proof? You see how silly this is? You pull out numbers from your imagination without any shred of evidence or trace of knowledge and I am the one who should provide the proof your claim is not true?

This keeps getting better and better. There is no end for Vega reigns supreme in here I guess.

Yes, if you are a part of discussion it would be good to provide proofs for your statements. I brought them. All you do is this:

Vega will not be faster because it won't be faster.
Nvidia is better because it is better.

1.6 GHz Fury X will be on par with GTX 1080 Ti. 1.6 GHz Vega with Fiji layer of abstraction will not be faster, because it is bottlenecked by the drivers. It is this simple. That is also why Clock for clock, core for core, we have seen in Vega FE reviews decrease per clock compared to Fury X in gaming.

 

[PhonakV30]

A 1.6GHz Fury is barely 1080 level. That's what Vega is really.

So now I am the one who has to provide a proof? You see how silly this is? You pull out numbers from your imagination without any shred of evidence or trace of knowledge and I am the one who should provide the proof your claim is not true?

This keeps getting better and better. There is no end for Vega reigns supreme in here I guess.

not sure , He's troll or serious.

 

[HurleyBird]

Ok, I will admit I didn't know enough about 7970 vs 580. Now I do thanks.

Thanks. Although that's not really relevant. What is relevant is that large performance increases via drivers do have precedent.

In the interest of playing by my own rules, the entire "7970 nipping at the heels of OG Titan" is admittedly something of an exaggeration.

Rational doesn't describe a statement of a "30% faster than 1080Ti".

I never asserted that, my posts were directed at Glo for his irrational and overhyped analysis of the situation. For Vega to be 30% faster than 1080Ti, it will have to pick up at least 60% more performance.

I'll admit that now that I've read them, Glo's predictions do seem a bit... optimistic. Not out of the realm of possibility, but maybe not exactly probable. Despite that, you do seem to be saying that people who think that Vega can gain a 20% improvement through drivers need to "grow up":

I would be a fool to believe each new piece of hardware is going to pick 20% more performance through drivers. That's just unprecedented and just doesn't happen. FX 5000 didn't manage it, HD 2900 didn't manage it. Grow up

 

[Tup3x]

Tell this to software developers at AMD . And tell this to Game Developers, as well .

P.S. What do you have a proof for your statement that with properly optimized software, Vega will not be 30% faster than GTX 1080 Ti?

Let me give it to you this way:

Putting Fury X on 1.6 GHz core clock already would tie it with GTX 1080 Ti. How much faster will be Vega with all of the architecture improvements, 2.5 times higher Geometry Throughput and load balancing?

The fact is that AMD compares Vega to GTX 1080. GTX 1080 Ti is clearly ahead. It just doesn't make much sense to make claims that it would suddenly start beating GTX 1080 Ti by a massive margin when it has trouble beating GTX 1080. This is the Polaris episode all over again ("it will beat GTX 980 Ti").


By the way, architectures that need heavy optimisations in order to perform well tend to fail and become obsolete. Devs just don't want to spend extra time if that doesn't help the majority at all. It's a bit different story if hardware vendor sponsors them but still. Majority of games will not get any miracle optimisations. Even in games that do get 'em Vega is not going to beat GTX 1080 Ti by 30%.

Making up ridiculous expectations doesn't help AMD at all.

 

[Muhammed]

What do you call:

I meant "each" as in each new GPU, CPU ..etc will pick up 20% more performance. I meant as in an inescapable fact of life. I might have worded it wrongly though, and I apologize for that. English is not my first tongue. But still my statement doesn't even begin to compare to his claim Vega will be 70% faster out of thin air.

In the interest of playing by my own rules, the entire "7970 nipping at the heels of OG Titan" is admittedly something of an exaggeration.

Thanks again

1.6 GHz Fury X will be on par with GTX 1080 Ti.

Proof? test? benchmarks? anything? What we do have is Vega with all the improvements over Fiji and with a clock speed of 1.6GHz, and according to AMD's own estimations is barely 1080 level, but now we have you here claiming Fiji and Vega and perhaps even Polaris is 1080Ti level.

This is has clearly descended into the territory of madness.

 

[Krteq]

What we do have is Vega with all the improvements over Fiji and with a clock speed of 1.6GHz, and according to AMD's own estimations is barely 1080 level...

Can you please post a link to that? As I know, AMD never told us it's "barely 1080 level". Only performance figures they showed were comparison of minimum framerates for Adaptive-sync/G-sync.

 

[Qwertilot]

Their marketing department might not be hyper efficient but they're not insane. If it beat the 1080 in a meaningful sense, even on a few cherry picked games, they'd have said that.

 

[Muhammed]

Can you please post a link to that? As I know, AMD never told us it's "barely 1080 level". Only performance figures they showed were comparison of minimum framerates for Adaptive-sync/G-sync.

Gladly.

what we’ve been told is to expect the Vega 64 to “trade blows” with NVIDIA’s GeForce GTX 1080.

http://www.anandtech.com/show/11680...rx-vega-64-399-rx-vega-56-launching-in-august

 

[Krteq]

Heh, so nothing concrete then. Well, we will see in less than a week

 

[Glo.]

In DX11 games and current slack of games not optimized for Vega - this is what we will see. I have written this, before. Games have to be rewritten to use Vega features, that bring highest improvement in geometry throughput. In DX11 and current slack of games we will see GTX 1080 level of performance. When games will be updated to use Primitive Shaders and Programmable Geometry Throughput - it will be faster than GTX 1080 Ti.

If Vega has higher core throughput than Nvidia GPUs. If Vega has higher Geometry Throughput than Nvidia GPUs. If Vega has the same type of graphics pipeline as Nvidia GPUs.

Why does anyone believe that Vega will not be faster than GTX 1080 Ti? On hardware level there is absolutely no reason why Vega would not be faster than GTX 1080 Ti.

 

[NickyVida]

https://www.facebook.com/Bnhtpl

Sapphire Radeon RX VEGA 64 8GB HBM2 HDMI/Triple DP / Liquid Cooled
1677MHz / 1406MHz / 945MHz
SRP : S$1149/- (846 USD)

Sapphire Radeon RX VEGA 64 8GB HBM2 HDMI /Triple DP / ( Air-Cooled )
1546MHz / 1247MHz / 945MHz
SRP : S$999/- (734.81 USD)

Welp. It'd better perform close to a 1080Ti for that pricing. Else it's just gonna be DOA.

 

[Krteq]

Where is this from? Can you guys start to post a links please?