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

[JDG1980]

AMD need a post GCN architecture. That is painfully obvious. Going back to the drawing board with a clean sheet is the only way out. There are risks but thats the only option for long term survival of RTG.

GCN is a compute-first architecture, and in that regard, Vega still looks to be competitive. For gaming, it was competitive with Kepler, but Maxwell and Pascal eclipsed it.

It's important to consider that, at this point, Nvidia's professional and gaming cards are basically two different architectures. Sure, GP100 and GP102/4/6/7/8 are all called Pascal, but GP100 is very different from the others in nontrivial ways. There's every reason to think the same will be true of Volta as well. The problem is that AMD, unlike Nvidia, almost certainly can't afford to develop two new GPU architectures. There are three other possible ways they could proceed:

1. Try to develop a new "jack-of-all-trades" architecture that can beat Volta (and whatever comes after) in both gaming and professional applications. This would be very challenging, especially with limited R&D resources. One of the major reasons why Maxwell took such a leap forward is that it basically jettisoned everything (such as hardware scheduling) that wasn't vital for gaming. A "jack-of-all-trades" architecture is bound to look a lot like GCN and to have many of GCN's shortcomings.
2. Keep GCN as the compute architecture and continue to do minor refinements, die-shrinks, etc. on that for future compute cards, but develop a new ground-up architecture for gaming cards. This would basically be the "two architecture" strategy, and it's questionable if AMD could afford to do it. It's not just hardware R&D but also software development that would be stretched too thin - AMD's driver team is clearly incredibly overworked as things already are.
3. Keep refining GCN for compute and just give up on high-end gaming. This seems to be the strategy that AMD has chosen. From a business perspective, it makes sense - I just wish they didn't have dishonest marketing that would lie about it all the time.

 

[Veradun]

I don't know why you people complicate matters with Nvidia's lineup so much. There are basically 4 tiers:

I. Low-end
GT 1030
GTX 1050

II. Mid-range
GTX 1050 Ti
GTX 1060

III. High-end
GTX 1070
GTX 1080

IV. Enthusiast
GTX 1080 Ti
Titan Xp

nVidia thinks 1080 and 1080Ti are on the same tier:

http://www.nvidia.it/graphics-cards/geforce/pascal/compare-specs/

So the people in this very thread that are saying they aren't happy but have to go with Vega because they already have Freesync displays aren't locked in.

https://en.wikipedia.org/wiki/Confirmation_bias

Nope.

Last Christmas I wanted a new monitor. I checked prices and went for the freesync version as the gsync version would have cost me far more as the former is a royalty free standard and the latter a proprietary standard with royalties I have to pay for.

Now I want a new card. I have the choice of choosing the card that support the royalty free standard or the card that doesn't support it because it has to push the consumer to pay 200€ more.

Never in the world I would pay those bucks. If I want an adaptive sync monitor + gpu combo the only reasonable choice for me is to spend 200€ less to have the same result.

 

[coercitiv]

Keep refining GCN for compute and just give up on high-end gaming. This seems to be the strategy that AMD has chosen. From a business perspective, it makes sense - I just wish they didn't have dishonest marketing that would lie about it all the time.

Thumbs up. The only thing that defies logic is their marketing.

 

[xpea]

Whatever way you slice it, NVIDIA is the one responsible for segregation in the VRR market.

Nvidia created this market, G-Sync monitors were available 2 years before Adaptive sync ones and they have 70% of the gaming market. Sorry nothing wrong here on their side. It's the benefit of being the leader and they simply protect their investment.

 

[SpaceBeer]

But if you are convinced your standard is much better, than why don't you support both and let users decide. Even if they only supported FS with LFC, it would be much better for users. Even though better FS(2) monitors are not so much cheaper than G-sync ones (more like $50 than $200).

If users (gamers) only needed the best possible experience, low-end and mid-range products wouldn't exist. Some people don't think 4K HDR, 60+ FPS at Ultra settings is necessary to enjoy the games

 

[PeterScott]

But if you are convinced your standard is much better, than why don't you support both and let users decide. Even if they only supported FS with LFC, it would be much better for users. Even though better FS(2) monitors are not so much cheaper than G-sync ones (more like $50 than $200).

If NVidia supported FS, there would never be another monitor built with GS. Manufacturers don't like paying royalties. So it would be less royalties for NV. Simple as that.

 

[CatMerc]

Nvidia created this market, G-Sync monitors were available 2 years before Adaptive sync ones and they have 70% of the gaming market. Sorry nothing wrong here on their side. It's the benefit of being the leader and they simply protect their investment.

Doesn't change what I said at all. Sure NVIDIA pioneered it, and I thank them for that. Doesn't mean that they aren't responsible for segregation in the current market.

 

[Glo.]

Nvidia created this market, G-Sync monitors were available 2 years before Adaptive sync ones and they have 70% of the gaming market. Sorry nothing wrong here on their side. It's the benefit of being the leader and they simply protect their investment.

70%?

There is 30 G-Sync monitors in the market, and the number of Freesync monitors is coming to 100.

And those G-Sync monitors cost MUCH MORE, than Freesync monitors.

yeah, G-Sync has 70% of market share, for sure.

 

[thilanliyan]

^I think he meant 70% of the discrete GPU market.

 

[Malogeek]

Yeah and there are likely a lot more people with Freesync monitors on Nvidia GPUs than G-sync.

 

[raghu78]

GCN is a compute-first architecture, and in that regard, Vega still looks to be competitive. For gaming, it was competitive with Kepler, but Maxwell and Pascal eclipsed it.

It's important to consider that, at this point, Nvidia's professional and gaming cards are basically two different architectures. Sure, GP100 and GP102/4/6/7/8 are all called Pascal, but GP100 is very different from the others in nontrivial ways. There's every reason to think the same will be true of Volta as well. The problem is that AMD, unlike Nvidia, almost certainly can't afford to develop two new GPU architectures. There are three other possible ways they could proceed:

1. Try to develop a new "jack-of-all-trades" architecture that can beat Volta (and whatever comes after) in both gaming and professional applications. This would be very challenging, especially with limited R&D resources. One of the major reasons why Maxwell took such a leap forward is that it basically jettisoned everything (such as hardware scheduling) that wasn't vital for gaming. A "jack-of-all-trades" architecture is bound to look a lot like GCN and to have many of GCN's shortcomings.
2. Keep GCN as the compute architecture and continue to do minor refinements, die-shrinks, etc. on that for future compute cards, but develop a new ground-up architecture for gaming cards. This would basically be the "two architecture" strategy, and it's questionable if AMD could afford to do it. It's not just hardware R&D but also software development that would be stretched too thin - AMD's driver team is clearly incredibly overworked as things already are.
3. Keep refining GCN for compute and just give up on high-end gaming. This seems to be the strategy that AMD has chosen. From a business perspective, it makes sense - I just wish they didn't have dishonest marketing that would lie about it all the time.

Actually I think AMD can build separate chips for gaming and for HPC / Deep learning. AMD is most likely to move to a multi die GPU using Infinity Fabric with Navi generation. AMD will need 3 dies - The stack would look like this

Navi 11 - 100 sq mm (Compete with Nvidia 107 series)
Navi 10 - 200 sq mm (Compete with Nvidia 106 series)
Navi 10 x 2 - 400 sq mm (Compete with Nvidia 104 series)
Navi 10 x 4 - 800 sq mm (Compete with Nvidia 102 series)

Navi 20 HPC die x 4 - (Compete with Nvidia 100 series)

The key here is the building block must be extremely efficient for area and power like the Zeppelin die is for Ryzen, Threadripper and EPYC. In fact if AMD is smart they should look to have APUs replace Navi 11. The key requirement for that is APUs with HBM. This has long been the prerequisite for APUs to replace the entry level discrete GPU. One thing is sure - a post GCN architecture which is extremely efficient and scalable is needed for both AMD's future GPU roadmap and the next gen consoles which should arrive in 2020/2021.

 

[Dribble]

Yeah and there are likely a lot more people with Freesync monitors on Nvidia GPUs than G-sync.

And most of them have rubbish freesync. Even the monitor amd is advertising specifically for freesync in the Vega discount is useless (having and effective 80-100hz vbr range). I think if you removed all the freesync monitors that didn't work properly, and I define that by meeting the gsync min spec which all gsync monitors meet (i.e. 30-max hz range, overdrive support, etc) you'd find the numbers are completely different.

 

[Tup3x]

It looks like VEGA FE:s performance advantage in professional apps just came to an end. NVIDIA released new drivers (385.12) that do not artificially hinder the performance (at least for Titan Xp, not sure about the rest). It should now at least make performance easier to compare. In any case though, consumers win.

 

[nathanddrews]

It looks like VEGA FE:s performance advantage in professional apps just came to an end. NVIDIA released new drivers (385.12) that do not artificially hinder the performance (at least for Titan Xp, not sure about the rest). It should now at least make performance easier to compare. In any case though, consumers win.

Called it.

From NVIDIA:

TITAN Xp is the most powerful GPU you can put in your PC, and now we’re enabling even better performance with our latest TITAN drivers.

We built TITAN Xp for people who design and create — and, of course, play games. And it’s always getting better.

Our latest driver — available today — delivers 3x more performance in applications like Maya to help you create and design faster than ever.

You can now also use TITAN Xp with more PCs, including thin & light laptops, than ever before, and with external GPU chassis from a wide range of manufacturers including Asus, HP, Powercolor and Razer

And when you’re done if you want to rip through Battlefield 1 at 100 frames per second – no problem.

Return shots fired!

 

[guachi]

So nVidia was intentionally crippling their product?

 

[Crumpet]

So nVidia was intentionally crippling their product?

Yes.

Would be interesting to test them head to head now.

 

[Malogeek]

Does this cause conflict with the performance/price of their Quadro product? Are they kicking themselves in the foot? I know Titan isn't certified but Nvidia has always been one to cripple hardware/software to ensure product space.

 

[PeterScott]

Does this cause conflict with the performance/price of their Quadro product? Are they kicking themselves in the foot? I know Titan isn't certified but Nvidia has always been one to cripple hardware/software to ensure product space.

It's the exact same thing AMD does, and changes with the the Frontier Edition. This brings them back into parity.

They now have expensive card that performs better on workstation tasks, but isn't certified.

There are still high end applications that require certified drivers, so those user still have to pay the even higher prices. Likely they were the marjority of the market that buys the high priced cards anyway so not much will change.

 

[Flapdrol1337]

Does this cause conflict with the performance/price of their Quadro product? Are they kicking themselves in the foot? I know Titan isn't certified but Nvidia has always been one to cripple hardware/software to ensure product space.

It doesn't really change all that much, even without driver work geforce products had better price/performance in cad software since quadro is so expensive. Maybe they'll make more money this way since people that used to buy geforce might now buy a titan.

 

[n0x1ous]

so titan is getting its own driver separate from geforce?

 

[Cookie Monster]

So nVidia was intentionally crippling their product?

More like the power of competition

We need more of this.

 

[Malogeek]

And most of them have rubbish freesync. Even the monitor amd is advertising specifically for freesync in the Vega discount is useless (having and effective 80-100hz vbr range). I think if you removed all the freesync monitors that didn't work properly, and I define that by meeting the gsync min spec which all gsync monitors meet (i.e. 30-max hz range, overdrive support, etc) you'd find the numbers are completely different.

Completely agree on this. The G-sync product is a very tight specification with assistance hardware guaranteeing specific results. Freesync is all over the place but it's a start at least and provides an alternative in the VRR space without lock-in to a specific brand, which I like. Of course one could argue that the lock-in is there still to AMD for Freesync but that's not AMD's intention and is only true by virtue of Nvidia having their premium product (and thus refusal to support Freesync) and Intel... well nevermind.

 

[Paratus]

I do wonder why isn't there a monitor that supports both G-sync & Freesync?
Seems they could wire up a bypass for G-sync if needed.

Also, this came out today... (shots fired!)
3x the performance with new drivers... it is almost like they were waiting for something to launch...


https://blogs.nvidia.com/blog/2017/...ls-of-performance-for-creative-professionals/

Wow. So they've been sitting on a 3X improvement for their $1200 card. That's really crappy.

Guess Titan Xp ownwers need to thank AMD for their performance.

Titan X Pascal owners are out of luck I guess.

 

[Malogeek]

Wow. So they've been sitting on a 3X improvement for their $1200 card. That's really crappy.

It's called product segmentation and I'm sure AMD wish they could do it on the level that Nvidia does with their professional products.

 

[Veradun]

So Vega is not competitive. That's why nVidia released an unlocked driver for their Titan. Right.