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

[mohit9206]

Just watched the Adoredtv video titled "Poor Vega".
He called Vega a bigger fail than Fermi.
Atleast Fermi had competitive performance, Vega doesn't even have that.
So much for everybody riding the Vega hype train. Not surprisingly the train got derailed.

 

[Elixer]

There aren't many benches with pro cards on gaming, but from the ones I've seen they do worse than the much cheaper gaming cards.

Let me help you.

are just a few...

My main point is, these days, it is the exact same silicon down the line, with things gimped the lower down the ladder you go. Nvidia is famous for gimping drivers & silicon to prevent people from "crossing lines" into a higher bracket.
There would be no good reason here for AMD to be gimping their drivers so people that want to pay more get less performance than a Vega RX.

Vega FE beats the Titan XP in workstation applications, computing, AI, VR, etc... Again not every company needs to run airplane engine simulations or biomolecular simulations where you need 100% reliability. This the Vega FE makes sense for work in companies that don't need the certified drivers and can't afford the more expensive pro cards, but still need the speed of those, which the Vega FE provides.

Where are the facts to back this up?
Have you seen these applications / computing / AI / VR where Vega FE actually shines? No, I don't mean AMD PR's team generated "wins". Those don't count.
AFAIK, we have only seen benchmarks of SPECviewperf, where it loses to a Titan Xp 3 times and is more or less a wash.
We haven't seen any good AI/VR benchmarks as of yet.

Yes, I do realize that by the time Vega RX finally ships, and the drivers have been tuned a bit more, it could have better scores. I'm not arguing that at all.
I am saying that those same drivers running on the Vega FE should be pretty much equal, assuming clock speeds are the same.

 

[Kallogan]

indeed poor amd whose gpus are sold out cause of mining

 

[Bacon1]

I see a $2000 P5000 is slower than the 1080 which has half the VRAM but otherwise the same specs.

Where are the facts to back this up?
Have you seen these applications / computing / AI / VR where Vega FE actually shines? No, I don't mean AMD PR's team generated "wins". Those don't count.
AFAIK, we have only seen benchmarks of SPECviewperf, where it loses to a Titan Xp 3 times and is more or less a wash.
We haven't seen any good AI/VR benchmarks as of yet.

http://imgur.com/a/vzvhf

Those were all tested by PCPER, looks like VEGA is a much better buy with only 2/8 of the tests having the Titan XP faster and many it was much slower.

 

[tential]

So titanxp gives better performance in 3/8 of the tests and gives better gaming performance 100% of the time? How is that good for the fe card?

 

[Bacon1]

So titanxp gives better performance in 3/8 of the tests and gives better gaming performance 100% of the time? How is that good for the fe card?

Guess you missed the cost difference and it being faster in 5/8 pro tasks tested...

 

[Bouowmx]

With some of these tests, savings of AMD Vega Frontier Edition over NVIDIA big Quadro is irrelavent next to software costs. And the same stuff you heard numerous other times, like Vega Frontier Edition not being pro without application and driver certification, and Vega Frontier Edition appealing to the "not-professional professional" in cost (sorry for mentioning again).

Can NVIDIA Quadro comparisons be left to AMD Radeon Pro WX9100?

 

[Bacon1]

With some of these tests, savings of AMD Vega Frontier Edition over NVIDIA big Quadro is irrelevant next to software costs.

Are you saying that for PRO work hardware costs don't matter?

If that is the case, why are there so many Quadro cards ranging from $400, $900, $2000, and $5000? Why wouldn't all companies just buy the $5000+ GPUs if cost doesn't matter?

Not all "professionals" need the higher cost software or driver certification, that's why they buy things like the Titan XP. Vega FE compares to the Titan XP and beats it in these tasks. It also games better than the P4000. Yes the WX9100 will be even more "Pro" focused, and I'm guessing the certified drivers will also come to the FE just like the RX drivers will when they are released.

 

[Bouowmx]

I am not saying that.

Cool, I saved 1000 USD getting AMD Vega Frontier Edition over NVIDIA Quadro P5000, when software licenses cost ~ 1500 USD (Autodesk 3ds Max and Maya, and Creo) and up (unknown? Catia and Siemens NX (SNX)). Now I can fancy not receiving application and driver certifications and support in running my expensive software.

Various NVIDIA Quadros exist because the processors have already been developed. Make a processor just for GeForce? Well, GP108 is just for GeForce (got me there).


> not a gaming card
> not a pro card

The oxymoron of Frontier Edition's reactions.

 

[tential]

Guess you missed the cost difference and it being faster in 5/8 pro tasks tested...

I didn't miss that, I'm saying are you considering that to be in the frontier editions favor in comparison to titan xp?

Titan xp
Faster in 100% of gaming cases.
Faster in 3/8 professional workloads
Consumes less power

Founders edition
Faster in 5/8 professional workloads
Performs like a ~gtx 1080 in gaming
Consumes a lot of power

I'm just saying are you genuinely thinking the frontier edition is offering a good overall package.
Not to just explain use cases where someone may in theory pick the frontier edition.


Since the closest competitor is a titan xp, and both cards game and do pro tasks, then a large elephant in the room is that the card just outright loses in ALL gaming benchmarks. The fe may cost less, but the gaming deficit is tremendous.

Power consumption is a massive issue as well.
While it may be cheaper if you end up having to also buy a new power a supply and the huge power consumption is a not good for multiple cards

The only benefit the fe has is a couple of pro benchmarks. I don't think that's worth a tiny drop in price.

I honestly won't be surprised if you don't see much of this card in the future, or if the price is further reduced as at the current price it's still a meh product.

$700-800 is the price I expect this to go for very soon.

 

[Bacon1]

I didn't miss that, I'm saying are you considering that to be in the frontier editions favor in comparison to titan xp?

If anything, this card makes the Titan XP completely unneeded.

1080 Ti is much better for gaming than Titan XP, its faster, doesn't throttle, has custom coolers, and is half the price.

Vega FE is better in most Pro uses, and is also cheaper.

Both use about the same power FYI.

Adding the Titan Xp stock card to the test, we see an average power consumption of about 357W – so that’s 27W lower than the Vega: FE system

The Titan Xp runs lower power consumption, but also higher thermals.

Vega's cooler is also much better than Titan XPs:

They also found that trying to OC the core clocks would always reduce HBM down to 500 (Fury's speeds) from 945 which caused a massive downclock by reducing HBM almost 50%. Not sure why no one else reported this problem.

So why would anyone buy a Titan XP over Vega FE for Pro tasks, or a 1080 Ti for gaming / some pro stuff (its not much weaker than Titan XP in Pro stuff after all!)?

 

[Guru]

You have mentioned this fact multiple times, but the only professional benches. I only seem to recall the initial pcper review, were it was a wash (either XP or Quadro being faster, sometimes significantly so). Have I missed some benches were VEGA wins Titan XP in all those workloads?

Even the pcper ones showed Vega FE winning significantly in applications, but you've missed several others where the Vega FE wins in the majority of the cases against the Titan XP.

 

[Dribble]

So why would anyone buy a Titan XP over Vega FE for Pro tasks, or a 1080 Ti for gaming / some pro stuff (its not much weaker than Titan XP in Pro stuff after all!)?

No one will buy either for pro tasks, they will buy a quadro because certified drivers are much more important then a little extra performance. The whole Vega FE for pro is a complete red herring fired out their by AMD marketing to make Vega look better at something. Both the Titan and the Vega FE are gaming cards. The titan sells because it is the absolute fastest, Vega FE won't because it isn't.

 

[SpaceBeer]

Radeon Pro Duo (both Fiji and Polaris) is certified for stuff like 3ds max, Maya, Premiere Pro.... So I don't see reason why FE won't be.

We know any gaming card can be used for CAD/GPGPU work, and any Pro card can be used for gaming. But the most important thing is to know that in one case, your driver is optimized for games (lot of frames per second in DX11/12) and in other for pro applications (less FPS, more accurate model display in openGL). And if AMD could give me pro driver for RX 560 so it would behave EXACTLY like WX 4100, I wouldn't mind if such card is not on Siemens' or Autodesk's certified cards list

 

[Glo.]

Radeon Pro Duo (both Fiji and Polaris) is certified for stuff like 3ds max, Maya, Premiere Pro.... So I don't see reason why FE won't be.

We know any gaming card can be used for CAD/GPGPU work, and any Pro card can be used for gaming. But the most important thing is to know that in one case,your driver is optimized for games (lot of frames per second in DX11/12) and in other for pro applications (less FPS, more accurate model display in openGL)

Radeon Pro Duo is part of Radeon Pro brand.

Radeon Vega Frontier Edition is not part of this brand. Radeon Pro WX9100 will be part of Radeon Pro brand, and will be based on Vega 10 chip and will have signed drivers.

 

[Geegeeoh]

Not sure if it deserves its own topic:
AMD RX Vega Leaked Benchmark Shows It Ahead Of GTX 1080 – Specs Confirmed, 1630MHz Clock Speed, 8GB HBM2 & 484GB/s Of Bandwidth

RX Vega, not Vega FE.

 

[JDG1980]

Someone on Reddit pointed out that the official Vega architecture page contains NO mention of the Draw Stream Binning Rasterizer, despite that feature's inclusion in the slide deck Anandtech posted a while back.

This reinforces the suspicion that something is badly wrong with Vega at the silicon level, and tiled rendering just doesn't work on current chips. That could be a very serious problem if it affects Raven Ridge's Vega iGPU as well, since that will rely heavily on tiled rendering for performance improvements over Bristol Ridge due to the memory bandwidth bottleneck.

 

[3DVagabond]

No one will buy either for pro tasks, they will buy a quadro because certified drivers are much more important then a little extra performance. The whole Vega FE for pro is a complete red herring fired out their by AMD marketing to make Vega look better at something. Both the Titan and the Vega FE are gaming cards. The titan sells because it is the absolute fastest, Vega FE won't because it isn't.

Fake news. You are just flat making things up.

 

[Veradun]

Someone on Reddit pointed out that the official Vega architecture page contains NO mention of the Draw Stream Binning Rasterizer, despite that feature's inclusion in the slide deck Anandtech posted a while back.

This reinforces the suspicion that something is badly wrong with Vega at the silicon level, and tiled rendering just doesn't work on current chips. That could be a very serious problem if it affects Raven Ridge's Vega iGPU as well, since that will rely heavily on tiled rendering for performance improvements over Bristol Ridge due to the memory bandwidth bottleneck.

This also reinforces my (and someone else's) guess that the RX is late because of a necessary silicon respin :>

 

[Head1985]

Just watched the Adoredtv video titled "Poor Vega".
He called Vega a bigger fail than Fermi.
Atleast Fermi had competitive performance, Vega doesn't even have that.
So much for everybody riding the Vega hype train. Not surprisingly the train got derailed.

Link https://www.youtube.com/watch?v=8iaUjU-2Jmc&t
btw 40%faster than gtx1070 for 400usd?I think that is too good to be true.FURYX had worst performance/cost on market and i think vega will be same.HBM2+huge die...

 

[Glo.]

Someone on Reddit pointed out that the official Vega architecture page contains NO mention of the Draw Stream Binning Rasterizer, despite that feature's inclusion in the slide deck Anandtech posted a while back.

This reinforces the suspicion that something is badly wrong with Vega at the silicon level, and tiled rendering just doesn't work on current chips. That could be a very serious problem if it affects Raven Ridge's Vega iGPU as well, since that will rely heavily on tiled rendering for performance improvements over Bristol Ridge due to the memory bandwidth bottleneck.

It is mentioned...

Pixel Engine. Part of it is the Draw Stream Binning Rasterizer...

P.S. It appears that all of the features for Vega are apparent in drivers. The thing is, software is still reading it as Fiji chip, because of such similarity in product design(4 shader engines, 64 CU's, etc).

Developers will have to rewrite the software to take benefit of the architecture.

One of examples. 16 bit float is not read by software on the go.

 

[leoneazzurro]

Someone on Reddit pointed out that the official Vega architecture page contains NO mention of the Draw Stream Binning Rasterizer, despite that feature's inclusion in the slide deck Anandtech posted a while back.

This reinforces the suspicion that something is badly wrong with Vega at the silicon level, and tiled rendering just doesn't work on current chips. That could be a very serious problem if it affects Raven Ridge's Vega iGPU as well, since that will rely heavily on tiled rendering for performance improvements over Bristol Ridge due to the memory bandwidth bottleneck.

There are a lot of other things not mentioned there (i.e. primitive shaders, hw support for all DX12 tiers, and so on), but it does not mean they are not supported.

 

[Glo.]

There are a lot of other things not mentioned there (i.e. primitive shaders, hw support for all DX12 tiers, and so on), but it does not mean they are not supported.

Primitive Shaders are mentioned. Just under the Pixel engine.

It is all part of Geometry Pipeline.

 

[beginner99]

Someone on Reddit pointed out that the official Vega architecture page contains NO mention of the Draw Stream Binning Rasterizer, despite that feature's inclusion in the slide deck Anandtech posted a while back.

This reinforces the suspicion that something is badly wrong with Vega at the silicon level, and tiled rendering just doesn't work on current chips. That could be a very serious problem if it affects Raven Ridge's Vega iGPU as well, since that will rely heavily on tiled rendering for performance improvements over Bristol Ridge due to the memory bandwidth bottleneck.

Yeah saw that too. This is a troubling observation and might also explain the delays. They knew but just did not get it to work.

 

[SPBHM]

Not sure if it deserves its own topic:
AMD RX Vega Leaked Benchmark Shows It Ahead Of GTX 1080 – Specs Confirmed, 1630MHz Clock Speed, 8GB HBM2 & 484GB/s Of Bandwidth

RX Vega, not Vega FE.

this is exactly what I expected from Vega, based on the "13TF" thing and how AMD compared with previous NV products