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

[Head1985]

We still dont know how Vega overclocks. PCPer tried it. Overclocked the Vega FE to 1682MHz but got almost no gain in FPS. Power draw increased from stock 280W to 350W. GTX 1080 can easily do +20% more FPS with overclock.

Thats sign of memory bandwidth bottleneck.I am telling this again and again.Vega is memory bandwidth bottleneck.Less bandwidth than furyX and also no new delta color compression.Vega using old polaris delta color compression whitch is far worse than pascal one.(rx580 have same bandwidth as GTX1070 yet its far slower and rx580 is also memory bandwidth bottleneck)
So yeah vega is very hard memory badwidth bottleneck.Thats why it suck in games at the moment.Tile rasterizer should help with that btw(we dont know if its Hw bug and its not working at all or its working in games but not in triangle test or amd dont hnow how code it in drivers)

Btw no GTX1080 dont gain 20% from OC.Its 10-12% like GTX1070.

 

[raghu78]

Vega FE and Vega RX will have the same cores, the same silicon.
AMD used a dual driver setup with Vega FE. One that utilize the professional side to do professional workloads. Like designing a game. The other one runs a gaming driver code, for the same developer to run his game to see how smooth it will run on a gaming card. Thats what AMD themselves state and the idea behind this is kinda cool.

Why would Vega RX run much different then? Vega FE is 10-30% below GTX 1080 in games in gaming mode. At best it sounds to me that it can get close to GTX 1080. But even that looks like a stretch.

PcPer say Vega RX most likely will get around 10% more than Vega FE.

Because this driver does not have the gaming optimizations for the Vega architecture. As for pcper lets say they have no knowledge in the subject of Vega GPU architecture design or Vega GPU driver design.

 

[Peicy]

Because this driver does not have the gaming optimizations for the Vega architecture. As for pcper lets say they have no knowledge in the subject of Vega GPU architecture design or Vega GPU driver design.

There is hope that a lot of potential performance is not used at the moment. We dont know for certain if thats the case or not. I find it unthinkable that the RX variant will perform like this in games, but then again every couple of years a Bulldozer, Pentium 4 or Geforce FX happens.

Either way, they really botched this hard. If there is a 30 to 40% performance uplift for the RX variant in the cards, they would know that by now. And maybe...if they are smart, issue a statement? Because right now, Vega is a PR desaster.

 

[french toast]

There is hope that a lot of potential performance is not used at the moment. We dont know for certain if thats the case or not. I find it unthinkable that the RX variant will perform like this in games, but then again every couple of years a Bulldozer, Pentium 4 or Geforce FX happens.

Either way, they really botched this hard. If there is a 30 to 40% performance uplift for the RX variant in the cards, they would know that by now. And maybe...if they are smart, issue a statement? Because right now, Vega is a PR desaster.

This. There is no doubt there is a large performance uplift there not activated by drivers, how much is un known and how long to we see it is also important.
BUT amd has had silicon running since November/December, drivers should be fairly respectable by now and if there is a large uplift arriving in a month they should offer up some statement as this is the kind of pr disaster that nvidia seems to avoid with titan - titans launch and are judged on gaming, when your charging top $ for a non professional card gaming performance DOES matter on launch no matter if it's prosumer or not, it should be judged against its competitors in its price range to the same standard.

I believe we will see better performance with Rx Vega, I certainly hope so.

 

[Cloudfire777]

There is hope that a lot of potential performance is not used at the moment. We dont know for certain if thats the case or not. I find it unthinkable that the RX variant will perform like this in games, but then again every couple of years a Bulldozer, Pentium 4 or Geforce FX happens.

Either way, they really botched this hard. If there is a 30 to 40% performance uplift for the RX variant in the cards, they would know that by now. And maybe...if they are smart, issue a statement? Because right now, Vega is a PR desaster.

That 30% increase is just wishful thinking.
AMD giving a professional card that people paid $1000+ for and giving it less performance than it should? Yeah lets piss off the buyers that is willing to dish out the most...how is that even logical?

I will post this quote once again.
The only difference between RX Vega and FE Vega is that RX Vega will ship with a newer driver. As in the difference between a January 2017 and a say April 2017 driver for any video card

You cant get it more clearer than this.

Maybe @rys could shine a light on the matter, just one question: is the Vega FE gaming driver so far gimped compared to the Vega RX driver coming a month later?

It's not gimped (that would be completely ridiculous), but it is older.

https://forum.beyond3d.com/threads/...rs-and-discussion.59649/page-119#post-1989522

 

[Magee_MC]

If, as Raja said, creating new drivers for Vega is extremely difficult, then it's not surprising that the drivers that the FE shipped with are slower than the RX drivers will be. In professional applications, stability is king, where with gaming speed is what you focus on. Which would professionals rather have faster, more unstable drivers or slower stable ones.

My take is that what we have here is a mix of all of the discussed problems. A hardware issue that required a new stepping that pushed back the launch date. An absolute need to meet the 1H release promised to investors. Incredibly difficult to code for drivers for all of the new additions to Vega, and possibly a need to rework all of the driver work previously done for the new stepping.

I expect that the FE drivers will significantly increase in speed in a little bit, but that will take time to make sure that they are absolutely stable for the pro users. I also expect that the RX drivers will make use of all or most of the new parts of Vega at release and that they will be significantly better than what we have seen to date with the FE. It's also possible that a lot of the new power management systems aren't being used yet, so what we are seeing is pretty much a raw, stable and functional Vega, but not what it will be in the near future once it's running on all cylinders.

 

[piesquared]

That 30% increase is just wishful thinking.
AMD giving a professional card that people paid $1000+ for and giving it less performance than it should? Yeah lets piss off the buyers that is willing to dish out the most...how is that even logical?

I will post this quote once again.
The only difference between RX Vega and FE Vega is that RX Vega will ship with a newer driver. As in the difference between a January 2017 and a say April 2017 driver for any video card

You cant get it more clearer than this.




https://forum.beyond3d.com/threads/...rs-and-discussion.59649/page-119#post-1989522

Have you considered that Vega FE will also get the driver updates? Why would that piss off buyers of Vega FE?

The people buying those are buying it for its professional cababilities not its gaming capabilities, or at least they should considereing that is what AMD has been saying since the Frontier Edition's introduction.

 

[raghu78]

Have you considered that Vega FE will also get the driver updates? Why would that piss off buyers of Vega FE?

The people buying those are buying it for its professional cababilities not its gaming capabilities, or at least they should considereing that is what AMD has been saying since the Frontier Edition's introduction.

We know few users are here to create FUD and not have any meaningful or logical discussions.

 

[Elfear]

An interesting assessment by one of the technical gurus over at B3D.

"Well, I agree that it is very optimistic to hope for this [meaning a 30% performance increase of RX over FE]. But assuming that the current state is really not using any of the architectural improvements over Fiji (besides the L2 backing of the ROPs) as the results very much are identical to what to expect from a heavily overclocked FuryX (~1.4Ghz with slightly more exotic cooling), it wouldn't be completely impossible. That assumption requires quite a bit of faith, though. But keep in mind that nV claims their tiled approach saves them ~60% (!) of the external memory bandwidth (in combination with DCC, which Fiji was barely taking advantage of). Add in some saved work through the means of HSR within bins when applicable, the allegedly overhauled work distribution in the frontend, some slight IPC improvements of the CUs, and it may start to appear feasible. The alternative would be that AMD has severely messed up the Vega design."

So a big increase is possible given the architectural improvements AMD has touted but I still think the probability is low for the RX to come out and compete with GP102.

 

[beginner99]

You sure about that?

Specviewperf12
Vega wins 6 out of 9 vs the Titan Xp.
Vega wins 7 out of 9 vs the P4000.

Yeah but P400 is cheaper.

 

[french toast]

Is there a chance that some of the lower precision compute resources could be fused off/disabled? Thereby using slightly less power or allowing for higher clocks?
Edit: with the Rx Vega.

 

[Hi-Fi Man]

The speculations here are insane, let's stick with facts instead. Here is a dose of reality.

This isn’t a gaming card.

The Radeon Vega Frontier Edition is billed a card for creators that game, and gamers that create. Or rather, the Radeon Pro Duo was. But the audience segment matches up with the language AMD has been using for FE. There is a fine line between the high end of the gaming market and the professional market made up of Quadro and Radeon Pro (previously FirePro). NVIDIA has skirted that with the Titan product family for some time, but the Frontier Edition of Vega is the first time we are seeing that from AMD.

Calling this “not a gaming card” is a fair statement, as long as you also agree then the GTX Titan, Titan X, Titan Xp are also “not gaming cards.” But they are, despite NVIDIA segmenting it off as well. Plenty of professionals will buy this hardware, but discerning gamers that want the best of the best will also be purchasing Titans and FEs well into the future. “Professional graphics cards” have certified drivers and specific code paths in place for applications like 3ds Max, Maya, etc. Neither Titan nor Vega FE have that and instead will depend on the driver stacks we are used to seeing in GeForce and Radeon systems.

Even if it weren’t a gaming card, we have every right to test it with gaming applications! Is a professional developer, gamer or not, going to buy a $4000 Quadro and THEN a GTX 1080 Ti to game on? Nope, they are going to get double duty of that card.

Once the Radeon RX Vega graphics card hits the market, then the Vega Frontier Edition can exist in its vacuum where it only addresses the professional market. Until then, the Vega FE will be the pseudo-representative of how the rest of us are expecting the consumer gaming card variants to perform.

The drivers are old.

There were concerns over the last couple of days that the driver for Vega FE on the site was from January of this year. As it turns out, reading version numbers of the Radeon driver package is difficult, and the driver we used in our testing, and the ONLY driver that supports Vega FE today, is not old. To be clear though, the driver IS from a different branch than the currently available Radeon RX 500-series driver, but the exact time of that branch, and how it affects performance on games or other workloads, isn’t information AMD is interested in sharing at this time.

The driver isn’t optimized for gaming.

I saw this pop up a lot during our stream yesterday, that the driver isn’t meant for gaming so it hasn’t been optimized for gaming. Instead, it’s only targeting “professional” level applications. First, that’s not the case and AMD has confirmed that. The driver has all the gaming optimizations that the other Radeon drivers would include up until at least the driver branching mentioned above. After that time, optimizations may or may not have made it in, as AMD tells it.

The games we are using for this review were not released in the last 30 days or anything like that. GTA V, Rise of the Tomb Raider, Witcher 3; these are all games that have been out for some time, were around for AMD to address in both Radeon RX 500 and Vega-series drivers for many, many months.

The one caveat to this is that the Vega architecture itself is still unoptimized in the driver. But this would affect gaming and non-gaming workloads most of the time. So if the driver isn’t ready for ANYTHING, then that’s a valid concern, but it applies to ALL workloads and not just games.

 

[Bouowmx]

Is there a chance that some of the lower precision compute resources could be fused off/disabled? Thereby using slightly less power or allowing for higher clocks?
Edit: with the Rx Vega.

http://www.anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser/2

I don't think there are lower-than-32-bit floating-point cores to fuse off. One 32-bit core handles one 16-bit operation, or two 16-bit operations packed together.

Spoiler: AMD Vega packed math

 

[Paratus]

An interesting assessment by one of the technical gurus over at B3D.

"Well, I agree that it is very optimistic to hope for this [meaning a 30% performance increase of RX over FE]. But assuming that the current state is really not using any of the architectural improvements over Fiji (besides the L2 backing of the ROPs) as the results very much are identical to what to expect from a heavily overclocked FuryX (~1.4Ghz with slightly more exotic cooling), it wouldn't be completely impossible. That assumption requires quite a bit of faith, though. But keep in mind that nV claims their tiled approach saves them ~60% (!) of the external memory bandwidth (in combination with DCC, which Fiji was barely taking advantage of). Add in some saved work through the means of HSR within bins when applicable, the allegedly overhauled work distribution in the frontend, some slight IPC improvements of the CUs, and it may start to appear feasible. The alternative would be that AMD has severely messed up the Vega design."

So a big increase is possible given the architectural improvements AMD has touted but I still think the probability is low for the RX to come out and compete with GP102.

It's possible but not likely. However when Fiji was designed they literally ran out of room to cram anything more into the die. It sort of made sense that Fiji was somewhat unbalanced, even if it was almost exactly twice a Tahiti. It did compete with the 980Ti at launch (at least stock vs stock)

Vega on the other hand kept the same basic layout and added roughly 60% more transistors. Vega appears to perform like nothing more than an oc'd Fiji. If that's all it is why not just die shrink Fiji and put out a 362mm GPU that comes in between a 1070 and 1080 instead of a 562mm GPU that comes in at the same performance.

If it's not, what are they doing with the extra transistor budget? I'm unaware of any significant new functionality that would take most of that budget AND not provide an IPC lift. I can't believe they would chase a bad design this long so there must be a bottleneck somewhere and yet they should have alleviated that bottleneck with the increased transistor budget.

So I'm stumped.

I guess I'm leaning towards drivers. It performs well enough professionally to have a niche so maybe that's where their man powe went. Maybe we will see a drastic improvement in gaming performance with RX launch drivers.

If not it's a design failure but I'm struggling to see how that's possible.

 

[french toast]

http://www.anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser/2

I don't think there are lower-than-32-bit floating-point cores to fuse off. One 32-bit core handles one 16-bit operation, or two 16-bit operations packed together.

Spoiler: AMD Vega packed math

Thanks, but what about 4x integer 8 bit?

 

[Cloudfire777]

Have you considered that Vega FE will also get the driver updates? Why would that piss off buyers of Vega FE?

The people buying those are buying it for its professional cababilities not its gaming capabilities, or at least they should considereing that is what AMD has been saying since the Frontier Edition's introduction.

I am saying that the gaming capabilities of Vega FE is just as good as RX Vega. Right now its just using a bit older driver than what RX Vega is shipping with.

That is what is taking AMD so long. I bet the delays is because AMD saw that they still didnt meet their performance target of Vega, and decided to delay the RX Vega to squeeze out as much performance as possible.
But even drivers have their limits on what is possible.

 

[Karnak]

The speculations here are insane, let's stick with facts instead. Here is a dose of reality.

Yep, from that quote.

The one caveat to this is that the Vega architecture itself is still unoptimized in the driver. But this would affect gaming and non-gaming workloads most of the time. So if the driver isn’t ready for ANYTHING, then that’s a valid concern, but it applies to ALL workloads and not just games.

And that's what we don't know.

 

[eek2121]

Well... NVIDIA has been on top since 980 Ti though. And it looks like that is not going to change. If you want the fastest you sure as hell will not buy Fury X.

Pleas don't say stuff like this in a factual way. The 980ti was pretty much neck in neck with the Fury-X exept in ROTR, an Nvidia optimized game. Source: http://www.anandtech.com/bench/product/1720?vs=1715








This isn't an Nvidia nor a 980ti thread.


esquared
Anandtech Forum Director

 

[Hi-Fi Man]

Yep, from that quote.


And that's what we don't know.

He also goes on to say,

There is still potential that AMD will be able to find some additional performance in the product before the consumer launch, either in drivers and efficiency improvements, higher clocks, or maybe better binning. At most, I could see AMD adding another 10% to the results we see with an odd case of a significantly under-developed part of the driver stack going beyond that.

Since AMD has had working drivers for the past 6+ months I highly doubt these drivers are that severely underdeveloped that we could see performance improvements past 10%.

 

[IllogicalGlory]

Pleas don't say stuff like this in a factual way. The 980ti was pretty much neck in neck with the Fury-X exept in ROTR, an Nvidia optimized game. Source: http://www.anandtech.com/bench/product/1720?vs=1715

Yeah, but the reference 980 Ti was terribly gimped by its cooling and PCB.

The 980 Ti came in beastly variants such as this: https://www.techpowerup.com/reviews/ASUS/GTX_980_Ti_Matrix/26.html

That one overclocks to about 30% faster than reference, and out of the box is over 20% faster at 1440p. Fury X just can't compete with that.

The 1080 Ti does not behave quite so well, so if by some miracle RX Vega reaches Fury X-ish performance relative to the 1080 Ti, it's much better positioned.

 

[eek2121]

You don't have to believe what AMD is saying, y'know. They are just making excuses because the gaming performance isn't enough to justify the prices they need to sell Vega at, that's why there are no gaming cards out now.

If that were the case, AMD would have released RX Vega at the same time. The gaming drivers aren't ready. I am not just taking their word for it, I know this to be fact.

Did you know that both AMD and Nvidia have optimized code paths for the top games? They may even use shader specific optimizations. None of that stuff is used in pro applications, but it IS used in gaming. All the stuff in the pro drivers is also used in the gaming drivers so which would you build first?

 

[railven]

Woof so much to catch up on. That die is HUGE!!!! Pretty sure I remember, and other's have said, Raja mentioned it was smaller?

Just another reminder AMD can't promote their own products without shooting a foot. AMD were the ones promoting Gaming Mode on their website. I don't think anyone coming into the launch of this product was going to take Gaming Performance serious (yes, yes the trolls would spread it), but AMD themselves basically saying "Gaming Mode, makes it a full fledged Radeon!!!" sort of kicked them in the jimmies.

Sure, argue old drivers all you want. AMD still put this thing in the wild, with the current limitations, and are now having their product dragged through the mud because of it. RX Vega's hill got even steeper and it has no one to thank but Vega FE for it.

Siggraph is going to have a lot of eyes on it.

 

[Hi-Fi Man]

If that were the case, AMD would have released RX Vega at the same time. The gaming drivers aren't ready. I am not just taking their word for it, I know this to be fact.

Did you know that both AMD and Nvidia have optimized code paths for the top games? They may even use shader specific optimizations. None of that stuff is used in pro applications, but it IS used in gaming. All the stuff in the pro drivers is also used in the gaming drivers so which would you build first?

In case you missed it from my previous post.

The driver isn’t optimized for gaming.

I saw this pop up a lot during our stream yesterday, that the driver isn’t meant for gaming so it hasn’t been optimized for gaming. Instead, it’s only targeting “professional” level applications. First, that’s not the case and AMD has confirmed that. The driver has all the gaming optimizations that the other Radeon drivers would include up until at least the driver branching mentioned above. After that time, optimizations may or may not have made it in, as AMD tells it.

 

[Erithan13]

I'm pretty well exhausted from following all the 'polite discussion' around this launch, and while I'm not expecting miracles from RX Vega, it's worth bearing in mind how dangerous it can be to jump to conclusions. Exhibit A:

Vega Crashes during blender benchmark Hundreds of people piling in with their incredibly witty and original observations about 'poor volta, wait for navi, another fail for AMD' etc etc

After this particular 'truth' has gotten the whole way around the world, this appears:

Crashes are caused by a bug in the benchmark, not the gpu/drivers, fix available

RE: The die size: Seems to be some confusion over this, last I heard it is indeed <500mm^2, and the measurements suggesting otherwise are in error (accidentally measured extra area outside the actual die, even a tiny error in the linear dimension measure makes a big difference to the area. I'll see if I can find the post)

 

[tential]

If that were the case, AMD would have released RX Vega at the same time. The gaming drivers aren't ready. I am not just taking their word for it, I know this to be fact. [/]

Did you know that both AMD and Nvidia have optimized code paths for the top games? They may even use shader specific optimizations. None of that stuff is used in pro applications, but it IS used in gaming. All the stuff in the pro drivers is also used in the gaming drivers so which would you build first?

Without a source or anyway to verify this comment its really detrimental to technical or any meaningful level of conversation here.