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

[Elixer]

The question that I don't know the answer to is, if AMD had to do a significant change to the layout of the silicon in order to remove some kind of show stopping problem, could that require significant rework of the drivers that they already had. The silicon that they had might have been nowhere near what they were expecting and needed, so they had no choice but to delay Vega and fix the problem and take the hit on the delay.

The basic design still should be the same, it is highly doubtful they found something that they would have to change completely after silicon has been made.

Though, come to think of it, it is possible there is indeed a hardware bug, and they are trying to work around it via drivers. That wouldn't be unheard of.
If we look back to the TLB bug on Phenom CPUs, that could cause a whopping 70% decrease in performance trying to fix that via software (BIOS).
That required a new stepping to fix without that massive performance hit.

Now, for that to be the case here, it takes ~16 weeks for a major respin, ~8 weeks for something minor. Then add a few weeks for the fab to stop & start again.
I can't seem to find the revision of the chip that is being used for Vega FE on any of the sites that bought the card, anyone know?

Anyway, it could very well be possible that they actually have a new Vega revision for RX,
the TLB I mentioned above, AMD stopped the fab orders in Nov. 2007, and it wasn't until April 2008 before the fixed version was out and shipping.
That would mean that AMD would have had to have a new version of Vega ready by Feb 2017 before we would see inventory at the end of July.
Yes, this scenario is far-fetched, in that, they would have to be shipping bugged Vega FE cards, and the specific workload only shows its ugly head on things that don't have to do with the more "professional" benchmarks that we have seen.

If that's not what happened, then I've got nothing else at the moment to explain this confusing morass of a release.

Can't disagree, this is one oddball release.

 

[.vodka]

https://semiaccurate.com/forums/showpost.php?p=292057&postcount=2463

Pictures of Vega FE from PCPer

Sorta throws a wrench in the RX is a respin theory some have mentioned.
Date code is 20th week of 2017, May 15-21. The only way a respin could happen for the RX launch would be if the date code isn't the silicon manufacture date but the package date.

Edit- But that would fly in the face of Charlie saying AMD was doing it right by delaying and waiting for enough supply for it to be a hardlaunch.

Hmm... He's right... These are pretty fresh out of the fab OR packaging facility, this makes all the difference in the world and we don't know exactly which one is it.

 

[Elixer]

https://semiaccurate.com/forums/showpost.php?p=292057&postcount=2463

Hmm... He's right... These are pretty fresh out of the fab OR packaging facility, this makes all the difference in the world and we don't know exactly which one is it.

Nice find!

 

[beginner99]

nyway, it could very well be possible that they actually have a new Vega revision for RX,
the TLB I mentioned above, AMD stopped the fab orders in Nov. 2007, and it wasn't until April 2008 before the fixed version was out and shipping.
That would mean that AMD would have had to have a new version of Vega ready by Feb 2017 before we would see inventory at the end of July.
Yes, this scenario is far-fetched, in that, they would have to be shipping bugged Vega FE cards, and the specific workload only shows its ugly head on things that don't have to do with the more "professional" benchmarks that we have seen.

I mean Vega release once was planned for end of 2016, then early 2017 and now it is mid 2017. So that would pretty much fit the respin time requirements. The semiaccurate image with the date (which i fail to see on the image) could very well be packaging only. So it doesn't really confirm or deny this rumour.

A possible defect could be the new rasterizer. That should not affect professional workloads much but of course will affect gaming heavily. Why? Because it reduces memory bandwidth requirements and in fact vegas bandwidth is rather small especially compared to previous AMD designs plus Vega FE sees large performance increase from memory OC (Fury did so as well and Polaris is somewhat also BW limited). The theory then is that Vega FE is bandwidth limited in gaming workloads and if the new rasterizer works, then this bottleneck is gone and performance could go up a lot (>20%).

But all in all this is just grasping at straws hoping that AMD did not do another faildozer. Wasn't there a like about a month back along the lines vega has high power use but also high performance?

EDIT:

Remaining question is if power use (GPU utilization) should not be much lower if there actually was a BW limit. But if the SPs are actively waiting, asking for data, then no. Power use would be just as high.

 

[Glo.]

Gamers Nexus in 4K/Ultra settings with AAx0 averaged 53 FPS, with dips under 50 FPS, while hovering possibly around 1.4 GHz core clock.

In AMD demo, DOOM was running at 4K/Ultra settings, with AAx8. And it averaged 70-75 FPS, while never dipping below 65 FPS.

https://www.youtube.com/watch?v=Y8tDaPLHxiE

And funniet part. By looking at the DeviceID, and its clock speed that leaked later - it is 1.2 GHz, 8 GB HBM2 version of the GPU demoed here.

Currently, Vega FE is running on Fiji drivers. Without any of improvements.

 

[Trumpstyle]

I did some checking from various reviews. The polaris architecture has around 60-70% flop scaling (If you double the teraflops you get between 60-70% performance gain). You can check reviews between radeon 460 and 470 for confirmation.

So this should put FE Vega slightly above geforce 1080 (it's not ofc) and the Rx vega (which almost certainly a cut-down version) should be around 10% below geforce 1080. This is if Vega has same ipc as polaris architecture.

Since Vega seems to have same IPC as old GCN I suspect vega might NOT been design for gaming at all (I could be wrong ofc).

 

[Cloudfire777]

Gamers Nexus in 4K/Ultra settings with AAx0 averaged 53 FPS, with dips under 50 FPS, while hovering possibly around 1.4 GHz core clock.

In AMD demo, DOOM was running at 4K/Ultra settings, with AAx8. And it averaged 70-75 FPS, while never dipping below 65 FPS.

https://www.youtube.com/watch?v=Y8tDaPLHxiE

And funniet part. By looking at the DeviceID, and its clock speed that leaked later - it is 1.2 GHz, 8 GB HBM2 version of the GPU demoed here.

Currently, Vega FE is running on Fiji drivers. Without any of improvements.

LOL since when was AMDs numbers during presentations/marketing any believable? They tweak games in a undisclosed setting where the cards perform better vs competition.

Remember Fury X? I do...

Here is the tricks they did:

Which is settings HardwareNexus or anyone would ever touch.

 

[iwulff]

LOL since when was AMDs numbers during presentations/marketing any believable? They tweak games in a undisclosed setting where the cards perform better vs competition.

Remember Fury X? I do...

Even with using these 'tricks' to look better the difference seems really big this time around for Vega. Like something is very off and Vega isn't performing as expected for whatever reason.

 

[Glo.]

LOL since when was AMDs numbers during presentations/marketing any believable? They tweak games in a undisclosed setting where the cards perform better vs competition.

Remember Fury X? I do...

Here is the tricks they did:

Which is settings HardwareNexus or anyone would ever touch.

Stop spreading misinformation about Vega. You can check the settings in the videos.

Settings are Ultra, 4K resolution with AAx8.
Settings for Gamers Nexus: 4K resolution, Ultra, AAx0.

So in more difficult circumstances, somehow RX Vega GPU, with 1.2 GHz is performing MUCH better than in Gamers Nexus review, in lower complexity of details, with higher core clocks.

It has nothing to do with tricks. It has everything to do with Architecture Features not being visible for software, which already has been proven based on Tile-Based Rasterization.

 

[Bacon1]

Which is settings HardwareNexus or anyone would ever touch.

Just anyone trying to actually play the game right?

Or you know, run 4k @ Ultra @ 25fps instead in the games, like how most review sites "play" the games right?

 

[Gideon]

Currently, Vega FE is running on Fiji drivers. Without any of improvements.

*If* this is true, then it is a huge marketing blunder. They take the time to design a game mode and a professional mode in to the drivers, in order to develop optimizations and test those optimizations on a new architecture. Yet in the very same drivers they have absolutely zero of those features (your claim == Fury driver).

And what's worse, they give absolutely 0 feedback of this major, deficiency in the FE edition. All the marketing statement right now are overly broad and borderline useless (e.g. "drivers are not gimped, just older", "don't compare the RX Vega too much to FE"), guaranteeing speculation like this.

If this is actually the case (which I hope, but am far from convinced), then they only need to utter two simple sentences, and all this drama would have been avoided: "Our Release Driver still has major performance features disabled, expect performance uplift in the future" or "RX Vega will have a noticeble performance uplift, compared to the FE edition".

What would be the downside of saying that? Tipping Nvidia off? ... I don't think they could really do much in a month, they haven't planned to do anyway. Besides, if the board partners already know the specs (which they do, judging by MSI comments), then Nvidia already knows too.

 

[Glo.]

*If* this is true, then it is a huge marketing blunder. They take the time to design a game mode and a professional mode in to the drivers, in order to develop optimizations and test those optimizations on a new architecture. Yet in the very same drivers they have absolutely zero of those features (your claim == Fury driver).

And what's worse, they give absolutely 0 feedback of this major, deficiency in the FE edition. All the marketing statement right now are overly broad and borderline useless (e.g. "drivers are not gimped, just older", "don't compare the RX Vega too much to FE"), guaranteeing speculation like this.

If this is actually the case (which I hope, but am far from convinced), then they only need to utter two simple sentences, and all this drama would have been avoided: "Our Release Driver still has major performance features disabled, expect performance uplift in the future" or "RX Vega will have a noticeble performance uplift, compared to the FE edition".

What would be the downside of saying that? Tipping Nvidia off? ... I don't think they could really do much in a month, they haven't planned to do anyway. Besides, if the board partners already know the specs (which they do, judging by MSI comments), then Nvidia already knows too.

There is a question at what group targeted is this GPU.

IMO. AMD decided to release the Vega FE for three reasons.
First - they had to meet the H1 2017 release target for Vega, to satisfy their shareholders.
Second - They wanted to earn MONEYZ!
Third - they wanted to bring a tool for developers to learn, and optimize software for this architecture, because it is so vastly different from previous gen. of GCN.

 

[Cloudfire777]

Even with using these 'tricks' to look better the difference seems really big this time around for Vega. Like something is very off and Vega isn't performing as expected for whatever reason.

The performance drop from 0xAA to 8xAA is too big that the demo AMD ran vs HardwareNexus result is believable if you ask me.

We are talking almost 2x the FPS of what HardwareNexus got if you take in AA in to account. No way Vega RX is 2x as fast as Vega FE. Stop dreaming

 

[Veradun]

*If* this is true, then it is a huge marketing blunder.

They clearly and explicitly said "If you are into gaming donàt buy this card as it's not for you".

This obviously doesn't mean RX will be any faster than FE, but that's not the point.

 

[Glo.]

The performance drop from 0xAA to 8xAA is too big that the demo AMD ran vs HardwareNexus result is believable if you ask me.

We are talking almost 2x the FPS of what HardwareNexus got if you take in AA in to account. No way Vega RX is 2x as fast as Vega FE. Stop dreaming

[Deleted]

If the GPU has inactive features that make Vega faster per clock in graphics - yes, it can be two times faster than Fiji, per clock.

To everyone, there are three videos that demo Vega in Doom in 4K ultra settings, with 8xAA. I posted them higher. I suggest checking them, yourself.

Always have to go the personal route, don't you? Keep it up.
-- stahlhart

 

[french toast]

*If* this is true, then it is a huge marketing blunder. They take the time to design a game mode and a professional mode in to the drivers, in order to develop optimizations and test those optimizations on a new architecture. Yet in the very same drivers they have absolutely zero of those features (your claim == Fury driver).

And what's worse, they give absolutely 0 feedback of this major, deficiency in the FE edition. All the marketing statement right now are overly broad and borderline useless (e.g. "drivers are not gimped, just older", "don't compare the RX Vega too much to FE"), guaranteeing speculation like this.

If this is actually the case (which I hope, but am far from convinced), then they only need to utter two simple sentences, and all this drama would have been avoided: "Our Release Driver still has major performance features disabled, expect performance uplift in the future" or "RX Vega will have a noticeble performance uplift, compared to the FE edition".

What would be the downside of saying that? Tipping Nvidia off? ... I don't think they could really do much in a month, they haven't planned to do anyway. Besides, if the board partners already know the specs (which they do, judging by MSI comments), then Nvidia already knows too.

Similar to yourself, whilst I firmly believe there is some major optimisation to be done, I feel if amd had significantly faster drivers in the lab even if they are not ready they would have the common sense to put out a statement saying so, this would largely stop the rampant negative speculation and in fact improve sales of Vega Fe, probably stop some people from jumping on a 1080ti instead of waiting to see what Rx Vega offers.

They have been working on drivers for 6 months now, whilst they might not be polished or ready for release they would certainly know the kind of gains expected with just a single month to go, for people just to think amd are just started writing the drivers now and massive gains are going to be made in one month is crazy.

The fact amd will not put out a statement leads me to believe there is not the massive gains on the table that Rx Vega needs on launch, either that or they are keeping some big secret to provide no gain what so ever,unlikely.

 

[Grubbernaught]

There is a question at what group targeted is this GPU.

IMO. AMD decided to release the Vega FE for three reasons.
First - they had to meet the H1 2017 release target for Vega, to satisfy their shareholders.
Second - They wanted to earn MONEYZ!
Third - they wanted to bring a tool for developers to learn, and optimize software for this architecture, because it is so vastly different from previous gen. of GCN.

Points one and two highly likely but as point for point three, how can one optimise for that which is not enabled?

I am not interested in future speculation, only what is on the table now, and it seems to perform very much like the other revisions of gcn.

Selling these cards without any confirmation or timeline to enable what is essentially promised on the box is a problem for me.

And yes acknowledge nv are also guilty of this (async, 3.5gb), but all the stuff that is the cornerstone of your new architecture? C'mon...

 

[Bacon1]

The performance drop from 0xAA to 8xAA is too big that the demo AMD ran vs HardwareNexus result is believable if you ask me.

We are talking almost 2x the FPS of what HardwareNexus got if you take in AA in to account. No way Vega RX is 2x as fast as Vega FE. Stop dreaming

So how do you explain what we saw in those demos then?

 

[Glo.]

So how do you explain what we saw in those demos then?

AMD cheated, by redesigning the game, to report false details, and resolution.

 

[advt.naveen]

I have few doubts, if amd released this card with Fiji drivers for developers, then without driver support for new features whether developers can able to optimize their code for vega architecture?

 

[Glo.]

If this is actually the case (which I hope, but am far from convinced), then they only need to utter two simple sentences, and all this drama would have been avoided: "Our Release Driver still has major performance features disabled, expect performance uplift in the future" or "RX Vega will have a noticeble performance uplift, compared to the FE edition".

They have said directly: Vega Frontier Edition is not for gamers.

How hard is it to understand that simple sentence?

It is people who jumped to conclusions about this GPU arch, completely ignoring what AMD has said on few ocasions: this is not meant for gamers, that created the drama.

P.S. We absolutely have no guarantee that with new drivers we will see performance uplift on Vega Frontier Edition.

 

[beginner99]

Third - they wanted to bring a tool for developers to learn, and optimize software for this architecture, because it is so vastly different from previous gen. of GCN.

As others have said: If this is a Fiji driver and all the new features are not active and hence performance is slow, how can you learn anything about Vega uArch if the new features aren't enabled and it acts like a Fiji-based GPU?

They have said directly: Vega Frontier Edition is not for gamers.

How hard is it to understand that simple sentence?

I mean I hope you are right and it will perform as indicated in the Doom benches. But I agree with Gideon and many others. This release was pretty bad. The could have added: "RX Vega will have significant higher gaming performance." That alone would have stopped any rumor mill and negative reviews/comments. It leads open what "significant" means but it indicates >= 10% at least. Or they could have actually mentioned issues with drivers.

Actually, was the rasterizer checked both in pro and gaming mode? Maybe the rasterizer poses an issue in pro workloads and must be disabled which leads to serious performance differences between pro and gaming cards. Add to that that gaming driver was not yet fully ready. Once it is, gaming mode will actually do something. But since Vega FE is a pro card, AMD decided to just release it with pro drivers only as these users would not care much about having to wait another month for gaming drivers while also meeting H1 release date.

 

[Veradun]

I mean I hope you are right and it will perform as indicated in the Doom benches. But I agree with Gideon and many others. This release was pretty bad. The could have added: "RX Vega will have significant higher gaming performance." That alone would have stopped any rumor mill and negative reviews/comments. It leads open what "significant" means but it indicates >= 10% at least. Or they could have actually mentioned issues with drivers.

They've done exactly that. It was Raja speaking on his reddit AMA.

 

[wanderica]

They have said directly: Vega Frontier Edition is not for gamers.

How hard is it to understand that simple sentence?

It is people who jumped to conclusions about this GPU arch, completely ignoring what AMD has said on few ocasions: this is not meant for gamers, that created the drama.

P.S. We absolutely have no guarantee that with new drivers we will see performance uplift on Vega Frontier Edition.

We understand. Really, we do. The issue is how is RX Vega different? The only indication we have as to the performance of RX Vega is the FE, the launch of which has been a marketing disaster. AMD also compared FE Vega to Titan XP. You can't have it both ways. FE beats Titan XP handily in workstation tasks and I, for one, think that's worthy of note. However, it gets trounced in gaming by the same Titan XP. Are we not allowed to point out that disparity and criticize AMD for it? After all, it was AMD that started that comparison. If we're supposed to ignore FE's abysmal gaming performance, then AMD should have chosen P4000 as a competitor, not nVidia's halo gaming product.

 

[Veradun]

We understand. Really, we do. The issue is how is RX Vega different?

Guess what? You have to wait for its launch. How strange :>