AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

[Geeforcer]

Alright.

Here is the original statement:

The funny part about "magic" drivers is that anything beyond 5-10%, maybe 15%, would be a fantasy. Greater than 20% seems to be becoming a trend of late. That argument was also founded on Vega competing with 1080ti, which we've already seen without all the features enabled.

With Forza off the table, is there any other game in which we've already seen Vega competitive with 1080 TI, without "all the featured enabled"?

 

[DavidGraham]

What shaky points? The ones that were validated and left your argument flat on it's face? Anyone that looked at the figures could see this coming. I've never said drivers are needed to catch Pascal, just that certain features clearly aren't enabled. That's a rather simple fact. Forza 7 gave competitive performance even without the drivers.

Validated with a single resolution in a single game? that's evidence now against dozens of games that show otherwise? What scientific method is this?

Empirical data was presented, I referenced that with simple facts and benchmarks seem to support my position.

That's one benchmark for you. And Nope, JUST ONE doesn't even qualify to be called empirical, it's called straw grasping.

So much for game ready drivers then. With multiple quick releases to fix major regressions with erratic performance that your competitor lacks.

So fixing bugs is now a problem, but releasing cards for months without a "supposed" imaginary magic drivers is not?

despite being a low level API where only scheduling should be an issue.

DX12 and Vulkan are not low level API, they are LOWER level API, which means they still need a lot of driver intervention.

The real discussion here is why the results are so erratic

Are we seeing the same tests? both vendors exhibited this erratic behavior in the HOCP article, while they didn't suffer it in the original ComputerBase test, so either it's a game problem introduced with the latest patches, or it's a problem with HOCP methodology.

 

[CSI PC]

Validated with a single resolution in a single game? that's evidence now against dozens of games that show otherwise? What scientific method is this?

That's one benchmark for you. And Nope, JUST ONE doesn't even qualify to be called empirical, it's called straw grasping.


So fixing bugs is now a problem, but releasing cards for months without a "supposed" imaginary magic drivers is not?


DX12 and Vulkan are not low level API, they are LOWER level API, which means they still need a lot of driver intervention.


Are we seeing the same tests? both vendors exhibited this erratic behavior in the HOCP article, while they didn't suffer it in the original ComputerBase test, so either it's a game problem introduced with the latest patches, or it's a problem with HOCP methodology.

Worth noting though one advantage of Vulkan is the use of extensions that expose hardware features for a specific manufacturer, Doom Vulkan was a great example the advantages this can bring when comparing AMD to Nvidia performance (in future may swing either way for new games depending upon what extensions are made available and when); although I do wonder if this will become a mess over the years for devs as both AMD and Nvidia spend more time with these *shrug*.
Cheers

 

[Infinisearch]

Is AMD's inferior geometry throughput only affecting framerate in games with high amounts of tessellation or is it a universal framerate sink? I ask because I'm wondering if primitive shaders being enabled will only help certain games or all games? The HBCC being enabled will likely only bring performance in cases when video memory is overcommitted. DSBR should affect all or most games with a performance benefit... I forgot is it enabled right now or not? So primitive shaders might only help in some cases, HBCC is relegated to "finewine" technology, and DSBR can affect all games but its questionable how much a performance boost it can provide. So I'm not so sure Vega will be able to catch up to the 1080ti with driver updates, except maybe in titles with high tesselation factors.

 

[Anarchist4000]

With Forza off the table, is there any other game in which we've already seen Vega competitive with 1080 TI, without "all the featured enabled"?

None that I'd consider reliable, however the design is based more around asynchronous behavior. Where a number of bottlenecks can be worked around. Very few current titles have the modernized engines. Mostly just DX12 tacked on or are asychronous, but not significantly so.

Worth noting though one advantage of Vulkan is the use of extensions that expose hardware features for a specific manufacturer, Doom Vulkan was a great example

I wouldn't say Doom is a great example, as the extensions used are more or less the console platform coming with the new shader model. All IHVs support the extensions which are becoming standard. A true extension mess would be something only one IHV reasonably supports, not that there aren't places for those. Primitive shaders for example are likely an extension, however it's unclear if they could even be abstracted onto other hardware. It needs to happen, but the graphics pipeline is changing with it.

Is AMD's inferior geometry throughput only affecting framerate in games with high amounts of tessellation or is it a universal framerate sink? I ask because I'm wondering if primitive shaders being enabled will only help certain games or all games?

Some games more than others based on optimizations. Any app with effective culling probably wouldn't see huge gains if there is nothing to discard.

The primitive rate is a secondary concern. Single vs multithreaded (async) would be a better analogy. So long as a frame isn't ridiculously bound by primitive rate it shouldn't matter much. Not allowing for primitive shaders being more efficient in terms of bandwidth. Primitive rate on a title with lots of async work shouldn't have a problem. The flip side is all the hardware for processing primitives is idle as there is nothing to do most of the time. Wasted die space.

The HBCC being enabled will likely only bring performance in cases when video memory is overcommitted.

Surprisingly no, going off some recent tests, as it's more or less compressing the streaming operations. Discarding and reordering pages based on actual use so there are less stalls. It's only when resources are fully committed, therefore no transfers, that it makes no difference.

 

[Deleted member 13524]

Is AMD's inferior geometry throughput only affecting framerate in games with high amounts of tessellation or is it a universal framerate sink? I ask because I'm wondering if primitive shaders being enabled will only help certain games or all games?

High-end GCN GPUs are generally limited by geometry performance in PC games.
At the same time, Vega 56/64 in particular seem to have an excess of compute performance for games, as seen by the fact that Vega 64 = 56 at ISO clocks, both cards have a hard time scaling >20% above Fiji despite up to 50% higher clocks and their performance-per-clock-per-ALU is below Polaris 10.
(Note: this last one may be because Polaris 10 has a fixed-function primitive discard accelerator, which Vega doesn't have because supposedly this would be done on the ALUs/NCUs).

AFAICT, all things point to Vega having been designed to use a portion of this overkill ALU throughput to reduce geometry, fillrate and bandwidth demands from raw scene to "viewable" scene output through the use of primitive shaders. Even more if primitive shaders can even partially make use of FP16.

Polaris 10's slide about the effects of the primitive discard accelerator says it all: it reduces geometry bottlenecks, fillrate bottlenecks (MSAA intensive scenes) and "frees internal bandwidth resources" (to which I'd guess it leads to less cache hits)



Now what the primitive shaders could/will do when active is a completely different discussion on when / if ever they will be active.
Plus, just because enabling primitive shaders would decrease bottlenecks on several fronts it doesn't mean we should expect gigantic performance upgrades on all games. But if it increases performance by 10% in general and 15% in some titles then it'll already be a significant difference in benchmarks against the competition, especially in the Vega 64 vs. GTX 1080 case.
The GTX 1070 Ti is also bound to bring a significant leap over the regular 1070 and therefore surpass the Vega 56, but put 10-15% more in the later and the competition could become fierce again.

The HBCC being enabled will likely only bring performance in cases when video memory is overcommitted.

Actually, there are reports of HBCC giving a ~12% performance uplift in PUBG. Even though the game is most probably not using >8GB of VRAM, it seems HBCC might be scrambling memory in a way that increases effective bandwidth.

So I'm not so sure Vega will be able to catch up to the 1080ti with driver updates, except maybe in titles with high tesselation factors.

I don't think AMD is reasonably expecting Vega 64 to compete with GP102 cards within its lifetime on anything other than "gaming experience" when using Vega+FreeSync vs GP102+VSync. Maybe it will be close in a couple of cases like Wolfenstein 2 that uses instrinsic shaders, FP16 and is extremely compute dependent:

https://twitter.com/x/status/913947204146073601

 

[Infinisearch]

Vega 56/64 in particular seem to have an excess of compute performance for games, as seen by the fact that Vega 64 = 56 at ISO clocks, both cards have a hard time scaling >20% above Fiji despite up to 50% higher clocks and their performance-per-clock-per-ALU is below Polaris 10.

I was actually wondering about that. Has anyone studied and compared AMD's 4096 core parts (fiji,vega64) vs their smaller parts in terms of scaling? I was thinking maybe the 4 "Shader Engine" of the big chips was holding back the performance of the chips. Nvidia seems to partition its big chips into 6 partition with less total cores than AMD's big chips... so basically I was thinking there are more bottlenecks in AMD's design.

 

[Bondrewd]

Has anyone studied and compared AMD's 4096 core parts (fiji,vega64) vs their smaller parts in terms of scaling?

Yes. Vega was hands down the worst.

 

[Deleted member 13524]

I was actually wondering about that. Has anyone studied and compared AMD's 4096 core parts (fiji,vega64) vs their smaller parts in terms of scaling? I was thinking maybe the 4 "Shader Engine" of the big chips was holding back the performance of the chips. Nvidia seems to partition its big chips into 6 partition with less total cores than AMD's big chips... so basically I was thinking there are more bottlenecks in AMD's design.

Looking at performance comparisons between Fiji, Polaris and Vega, I think it's easy to conclude that Fiji and Vega are limited by having 4 geometry engines, whereas Polaris 10 is not.

You can tell that by looking at a geometry-intensive game like Witcher 3 (at 1080p to remove bandwidth from the equation):

The Fury X has 40% more compute resources and 57% higher theoretical fillrate than the RX580 but they get the exact same score. In the same scenario, Vega 64 performs like a Fury X + 40% which is practically the clock difference between them.
Then Vega 64 gets only 63% higher performance than the RX580, despite having 100% more compute resources and 130% more fillrate.



And then if you look at a compute-intensive game like Vulkan Doom (1440p here because at 1080p Vega hits the 200FPS ceiling):

The Fury X has 40% more compute resources and performs 39.5% better than the RX580. Then Vega 64 has 100% more compute resources and performs 83% faster than the RX580 (and I'm guessing the average in Vega isn't higher because it's often hitting the 200FPS cap on its maximum framerates).



In the end, Vega is indeed behaving like an overclocked Fury X, which is turn has a significantly lower performance-per-clock-per-CU than Polaris 10.
And Polaris 10's only major architectural difference to Fiji was the introduction of the Primitive Discard Accelerator, which is something Vega doesn't have.. because it was supposed to have been replaced by running Primitive Shaders in the NCUs.

 

[ECH]

Is it really a 1080/1080 TI win in Forza 7? Their forums are littered with stutter, hitching and online game crashing since 387.92 was released.

So I've read somewhere on this forum that the new Windows update 1709 would fix alot of stuttering and such in Forza 7. Today the update arrived and I installed it immediately. The first thing I did when the installation was completed was to fire up Forza. I quickly noticed the almost constant stuttering and the game felt exactly like when I first installed Creators Update and hadn't turned off Game Mode. I went into windows settings to check if it had been re-enabled. Since update 1709 there is no longer an option to turn game mode off! And since there is no way to "disable fullscreen optimization" for any Windows Store apps, well.. We're screwed.

Before the 1709 update the game worked mostly fine (apart from the memory leak). No constant stuttering and I could race for maybe an hour before noticing any kind of stutter and had to restart the game. Now it's unplayable for me.

My setup:
Intel i7 4790K (clocked to 4.6ghz)
16GB DDR3
nVidia GTX 1080
144hz monitor (playing Forza with unlocked framerate)

https://forums.forzamotorsport.net/turn10_postst109783_-PC-Windows-10-update-1709-broke-Forza-7.aspx


https://forums.forzamotorsport.net/turn10_postst103899_Hitching-Stuttering-in-game.aspx

i Get solid 60fps but i'll get micro stutter here and there and its very annoying

i7 and 1080ti

https://forums.forzamotorsport.net/turn10_postst109840_Any-Stutter-Fix-Yet.aspx

It's truly littered with complaints. Not everyone posts their PC specs. But I still suspect it's related to the most recent drivers. Let's take a peek in the Official 387.96 driver forum.

https://forums.geforce.com/default/...hread-released-10-9-17-/post/5227778/#5227778
https://forums.geforce.com/default/...hread-released-10-9-17-/post/5227349/#5227349
https://forums.geforce.com/default/...hread-released-10-9-17-/post/5227578/#5227578
https://forums.geforce.com/default/...hread-released-10-9-17-/post/5227141/#5227141
https://forums.geforce.com/default/...hread-released-10-9-17-/post/5226888/#5226888
https://forums.geforce.com/default/...hread-released-10-9-17-/post/5226964/#5226964
https://forums.geforce.com/default/...hread-released-10-9-17-/post/5226743/#5226743
https://forums.geforce.com/default/...hread-released-10-9-17-/post/5226242/#5226242
https://forums.geforce.com/default/...hread-released-10-9-17-/post/5226097/#5226097
https://forums.geforce.com/default/...hread-released-10-9-17-/post/5226030/#5226030

I'll stop here, you get the point. I think HOCP might have inadvertently showed, as I don't see any mention of it, that although the 387.92 drivers increase FPS in Forza 7 the game is a lot less
smoother then it was from the previous driver.
https://www.hardocp.com/article/2017/10/16/forza_motorsport_7_video_card_performance_update/2

 

[DavidGraham]

I think HOCP might have inadvertently showed, as I don't see any mention of it, that although the 387.92 drivers increase FPS in Forza 7 the game is a lot less smoother then it was from the previous driver.

ComputerBase tested the new driver on both the 1060 and 1080, and found increased fps consistency and smoothness compared to the old driver.
https://www.computerbase.de/2017-09/forza-7-benchmark/3/

New Is it really a 1080/1080 TI win in Forza 7? Their forums are littered with stutter, hitching and online game crashing since 387.92 was released.

The stutter of the game is present since the demo was introduced, and it happens on several configurations from both vendors. It's effect appears after several races. It's probably due to Game Mode, the latest Windows update supposedly fixed the issue.

•We are investigating stutters reported in the PC demo on specific configurations.

https://forums.forzamotorsport.net/...a-Motorsport-7-Demo-FAQ-and-Known-Issues.aspx

If you are experiencing hitches/stuttering while playing Forza Motorsport 7, we suggest disabling Game Mode in Windows and close any overclocking, hardware monitoring, streaming, recording and any nonessential apps while we investigate this.

https://forums.forzamotorsport.net/turn10_postst103899_Hitching-Stuttering-in-game.aspx

we have an excellent piece of news for all Forza Motorsport 7 fans out there. The Windows 10 Fall Creator Update fixes the annoying stuttering issues affecting the game.

http://windowsreport.com/fix-forza-motorsport-stuttering/

 

[ECH]

ComputerBase tested the new driver on both the 1060 and 1080, and found increased fps consistency and smoothness compared to the old driver.
https://www.computerbase.de/2017-09/forza-7-benchmark/3/


The stutter of the game is present since the demo was introduced, and it happens on several configurations from both vendors. It's effect appears after several races. It's probably due to Game Mode, the latest Windows update supposedly fixed the issue.


https://forums.forzamotorsport.net/...a-Motorsport-7-Demo-FAQ-and-Known-Issues.aspx


https://forums.forzamotorsport.net/turn10_postst103899_Hitching-Stuttering-in-game.aspx


http://windowsreport.com/fix-forza-motorsport-stuttering/

I think we need to consider the multitude of people, from the exampled forums, that say otherwise. In cases that allowed diff. in drivers; some have shown improved
behavior not using 387.92 drivers. Which points to the driver itself as being the problem. Although the drivers improve performance. Which is not in question.

Also, HOCP results clearly show Forza 7 is stuttering even at the higher frame rates with 387.92 drivers. There is no question about it. What is the question is what's causing it?
Because to me, it looks like the results were on different racing tracks. Not saying that it was. The fluctuation is so bad it give me the appearance that they are from different racing tracks.

Perhaps a 3rd driver is needed for Forza 7 (as this is the 2nd one so far).

 

[DavidGraham]

I think we need to consider the multitude of people, from the exampled forums, that say otherwise.

What otherwise? the game stutters since the demo on all hardware, before the new drivers.

HOCP results clearly show Forza 7 is stuttering even at the higher frame rates with 387.92 drivers.

This has nothing to do with the drivers, the erratic behavior is happening on AMD cards as well.

 

[ECH]

What otherwise? the game stutters since the demo on all hardware, before the new drivers.

This has nothing to do with the drivers, the erratic behavior is happening on AMD cards as well.

Then perhaps we should understand what is it that you are disputing then?

As for the HOCP results. That's the results they got. If we remember PCGH we clearly see a difference.
https://www.computerbase.de/2017-09/forza-7-benchmark/3/

What 387.92 does, at least for PCGH, is bring the frametimes around Vega, roughly 12ms or so. At least for the track they benched on. There results are more controlled and consistent. Which is what you expect.
However, even though they are reporting that many are still reporting stutter while using 387.92. And, it's not just in Forza 7 they are experience it in.
Nvidia's own 387.92 thread is also rifted with complaints about stutter with that driver.
That's why I'm focusing in on the 387.92 driver. This isn't a conspiracy. 387.92 gave increase in performance at the cost of immersion.

I'm sure it will be addressed in another driver update. But for now, it is what it is. shurg.

 

[DavidGraham]

However, even though they are reporting that many are still reporting stutter while using 387.92. And, it's not just in Forza 7 they are experience it in.

I use this driver, I experienced no stutters in a variety of games. The links you posted are a variety of driver problems not related to stutters. Only two mention stutters, one of them is related to triple surround and SLi setup, so I don't actually see the relation here. And again stuttering in Forza 7 is a documented wide spread problem since the days of the demo, the Fall windows update was to fix it. And it appears it did:

After more than an hour playing Forza Motorsport 7, we did not experience a single stutter. We don’t know what Microsoft has done, however the game stuttered like crazy prior to this update...

Thanks to the Windows 10 Fall Creators Update, we were able to unlock the framerate and retain really high framerates. Framepacing was great and there weren’t any stutters even after our extensive testing.

http://www.dsogaming.com/news/windo...-fixes-forza-motorsport-7s-stuttering-issues/

 

[Anarchist4000]

Looking at performance comparisons between Fiji, Polaris and Vega, I think it's easy to conclude that Fiji and Vega are limited by having 4 geometry engines, whereas Polaris 10 is not.

The 4SEs should scale by clockspeed though. With the Witcher3 numbers you posted, the 470, 580, Nano, and Fury are all ~58FPS. I highly doubt all of those are running similar core/memory clocks in testing. Then the Vegas vary a bit, but relative to Polaris they aren't 60% higher clocks. Barring the off chance Polaris only sustained 1GHz while Vega managed 1.6GHz. To be a 4SE limit some sort of latency independent of core and probably memory would need to be in play. SEs shouldn't be in their own clock domain either.

 

[Infinisearch]

Looking at performance comparisons between Fiji, Polaris and Vega, I think it's easy to conclude that Fiji and Vega are limited by having 4 geometry engines, whereas Polaris 10 is not.

I wasn't only referring to geometry throughput, I was also thinking about inter and intra Shader Engine buses/crossbars and caches/buffers. Basically I was wondering if the network supporting the CU's wasn't up to snuff.

 

[Deleted member 13524]

The 4SEs should scale by clockspeed though. With the Witcher3 numbers you posted, the 470, 580, Nano, and Fury are all ~58FPS. I highly doubt all of those are running similar core/memory clocks in testing.

The 470 isn't doing 58 FPS, it's doing 51 FPS or ~15% lower than a 580, which is close to their clock difference. Here we have the RX580 with more CUs than the 470 but their performance difference boils down to their clocks. And since both the 470 and 580 have the same number of geometry engines, this further implies that Witcher 3 is geometry intensive for GCN cards (like pretty much all Gameworks titles).
The Nano is a difficult card to evaluate because it can clock up to 1GHz so at 1080p if many blocks are just waiting for the results from the geometry engines then I'd guess it could be clocking high enough to get close to the Fury X.


As for the RX580 vs Fury vs Fury X, it indicates that while Polaris cards have to spend a lot more time with the compute/pixel shaders in each frame, they spend a lot less time on geometry (because of higher clocks + primitive discard accelerator), hence the similar performance between Polaris 10 and Fiji.

Then the Vegas vary a bit, but relative to Polaris they aren't 60% higher clocks. Barring the off chance Polaris only sustained 1GHz while Vega managed 1.6GHz.

I didn't compare Vega to Polaris in clocks so I don't know where you got those 60% from. I said Vega 64 has close to 100% more compute throughput and 130% more fillrate than RX580, and that's taken from the pre-calculated compute and fillrate numbers that AMD provides in their documentation (and you can find on wikipedia).

 

[ECH]

And again stuttering in Forza 7 is a documented wide spread problem since the days of the demo, the Fall windows update was to fix it. And it appears it did:

I don't have any stutter in that game. It's a pretty smooth experience for me. And there are other's that I've seen that echo that experience.
Just because a group with a certain hardware setup experiences a problem doesn't mean everyone else "must experience the problem". When it's clear they have a different PC setup.

Here is another way to look at this.
Below is a 1080 TI using the newly released Fall Creator's Update. This is his experience:

Performance has deterioted for this user for this user's experience.






Below is a review using the Vega RX 56 using the newly released Fall Creator's Update. This is his experience:

Performance has improved for this user's experience.

I'm quite taken by this. I've not seen anything like this from a Windows Update in...I do not know when...
Something worth investigating. If this holds out to be true it has the potential to defend itself against the 1070 TI.

 

[itsmydamnation]

I'm quite taken by this. I've not seen anything like this from a Windows Update in...I do not know when...
Something worth investigating. If this holds out to be true it has the potential to defend itself against the 1070 TI.

If the difference is real ( im not convinced) i think it will come down to scheduling differences. I tired the fall up date and in witcher 3 and superposition i saw 0 difference. In witcher 3 with croud/npc count set to ultra im CPU limited on 3770k @ 4.3 and Superposition barely uses 1/2 a core.