AMD Radeon RX Vega 64 and 56 Reviews [*UPDATED* Aug 28]

[Rasterizer]

he stated as fact that there are several disabled features that are performance enhancing. I just want proof of his statement

Providing good evidence that enabling NGG fast path in Vega drivers should provide a significant performance uplift is pretty straightforward: People have spent years talking about how Fiji was a fundamentally flawed GPU design that was severely front end bottlenecked, and that this is why it struggled against the 980ti. Vega, as it currently stands on the newest public driver, is at least as front end bottlenecked by culled geometry performance as Fiji was, as it is currently realising an identical ~3.75 triangles per clock of culled geometry throughput:

(Incidentally, the numbers given her for list culled geometry performance exactly match the Fiji and Vega native pipeline numbers from the Vega whitepaper). That Vega is, in practice, front end bottlenecked is also easily confirmed by noting that Vega 56 performs 100% the same as Vega 64 at the same clocks despite having 8 less CUs.

The Vega whitepaper claims that primitive shaders are able to significantly increase front end culled geometry throughput prior to any fixed function culling (of the type done in Polaris with the Primitive Discard Accelerator) from 4 triangles per clock to up to 17. If even half of this claimed gain was actually realised by implementing primitive shaders, this would literally double Vega's culled polygon throughput and should substantially alleviate Vega's front end bottleneck, allowing Vega to much more readily keep its shader array fully saturated.

Unless you intend to claim that the feature doesn't exist that RTG simply made it up in the whitepaper, or that Fiji wasn't extremely well known to be front end bottlenecked by its geometry performance and that doubling (or more!) Fiji's culled geometry performance wouldn't have substantially increased its performance in games, I don't see how you can claim that doubling (again, or more) Vega's culled polygon throughput won't substantively increase Vega's performance in games.

 

[Headfoot]

The only thing I can see is that they are having a very difficult time getting the drivers to automatically make use of the feature without performance regression, or they were overly optimistic on development time for the feature.

The only way it makes logical sense is if the primitive geometry feature actually doesnt make enough of an impact to withhold the whole launch, or they had some other hard deadline (big contracts?) they couldnt push back. Maybe they expect tuning this feature to be a 6 month long endeavor and getting the earliest first version out before launch isn't a priority given the long tail?? Very hard to decipher the reasoning

 

[geoxile]

Providing good evidence that enabling NGG fast path in Vega drivers should provide a significant performance uplift is pretty straightforward: People have spent years talking about how Fiji was a fundamentally flawed GPU design that was severely front end bottlenecked, and that this is why it struggled against the 980ti. Vega, as it currently stands on the newest public driver, is at least as front end bottlenecked by culled geometry performance as Fiji was, as it is currently realising an identical ~3.75 triangles per clock of culled geometry throughput:

(Incidentally, the numbers given her for list culled geometry performance exactly match the Fiji and Vega native pipeline numbers from the Vega whitepaper). That Vega is, in practice, front end bottlenecked is also easily confirmed by noting that Vega 56 performs 100% the same as Vega 64 at the same clocks despite having 8 less CUs.

The Vega whitepaper claims that primitive shaders are able to significantly increase front end culled geometry throughput prior to any fixed function culling (of the type done in Polaris with the Primitive Discard Accelerator) from 4 triangles per clock to up to 17. If even half of this claimed gain was actually realised by implementing primitive shaders, this would literally double Vega's culled polygon throughput and should substantially alleviate Vega's front end bottleneck, allowing Vega to much more readily keep its shader array fully saturated.

Unless you intend to claim that the feature doesn't exist that RTG simply made it up in the whitepaper, or that Fiji wasn't extremely well known to be front end bottlenecked by its geometry performance and that doubling (or more!) Fiji's culled geometry performance wouldn't have substantially increased its performance in games, I don't see how you can claim that doubling (again, or more) Vega's culled polygon throughput won't substantively increase Vega's performance in games.

I doubt the geometry throughout is the bottleneck, especially since games historically do their own culling before the gpu has to do anything. The packing of object instances sounds more likely to give a boost in performance. Iirc wavefronts (and warps) of x size have to have all x threads work in tandem, meaning an unfilled wavefront is the same as having idle threads and unused shaders. If you can fill the wavefront more efficiently you should be able to get more utilization out of the shaders, or make better use of the raw compute power of a gpu. For Vega that means tapping into more of that 13TF shader block. The best case for any gpu is to make 100% use of the shader block, considering it's usually the biggest and most power hungry part of a gpu.

That's assuming the new work distributor isn't active yet. If it is already working then Vega might be a screw up.

 

[swilli89]

I have missed a lot of this Vega news (in terms of spottily doing very well in some isolated cases) as I was forced to leave my home in Houston..

Finally found some shelter so I can relax and look into some tech stuff

Vega 56 obviously is the card to have (especially once AIBs come around), and I believe there are still some pretty major performance increases as the drivers continue to receive updates.

 

[Magic Hate Ball]

I have missed a lot of this Vega news (in terms of spottily doing very well in some isolated cases) as I was forced to leave my home in Houston..

Finally found some shelter so I can relax and look into some tech stuff

Vega 56 obviously is the card to have (especially once AIBs come around), and I believe there are still some pretty major performance increases as the drivers continue to receive updates.

Hey, glad you could join us again! Glad you're safe!

 

[Rifter]

Providing good evidence that enabling NGG fast path in Vega drivers should provide a significant performance uplift is pretty straightforward: People have spent years talking about how Fiji was a fundamentally flawed GPU design that was severely front end bottlenecked, and that this is why it struggled against the 980ti. Vega, as it currently stands on the newest public driver, is at least as front end bottlenecked by culled geometry performance as Fiji was, as it is currently realising an identical ~3.75 triangles per clock of culled geometry throughput:

(Incidentally, the numbers given her for list culled geometry performance exactly match the Fiji and Vega native pipeline numbers from the Vega whitepaper). That Vega is, in practice, front end bottlenecked is also easily confirmed by noting that Vega 56 performs 100% the same as Vega 64 at the same clocks despite having 8 less CUs.

The Vega whitepaper claims that primitive shaders are able to significantly increase front end culled geometry throughput prior to any fixed function culling (of the type done in Polaris with the Primitive Discard Accelerator) from 4 triangles per clock to up to 17. If even half of this claimed gain was actually realised by implementing primitive shaders, this would literally double Vega's culled polygon throughput and should substantially alleviate Vega's front end bottleneck, allowing Vega to much more readily keep its shader array fully saturated.

Unless you intend to claim that the feature doesn't exist that RTG simply made it up in the whitepaper, or that Fiji wasn't extremely well known to be front end bottlenecked by its geometry performance and that doubling (or more!) Fiji's culled geometry performance wouldn't have substantially increased its performance in games, I don't see how you can claim that doubling (again, or more) Vega's culled polygon throughput won't substantively increase Vega's performance in games.

Thats certainly one way to look at it, in a best case senario.

Another way to look at it is that they learned nothing from Fiji or at least were unable to correct it, and were forced to just release what they have as is and hope they find a way in software to fix it later. Which from the benchmarks im seeing so far looks to be exactly what happened, as it shows very little improvement over fiji.

Reality is likely somewhere in between those two options, and over the next several years AMD will likely increase performance like they always do, AMD has shown with historical evidence they are able to wring more performance from their hardware over time vs Nvidia. This im not disputing, prior evidence shows this is very likely, and this is the main reason i usually support AMD's GPU's and is why i have a rx480 right now, i was banking on it picking up more performance vs 1060 and being stronger in new DX12 titles. i just dont think there is going to be a big driver release that increases performance 20%+ across the board like some people think is coming.

AMD always talks big, even with ryzen which was a huge success saw its share of "wait for software" press releases from AMD about its performance in certain areas and with certain OS's. They dont always however deliver on these promises.

As it sits right now vega is years behind Nvidia in perf/watt, and at best AMD flagship card trades blows with NVidias third tier performing card the 1080, leaving the titan and 1080Ti clearly retaining both the performance and perf/watt crown. And this is with volta is just around the corner.

Look at it this way, we have people saying look at the white papers and look at what AMD is saying, and yes these features sure look good on paper, but so far paper has not translated at all into any real measurable performance gains in in real games. Remember these are the same people who ran a "poor volta" marketing campaign and then in reality released a card that doesnt even compare to the top two cards in the Pascal lineup let alone Volta. You cant always or even mostly trust a companies own marketing about their own products. Independent testing is always needed to form a unbiased opinion.

I suggest we do away with all this marketing BS, and rate and judge the card on its actual real world performance verified by independent testing , in actual real world games. Instead of clinging to any hopes and dreams AMD is trying to get people to buy into. Real world GPU performance plays your games, not hopes and dreams.

Hopefully im wrong, and if i am ill be back to say so, but i dont see a magical huge performance increasing driver coming out to fix this botched launch. Best case IMO is they eventually improve performance to get to 1080Ti levels in 1-3 years time, likely with more power usage though. And rely on Navi to really compete with volta.

 

[Rasterizer]

I doubt the geometry throughout is the bottleneck

People have literally been talking about how Fiji was front end geometry bottlenecked for two years. Here is a just a small selection from this very forum:

I don't believe ROPs are the bottleneck but I'd wager on it being the front end along with the somewhat anemic geometry engines ...

Full Fiji diagram.Its bottleneck by front-end(only 4x pipelines for Too much CU/Shaders)

David Kanter says he think this too. Fiji got bottlenecked with Frontend, but as AMD ran out of die area they had to bake the chip this way.

Fiji is front-end bottleneck (refer to original Fury X review at Computerbase.de where they talked to AMD directly about design challenges & tradeoffs, AMD is fully aware of the potential bottleneck), same setup for many more shaders. AMD do need to optimize it better to extract peak performance.

Yep, and we seem to be seeing the same general trends with Fury X. The front end simply isn't big enough to feed all those shaders (pitcarin also defeated the 7950 in certain games - not just synthetic benchmarks).

It should. It's a larger die (over 500mm^2, compared to 471mm^2 for GP102) and with more raw power (13 TFlops FP32, compared to about 11.3 for GTX 1080 Ti). And the implementation of tiled rendering should have closed the perf/TFlop gap with Nvidia's offerings. If it's lagging this badly, then either there are hardware bottlenecks (most likely the fact that the front end is still as starved as Fiji, with only 4 shader engines we've seen on the Linux drivers) or else the drivers suck.

Primitive shaders are directly and explicitly targeted at overcoming this 4 triangles per clock front end bottleneck by incorporating geometry culling to an earlier stage of the rendering pipeline and by deferral of attribute shading until after that culling. As per AMD themselves acknowledge in the Vega whitepaper:

The “Vega” 10 GPU includes four geometry engines which would normally be limited to a maximum throughput of four primitives per clock, but this limit increases to more than 17 primitives per clock when primitive shaders are employed.

 

[Rasterizer]

Look at it this way, we have people saying look at the white papers and look at what AMD is saying, and yes these features sure look good on paper, but so far paper has not translated at all into any real measurable performance gains in in real games.

So your contention then is that RTG is simply lying in the whitepaper and that primitive shaders cannot even deliver half of the claimed improvement in culled geometry throughput? Because even half of the claimed benefits would mean literally doubling Vega's culled polygon throughput on an architecture widely known to be front end bottlenecked by its four geometry engines.

 

[geoxile]

Front-end is more than just the geometry shaders... Front end also includes the work distributor, which ngg purports to improve. Higher geometry throughput is just another benefit, and even if you improvement throughput if you don't fill the wavefronts you won't be able to fully utilize the shaders anyway as far as I'm aware, meaning you still end up with a bottleneck due to shader underutilization

 

[Konan]

One function is this from the whitepaper. AMD has always suffered from poor utilization vs Nvidia. One reason why it took many more AMD shaders to equal GeForce GPUs. This can close the gap. There was an AMD patent discussed here several years ago with variable shader clusters for different sized wavefronts. This appears to have the same end result in packing a full 64 wavefront with different instances instead of breaking up the wavefronts into smaller values.

"Another innovation of “Vega’s” NGG is improved load balancing across multiple geometry engines. An intelligent workload distributor (IWD) continually adjusts pipeline settings based on the characteristics of the draw calls it receives in order to maximize utilization. One factor that can cause geometry engines to idle is context switching. Context switches occur whenever the engine changes from one render state to another, such as when changing from a draw call for one object to that of a different object with different material properties. The amount of data associated with render states can be quite large, and GPU processing can stall if it runs out of available context storage. The IWD seeks to avoid this performance overhead by avoiding context switches whenever possible. Some draw calls also include many small instances (i.e., they render many similar versions of a simple object). If an instance does not include enough primitives to fill a wavefront of 64 threads, then it cannot take full advantage of the GPU’s parallel processing capability, and some proportion of the GPU's capacity goes unused. The IWD can mitigate this effect by packing multiple small instances into a single wavefront, providing a substantial boost to utilization."

Regarding IWD (intelligent workload distributor) GCN works best if you feed it a full complement of 64 threads in a wavefront. And like NVIDIA's GPUs, GCN schedules workloads based on a wavefront (AMD), or warp (NVIDIA), or whatever the the two companies want to call "a group of threads".

But GCN works similar to Nvidia GPU's by executing in single instruction, multiple data (or thread) fashion. So what this does is it figures out if the work the drivers submitted to the GPU can fill a wavefront. If it can't, then it sees if another wavefront can fill in the rest of the empty slots. Pretty much like how HyperThreading works (which is kind of funny)
Traditionally, GPU programs through API, which go to drivers, which then compile what the API has given it into what the GPU understands. The API however does define what sort of hardware features the GPU must support in order to be compliant, but those features aren't necessarily tied to an API. For example, DirectX 9.0 required 24-bit minimum precision on floating point calculations. But that doesn't mean 24-bit floating point calculations is specific to DirectX 9.0. DirectX 10 required programmable pipelines, but that's not something specific to DirectX 10. OpenGL or Vulkan can use those same hardware features.

Or in the words, Intelligent Workload Distributor (AMD) figures out how assign Compute Units beforehand for maximum throughput. The IWD compute looks to work on the warp level.
Interestingly with Nvidia and what they have done with Volta, they have taken to innovate further by introducing Independent Thread Scheduling (ITS) which is there to prevent excessive stalling in the execution chain by allowing CUDA Cores to switch to another thread if there is a stall. Nvidia have moved to look at the Thread level now that their version of the IWD at the Warp level works well.

 

[Glo.]

Front-end is more than just the geometry shaders... Front end also includes the work distributor, which ngg purports to improve. Higher geometry throughput is just another benefit, and even if you improvement throughput if you don't fill the wavefronts you won't be able to fully utilize the shaders anyway as far as I'm aware, meaning you still end up with a bottleneck due to shader underutilization

This is correct. There is no point in increasing the throughput of the cores, by increasing amount of the work they can do each clock cycle, when you are not able to fill them with work. And filling the cores with work is the role of all of new graphics GCN features: Draw Stream Binning Rasterizer, Primitive Shaders, Intelligent Work Distributor, high Bandwidth Cache controller. All of those features are designed to increase the work available to the cores, each time, each second to avoid stalls in the pipeline.

 

[Elfear]

Thats certainly one way to look at it, in a best case senario.


I suggest we do away with all this marketing BS, and rate and judge the card on its actual real world performance verified by independent testing , in actual real world games. Instead of clinging to any hopes and dreams AMD is trying to get people to buy into. Real world GPU performance plays your games, not hopes and dreams.

Done.

Now can we get back to discussing the new tech baked into Vega?

 

[maddie]

Thats certainly one way to look at it, in a best case senario.

Another way to look at it is that they learned nothing from Fiji or at least were unable to correct it, and were forced to just release what they have as is and hope they find a way in software to fix it later. Which from the benchmarks im seeing so far looks to be exactly what happened, as it shows very little improvement over fiji.

Reality is likely somewhere in between those two options, and over the next several years AMD will likely increase performance like they always do, AMD has shown with historical evidence they are able to wring more performance from their hardware over time vs Nvidia. This im not disputing, prior evidence shows this is very likely, and this is the main reason i usually support AMD's GPU's and is why i have a rx480 right now, i was banking on it picking up more performance vs 1060 and being stronger in new DX12 titles. i just dont think there is going to be a big driver release that increases performance 20%+ across the board like some people think is coming.

AMD always talks big, even with ryzen which was a huge success saw its share of "wait for software" press releases from AMD about its performance in certain areas and with certain OS's. They dont always however deliver on these promises.

As it sits right now vega is years behind Nvidia in perf/watt, and at best AMD flagship card trades blows with NVidias third tier performing card the 1080, leaving the titan and 1080Ti clearly retaining both the performance and perf/watt crown. And this is with volta is just around the corner.

Look at it this way, we have people saying look at the white papers and look at what AMD is saying, and yes these features sure look good on paper, but so far paper has not translated at all into any real measurable performance gains in in real games. Remember these are the same people who ran a "poor volta" marketing campaign and then in reality released a card that doesnt even compare to the top two cards in the Pascal lineup let alone Volta. You cant always or even mostly trust a companies own marketing about their own products. Independent testing is always needed to form a unbiased opinion.

I suggest we do away with all this marketing BS, and rate and judge the card on its actual real world performance verified by independent testing , in actual real world games. Instead of clinging to any hopes and dreams AMD is trying to get people to buy into. Real world GPU performance plays your games, not hopes and dreams.

Hopefully im wrong, and if i am ill be back to say so, but i dont see a magical huge performance increasing driver coming out to fix this botched launch. Best case IMO is they eventually improve performance to get to 1080Ti levels in 1-3 years time, likely with more power usage though. And rely on Navi to really compete with volta.

This is actually a fair statement of doubt. Yes, the onus is on AMD to provide the performance.

On the other hand, this is a forum of enthusiasts talking about new features and what they might mean. Was it Zlatan (I forget), who claimed that these new features remove most if not all of previous AMD vs Nvidia bottlenecks?

What I dislike are posters outright lying, and for example, claiming that the feature is already enabled and it led to no performance increase, when we have it from sources that it's not enabled in release drivers.

 

[Rifter]

Done.

Now can we get back to discussing the new tech baked into Vega?

Right, so you agree with me that Vega was a failure and cannot compete with current Nvidia offerings either in perf/watt or overall performance. And at current street prices not in perf/$ either.

 

[maddie]

I have missed a lot of this Vega news (in terms of spottily doing very well in some isolated cases) as I was forced to leave my home in Houston..

Finally found some shelter so I can relax and look into some tech stuff

Vega 56 obviously is the card to have (especially once AIBs come around), and I believe there are still some pretty major performance increases as the drivers continue to receive updates.

Great you being safe. Is it as bad as it looks?

 

[Rifter]

So your contention then is that RTG is simply lying in the whitepaper and that primitive shaders cannot even deliver half of the claimed improvement in culled geometry throughput? Because even half of the claimed benefits would mean literally doubling Vega's culled polygon throughput on an architecture widely known to be front end bottlenecked by its four geometry engines.

My contention is that at this current time none of AMD's hopes and dreams have translated into any measurable real world performance increase. And that im doubtful that a major performance increase is around the corner.

 

[cfenton]

I find it very difficult to believe that AMD has a bunch of half-finished features that substantially improve performance. Or, at least, I have a hard time believing that any of them are even close to being implemented. If Vega 64 was closer to a 1080TI, then they could sell it for closer to 1080TI prices. They would basically be throwing away money by selling these cards at their current prices if a big improvement was coming soon. They are also now stuck at $499 for Vega 64, regardless of any improvements they make. Raising MSRP because they fixed their botched launch would be a PR nightmare. I just don't think they'd release Vega when they did if there were substantial improvements coming soon.

 

[Phynaz]

Providing good evidence that enabling NGG fast path in Vega drivers should provide a significant performance uplift is pretty straightforward: People have spent years talking about how Fiji was a fundamentally flawed GPU design that was severely front end bottlenecked

snip

Unless you intend to claim that the feature doesn't exist that RTG simply made it up in the whitepaper, or that Fiji wasn't extremely well known to be front end bottlenecked by its geometry performance and that doubling (or more!) Fiji's culled geometry performance wouldn't have substantially increased its performance in games, I don't see how you can claim that doubling (again, or more) Vega's culled polygon throughput won't substantively increase Vega's performance in games.

From Sebbi:

Modern games don't spend their whole frame time rasterizing geometry. Lighting, post processing, etc take significant chunk of GPU time (up to 50%), and during that time the geometry pipelines are idling. Only a small part of the frame is geometry bound.

https://forum.beyond3d.com/posts/1997086/

 

[kawi6rr]

Right, so you agree with me that Vega was a failure and cannot compete with current Nvidia offerings either in perf/watt or overall performance. And at current street prices not in perf/$ either.

In the reviews I read Vega 56 is out performing the 1070 so part of your statement above is false. There is a current review on TH that will show you where these two cards stand in terms of outright performance.

 

[Arachnotronic]

I find it very difficult to believe that AMD has a bunch of half-finished features that substantially improve performance. Or, at least, I have a hard time believing that any of them are even close to being implemented. If Vega 64 was closer to a 1080TI, then they could sell it for closer to 1080TI prices. They would basically be throwing away money by selling these cards at their current prices if a big improvement was coming soon. They are also now stuck at $499 for Vega 64, regardless of any improvements they make. Raising MSRP because they fixed their botched launch would be a PR nightmare. I just don't think they'd release Vega when they did if there were substantial improvements coming soon.

It's just wishful thinking, which is the phrase that I think best characterizes this whole multi-year Vega debacle.

 

[Hitman928]

I find it very difficult to believe that AMD has a bunch of half-finished features that substantially improve performance. Or, at least, I have a hard time believing that any of them are even close to being implemented. If Vega 64 was closer to a 1080TI, then they could sell it for closer to 1080TI prices. They would basically be throwing away money by selling these cards at their current prices if a big improvement was coming soon. They are also now stuck at $499 for Vega 64, regardless of any improvements they make. Raising MSRP because they fixed their botched launch would be a PR nightmare. I just don't think they'd release Vega when they did if there were substantial improvements coming soon.

I'm not saying that I buy into the belief that a large (25% +) amount of performance is being left out of Vega due to these disabled features (I'd guess it's more like 12% on average with some outliers maybe hitting 25%), but if we assume that Vega does have 25% + performance available with these features being enabled, they could easily fix the MSRP problem with a product refresh. Get it out before Volta and it should be fine.

 

[maddie]

I find it very difficult to believe that AMD has a bunch of half-finished features that substantially improve performance. Or, at least, I have a hard time believing that any of them are even close to being implemented. If Vega 64 was closer to a 1080TI, then they could sell it for closer to 1080TI prices. They would basically be throwing away money by selling these cards at their current prices if a big improvement was coming soon. They are also now stuck at $499 for Vega 64, regardless of any improvements they make. Raising MSRP because they fixed their botched launch would be a PR nightmare. I just don't think they'd release Vega when they did if there were substantial improvements coming soon.

This same pricing argument was stated before the Polaris and Ryzen launch. A few laughed at the possibility of a $200 Polaris and a $500 Ryzen.The new AMD management appears to be approaching pricing differently than the past, at least until marketshare improves. For certain, they have set up R&D and production to be very lean, and I'm certain that quite a few will also say that the marketing dept might be too lean.

We'll see.

 

[Arachnotronic]

I'm not saying that I buy into the belief that a large (25% +) amount of performance is being left out of Vega due to these disabled features (I'd guess it's more like 12% on average with some outliers maybe hitting 25%), but if we assume that Vega does have 25% + performance available with these features being enabled, they could easily fix the MSRP problem with a product refresh. Time it right around the time Volta comes out and it would work out fine.

It's this kind of thinking that led to way overinflated expectations for Vega in the first place.

 

[Hitman928]

It's this kind of thinking that led to way overinflated expectations for Vega in the first place.

How so?

 

[Rifter]

In the reviews I read Vega 56 is out performing the 1070 so part of your statement above is false. There is a current review on TH that will show you where these two cards stand in terms of outright performance.

How is it false? The 56 is not higher performaing than the 1080,1080
Ti. And the 56 is not cheaper or lower power usage vs the 1070 its competing directly with.