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

Page 50 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Status
Not open for further replies.

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
That really doesn't add up.

The performance Delta Between Fiji and Vega FE in workstation benchmarks, are exactly like the performance deltas between Normal and "pro" AMD drivers.

There are two possibilities.

A) Vega FE effectively has Pro drivers, even if they are unsigned.
B) Vega FE has such radically different HW that it behaves just like is has Pro drivers without them.


Occam's razor says A) is the most likely solution (Being simple and straightforward).

B) is unlikely in the extreme. What do you think will happen when Vega FE gets "real" Pro drivers will it get the same kind of boost that Pro drivers give on top of the ones we see here. So the SNX-02 benchmark will go from 5 Times faster and get another 5 X boost and be 25 times faster?

No. AMD is selling this card as a Workstation card, and they gave it Workstations drivers that boost those workstation benchmarks, just like their Pro drivers do.

Thinking that the Workstation benchmarks represent an enormous HW boost over Fiji is just wishful thinking, leading you away from the obvious answer: It's just Workstation drivers.
So how come 2048 CUDA core GTX 980 with 4.2 TFLOPs of compute power, was faster or on par with 2880 CUDA core 5.5 TFLOPs GTX 780 Ti in compute oriented workloads?

http://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-review/20
67749.png

67747.png

67746.png

67745.png

Kepler had 256 KB registry File size available to 192 CUDA cores in each SM.
Maxwell has 256 KB Registry File size available to just 128 CUDA cores in each SM.

In essence, those 128 CUDA cores in Maxwell had 100% of performance of 192 CUDA cores from Kepler.

This is definition of throughput of GPU. This change affected not only gaming performance of the GPU but also compute performance. Nvidia essentially increased throughput of the GPU by making it less starved for resources.

As for Vega vs Fiji. Increase in throughput of the GPU you can achieve just by increasing the cache size versus previous version of GCN. Polaris increased L2 cache 4 times compared to previous versions of GCN, including Fiji, which was reflected in gaming benchmarks. What you see in the benchmarks quoted, by Gamers Nexus is the difference of increased throughput of GPU. If AMD changed the way GPU handles Registy Files, and increased size of them - you should also see reflection of this in the applications you use. Most likely - this is the reason why Vega is per clock faster than Fiji, and why AMD touted that it has higher IPC compared to previous generations of GCN.

In sheer compute throughput, Vega will be on par with Volta and GP100 chip.

Vega does not have signed professional drivers. Whole point of professional drivers is signing them. They are heavily optimized much more than consumer drivers. That is the reason why you pay 4000$ for essentially GTX 1080 with different name, and different drivers. They are designed for highest reliability and compute performance.
 

CatMerc

Golden Member
Jul 16, 2016
1,114
1,149
136
PCGH tested the card and now we at least know where the problem lies:
http://www.pcgameshardware.de/Vega-...elease-AMD-Radeon-Frontier-Edition-1232684/3/
They tested the b3d suite and you can clearly see that the effective bandwidth at the moment ist terrible. Black texture with color compression at fury level and random texture bandwidth even way lower than fury. Whatever the reason for this low bandwidth might be, it's the reason for the slow performance. It's just starving on bandwidth.
That is indeed awful bandwidth. 250GB/s out of theoretical 480GB/s?
 
  • Like
Reactions: french toast

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
That is indeed awful bandwidth. 250GB/s out of theoretical 480GB/s?
That would also tell why in applications that rely on bandwidth of memory, the GPU actually sees decrease in performance per clock.

Somehow in current state of software(both drivers and application) it performs just like Fiji.
 

PeterScott

Platinum Member
Jul 7, 2017
2,605
1,540
136
So how come 2048 CUDA core GTX 980 with 4.2 TFLOPs of compute power, was faster or on par with 2880 CUDA core 5.5 TFLOPs GTX 780 Ti in compute oriented workloads?

http://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-review/20
67749.png

67747.png

67746.png

67745.png

Kepler had 256 KB registry File size available to 192 CUDA cores in each SM.
Maxwell has 256 KB Registry File size available to just 128 CUDA cores in each SM.

In essence, those 128 CUDA cores in Maxwell had 100% of performance of 192 CUDA cores from Kepler.

This is definition of throughput of GPU. This change affected not only gaming performance of the GPU but also compute performance. Nvidia essentially increased throughput of the GPU by making it less starved for resources.

As for Vega vs Fiji. Increase in throughput of the GPU you can achieve just by increasing the cache size versus previous version of GCN. Polaris increased L2 cache 4 times compared to previous versions of GCN, including Fiji, which was reflected in gaming benchmarks. What you see in the benchmarks quoted, by Gamers Nexus is the difference of increased throughput of GPU. If AMD changed the way GPU handles Registy Files, and increased size of them - you should also see reflection of this in the applications you use. Most likely - this is the reason why Vega is per clock faster than Fiji, and why AMD touted that it has higher IPC compared to previous generations of GCN.

In sheer compute throughput, Vega will be on par with Volta and GP100 chip.

Vega does not have signed professional drivers. Whole point of professional drivers is signing them. They are heavily optimized much more than consumer drivers. That is the reason why you pay 4000$ for essentially GTX 1080 with different name, and different drivers. They are designed for highest reliability and compute performance.

Comparing NVidia to AMD like that just muddies the waters.

Comparing AMD to AMD like was done previously, makes what I stated (it's just the drivers) is totally obvious, and it perfectly explains the Vega workstation benchmarks are comparatively better than Fiji on Workstation benchmarks, while remaining similar in gaming benchmarks.

You just don't want to believe it, so you are grasping at straws.
 

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
Comparing NVidia to AMD like that just muddies the waters.

Comparing AMD to AMD like was done previously, makes what I stated (it's just the drivers) is totally obvious, and it perfectly explains the Vega workstation benchmarks are comparatively better than Fiji on Workstation benchmarks, while remaining similar in gaming benchmarks.

You just don't want to believe it, so you are grasping at straws.
Yes, it is obviously me, who does not want to believe. I do not have to believe. I worked in professional industry and I know how this works. I am trying to tell you that there is more to GPUs than just cores, ROPs, Schedulers, and memory bandwidth. Its you, who is refusing to learn this.

I am not comparing Nvidia to AMD. I am just showing you example how you increase the throughput of GPUs, and how they actually work. Its up to you to resist broader picture.

Both architectures are pretty similar in terms of what has to be done to achieve certain goal. Both architectures have to Registry File sizes, both architectures have L2, and L1 cache, both architectures have certain compute and graphical throughputs.

Once again. Vega Frontier Edition does not have signed professional drivers. It does not come with them. If it would come - it would wipe Quadro P5000 out of this world, because of the sheer compute horsepower difference. How can you forget about this? P5000 is just GTX 1080 with 9.2 TFLOPs of compute power. Vega, even with 1.44 GHz core clock has 11.7, and should be faster.

Vega does not have signed professional drivers. It is as simple as it can be. The difference per clock compared to Fiji in compute benchmarks is because of increased the L2 cache size in Vega, and most likely - increased Registry File size.

One more thing. AMD had to change how Vega approaches and handles Registry Files because of new scheduling and load balancing system, and the fact that Pixel Engine is right now client of L2 cache, rather than Memory Controller. And this last bit should tell you the most about the throughput of the architectures if you know how it works.
 
  • Like
Reactions: french toast

Head1985

Golden Member
Jul 8, 2014
1,863
685
136
PCGH tested the card and now we at least know where the problem lies:
http://www.pcgameshardware.de/Vega-...elease-AMD-Radeon-Frontier-Edition-1232684/3/
They tested the b3d suite and you can clearly see that the effective bandwidth at the moment ist terrible. Black texture with color compression at fury level and random texture bandwidth even way lower than fury. Whatever the reason for this low bandwidth might be, it's the reason for the slow performance. It's just starving on bandwidth.
I am telling this last 2 weeks that vega is memory bandwidth bottleneck lol.Also rest results is not better.
 

Magic Hate Ball

Senior member
Feb 2, 2017
290
250
96
Comparing NVidia to AMD like that just muddies the waters.

Comparing AMD to AMD like was done previously, makes what I stated (it's just the drivers) is totally obvious, and it perfectly explains the Vega workstation benchmarks are comparatively better than Fiji on Workstation benchmarks, while remaining similar in gaming benchmarks.

You just don't want to believe it, so you are grasping at straws.

AMD is not going to sell full workstation functionality and speed with a card as powerful as Vega FE for only $1000... not happening.

That would horrifyingly gut their profits on workstation cards by a factor of 3-4x or more!
 

french toast

Senior member
Feb 22, 2017
988
825
136
I am telling this last 2 weeks that vega is memory bandwidth bottleneck lol.Also rest results is not better.
It was pretty obvious to everyone that there was bandwidth issues, the puzzle was the reasoning behind it.
Thre seems to be a bug in the drivers which means Vega is effectively using HALF the bandwidth, this along with the bandwidth saving tiled rasterizer not being used is likely the cause of most of the weird performance.

There is more of this to play out.
 

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
Supposedly any clock regulation on core of Vega automatically puts the memory at 500 MHz.

Would't it be not strange if the software was reporting 945 MHz for memory, but effectively it would run at lower speeds, because of not yet ready BIOS of the GPU?


Only one question pops into my mind right now, after reading about all of problems of Vega.

What AMD was thinking with releasing this GPU?
 

Magic Hate Ball

Senior member
Feb 2, 2017
290
250
96
Supposedly any clock regulation on core of Vega automatically puts the memory at 500 MHz.

Would't it be not strange if the software was reporting 945 MHz for memory, but effectively it would run at lower speeds, because of not yet ready BIOS of the GPU?


Only one question pops into my mind right now, after reading about all of problems of Vega.

What AMD was thinking with releasing this GPU?

They promised a Vega product 1H 2017 to share holders... they had to deliver something.
 

Despoiler

Golden Member
Nov 10, 2007
1,966
770
136

96Firebird

Diamond Member
Nov 8, 2010
5,709
316
126
Its plain ridiculous to say a 500$ card is "midrange", as veradun said, what the hell is a gtx 1070? 1060? 1050? 1030?.
People have been sipping that jensen sponsored kool- aid :)

You do realize it is the AMD fanboys that call the 1080 a "mid-range card" because they like to complain that Nvidia is charging so much for a "mid-range card". Performance be damned, according to some mid-range is defined by die size and GPU codename...
 
  • Like
Reactions: tential

DiogoDX

Senior member
Oct 11, 2012
746
277
136
PCGH tested the card and now we at least know where the problem lies:
http://www.pcgameshardware.de/Vega-...elease-AMD-Radeon-Frontier-Edition-1232684/3/
They tested the b3d suite and you can clearly see that the effective bandwidth at the moment ist terrible. Black texture with color compression at fury level and random texture bandwidth even way lower than fury. Whatever the reason for this low bandwidth might be, it's the reason for the slow performance. It's just starving on bandwidth.
And the primitive dischage is also not working when compared with the 580 result in the polygon test.

Only ALU and fillrate testes seems OK with the GPU specs (realy sad to see only 64 ROPs).
 

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
And the primitive dischage is also not working when compared with the 580 result in the polygon test.

Only ALU and fillrate testes seems OK with the GPU specs (realy sad to see only 64 ROPs).
This only proves that Vega does NOT have enabled new features that increase performance. Zero. Nada. Nothing.
 

Samwell

Senior member
May 10, 2015
225
47
101
And the primitive dischage is also not working when compared with the 580 result in the polygon test.

Only ALU and fillrate testes seems OK with the GPU specs (realy sad to see only 64 ROPs).

How did you come to this conclusion? I think you need to look at the 100% strip results for that to work. There it's nearly 2x Fury X at the same clock. What seems not in there is primitive shader, but i don't know whether primitive shader needs application support.
 

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
How did you come to this conclusion? I think you need to look at the 100% strip results for that to work. There it's nearly 2x Fury X at the same clock. What seems not in there is primitive shader, but i don't know whether primitive shader needs application support.
Primitive Discard Accelerator from Polaris architecture in the work.

Has anybody spotted that per clock Vega has lower fillrate than Fiji...?
 

Elixer

Lifer
May 7, 2002
10,376
762
126
One more thing. AMD had to change how Vega approaches and handles Registry Files because of new scheduling and load balancing system, and the fact that Pixel Engine is right now client of L2 cache, rather than Memory Controller. And this last bit should tell you the most about the throughput of the architectures if you know how it works.
What do you mean by "Registry Files"?
 
  • Like
Reactions: Aenra

Samwell

Senior member
May 10, 2015
225
47
101
Has anybody spotted that per clock Vega has lower fillrate than Fiji...?

Yes and i don't think this will change with new drivers. It might be a trade-off for the rops now beeing the clients of the l2 cache. The bad random texture bandwidth also doesn't seem to me like something driver related. Turning on the Binning Rasterizer will push for higher random and black bandwidth, but this won't suddenly change the ratio between them. But that seems more like something, which you could maybe fix with a new bios. But i've the fealing that at RX Vega launch not all of this strange results will change.
 

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
What do you mean by "Registry Files"?
https://en.wikipedia.org/wiki/Register_file
Yes and i don't think this will change with new drivers. It might be a trade-off for the rops now beeing the clients of the l2 cache. The bad random texture bandwidth also doesn't seem to me like something driver related. Turning on the Binning Rasterizer will push for higher random and black bandwidth, but this won't suddenly change the ratio between them. But that seems more like something, which you could maybe fix with a new bios. But i've the fealing that at RX Vega launch not all of this strange results will change.
If it is affected by not available features - yes it can change.

I have never seen such unready product to be launched, to actually launch...
 

Elixer

Lifer
May 7, 2002
10,376
762
126
Those are completely different things (registry file vs register file) hence the confusion.

It makes zero sense, at all.

Sheer clock speed difference should put it between GTX 1080 and Titan X(much closer to Titan X). And that is even excluding the architecture improvements.

This whole "release" makes zero sense at all, it was clearly a rush job, only to meet a deadline that was set by the CEO.

Some people keep yapping about XY&Z being disabled in the drivers, but, AMD denied this at B3D forums. Drivers are "older" != Drivers are "disabled"/gimped. (It wouldn't make sense to do that to the drivers, the first silicon can't be THAT broken.)
If that was the truth, we have to wait until RX launch to find out.

On paper, Vega seems good. In the real world, depending on what is being run, it can be anywhere from great to good to OK to bad to poor.
I think we can all agree that it shouldn't be anything below OK.

The facts right now is, Vega runs hot when pushed, is a power monster when pushed, and drivers are "older".
 

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
Those are completely different things (registry file vs register file) hence the confusion.



This whole "release" makes zero sense at all, it was clearly a rush job, only to meet a deadline that was set by the CEO.

Some people keep yapping about XY&Z being disabled in the drivers, but, AMD denied this at B3D forums. Drivers are "older" != Drivers are "disabled"/gimped. (It wouldn't make sense to do that to the drivers, the first silicon can't be THAT broken.)
If that was the truth, we have to wait until RX launch to find out.

On paper, Vega seems good. In the real world, depending on what is being run, it can be anywhere from great to good to OK to bad to poor.
I think we can all agree that it shouldn't be anything below OK.

The facts right now is, Vega runs hot when pushed, is a power monster when pushed, and drivers are "older".
Yes, my bad for pronouncing it wrong. Its just how I used to call this over past years, in discussions. Register Files are correct form.

I also agree that it appears AMD decided to meet the timeline, for Vega, with FE, nothing else.

On the other hand, it delivers what AMD has promised in terms of compute performance(is it really...?), but lacks gaming performance, and this product was never targeted at gamers, after all.
 
  • Like
Reactions: Aenra
Status
Not open for further replies.