Overclock.net › Forums › Industry News › Hardware News › [wccftech] AMD RX Vega 64 Outperforms NVIDIA GTX 1080 Ti By Up To 23% In DX12 Forza 7
New Posts  All Forums:Forum Nav:

[wccftech] AMD RX Vega 64 Outperforms NVIDIA GTX 1080 Ti By Up To 23% In DX12 Forza 7 - Page 25

post #241 of 497
Quote:
Originally Posted by ToTheSun! View Post

Possibly pixel fill rate. It's the most apparent conclusion, anyway.
Should be less of an issue when DSBR is enabled. Unless it already is... Which I doubt.
Quote:
Originally Posted by geoxile View Post

Could be bandwidth. It's effectively less than the fury x's despite propping up a bigger, faster chip
Vega has been shown to get significant improvement by OC-ing the memory... So this might be true. Can be easily tested by anyone that has a Vega card with OC'd memory. We would have to look at the relative difference in performance increase between 1080p and 4k before and after OC of HBM.
post #242 of 497
Quote:
Originally Posted by geoxile View Post

Could be bandwidth. It's effectively less than the fury x's despite propping up a bigger, faster chip

Don't think bandwidth, both around the same with a bit of an overclock and stock.



post #243 of 497
Quote:
Originally Posted by NightAntilli View Post

The GTX 970 is performing very well in those PUBG benchmarks though, if we compare it to its direct competitor the R9 390. And the GTX 1060 is performing better than a Fury X... And still Vega is outperforming the GTX 1080 line... Most curious. Primitive discard accelerator at work?
havent you thought it could be

-Tessellation(Primitive discard)
-Rasterization(Rasterizer/DSBR) or Pixel Engine(ROPs)
-CPU Overhead(Lower on Vega at 1080)


Compare Fury X to 1060 at the same time Vega 64 vs 1080Ti in 1080/1440/4k
Wanted: [WTB] GPU upgrade
$210 (USD) or best offer
  
Reply
Wanted: [WTB] GPU upgrade
$210 (USD) or best offer
  
Reply
post #244 of 497
Quote:
Originally Posted by PontiacGTX View Post

havent you thought it could be

-Tessellation(Primitive discard)
-Rasterization(Rasterizer/DSBR) or Pixel Engine(ROPs)
-CPU Overhead(Lower on Vega at 1080)


Compare Fury X to 1060 at the same time Vega 64 vs 1080Ti in 1080/1440/4k
I mentioned the primitive discard as a possibility.

As for rasterization, I'm not sure if DSBR is enabled or not... Considering nVidia already does tile-based rasterization, I can't see why this would give Vega some sort of advantage. And if it is the case, why these games and not others?
ROPs improvements from Polaris to Vega aren't exactly great as far as I know... So I doubt that's the reason.
CPU overhead is another questionable one... Why would Vega suffer from less CPU overhead compared to Polaris? The GTX 970 is beating the RX 580.

The fill rate is another possibility though... It's one of the weaknesses of Vega compared to the nVidia cards. The GTX 1080 has a higher pixel fill rate than Vega 64 Liquid... And pixel fill rate becomes important at higher resolutions.
post #245 of 497





Throw a dart and see where it lands. That's my guess biggrin.gif.

But something is going on. And until AMD spills the beans or something is leaked there is really no way of telling.
The only real consistent trend is that they've nailed 1080p. Now it''s time to get 1440 and 4K results.
post #246 of 497
Primitive shaders and next gen geometry aren't even enabled yet
New and Shiny
(18 items)
 
  
CPUMotherboardGraphicsRAM
Ryzen 7 1700 ASUS ROG Crosshair VI Hero Sapphire RX Vega 56 G.SKILL TridentZ F4-3200C14D 
Hard DriveHard DriveHard DriveHard Drive
Kingston HyperX 3K Crucial MX300 Western Digital Black Western Digital Green 
Hard DriveCoolingOSMonitor
Western Digital Red Noctua NH-U14S Windows 10 Pro Dell U2414H 
MonitorKeyboardPowerCase
Dell P2414H Ducky One Corsair RM650x NZXT H440 White 
MouseAudio
Logitech G502 Proteus Spectrum Schiit Modi2U & Magni2U 
  hide details  
Reply
New and Shiny
(18 items)
 
  
CPUMotherboardGraphicsRAM
Ryzen 7 1700 ASUS ROG Crosshair VI Hero Sapphire RX Vega 56 G.SKILL TridentZ F4-3200C14D 
Hard DriveHard DriveHard DriveHard Drive
Kingston HyperX 3K Crucial MX300 Western Digital Black Western Digital Green 
Hard DriveCoolingOSMonitor
Western Digital Red Noctua NH-U14S Windows 10 Pro Dell U2414H 
MonitorKeyboardPowerCase
Dell P2414H Ducky One Corsair RM650x NZXT H440 White 
MouseAudio
Logitech G502 Proteus Spectrum Schiit Modi2U & Magni2U 
  hide details  
Reply
post #247 of 497
Quote:
Originally Posted by NightAntilli View Post

Vega has been shown to get significant improvement by OC-ing the memory... So this might be true. Can be easily tested by anyone that has a Vega card with OC'd memory. We would have to look at the relative difference in performance increase between 1080p and 4k before and after OC of HBM.

It does. Depends on the application, type of workload, obviously. I tried it out in Blackwake (an odd-ball game, I know) and from 945 to 1050 gained 7% and 1100 gain 9%. Directly tied to GPU bound frame rate.

In relation to this specific PUBG benchmark I don't think any of Vega's features or specs play the limiting factor however.
Quote:
Originally Posted by PontiacGTX View Post


Compare Fury X to 1060 at the same time Vega 64 vs 1080Ti in 1080/1440/4k

That is why I don't believe, in this benchmark using the 17.8.1 drivers, that the game had been even rectified at this point by AMD. The deficit that the Radeon cards exhibit is not normal, especially at 4K. It's beyond horrible but back in-line with the theoretical 'Fury on Overdrive' benchmarks given the GCN 1.3/4 vs. GCN 1.5 results from the Fury X to Vega 56. If Vega was using DSBR/PD or RPM you'd see the 64 do much better than sub-1070 at 4K.

Heck... if it was properly optimized for the game to begin with you'd at least see the 64 bounce around the vanilla 1080 (all other deficits aside).

At the middle, 580/Fury X/390, we're seeing the horrible GPU optimization because they are being (unoptimized but regardlessly artificially) GPU limited. At the high end we're seeing NVIDIA run up against that CPU sided limitation and Vega go beyond because that API sided limitation simply does not apply so soon here for AMD.
--
It is interesting that you mention DSBR and discard Pontiac:
I think it may be possible that Vega is showing a ever so slight CPU boundary in some applications and this may be where the draw stream and discard features become useful - theoretically they should alleviate draw call boundaries when using DX11, shouldn't they? Given that they throw away tons of polygons, models and game assets - everything rendered on screen being an actual draw call. I could see how that could potentially impact Vega when a game is using the DX11 rendering path -- alleviate the rendering thread>push the cpu limitation forward & decrease rendering work and memory usage on the GPU side>double rate fp16 (RPM) to increase rendering throughput
Edited by pengs - 10/5/17 at 11:18am
post #248 of 497
Quote:
Originally Posted by pengs View Post

It does. Depends on the application, type of workload, obviously. I tried it out in Blackwake (an odd-ball game, I know) and from 945 to 1050 gained 7% and 1100 gain 9%. Directly tied to GPU bound frame rate.
Was that at 1080p or 4k? What is the improvement on the opposite?
Quote:
Originally Posted by pengs View Post

It is interesting that you mention DSBR and discard Pontiac:
I think it may be possible that the 64 is showing a direct CPU boundary (in this benchmark and possibly others) and this may be where the draw stream and discard features become useful - theoretically they should alleviate draw call boundaries when the using DX11, shouldn't they? Given that they throw away tons of polygons, models and game assets - everything rendered on screen being an actual draw call. I could see how that could potentially impact Vega when a game is using the DX11 rendering path -- alleviate the rendering thread>push the cpu limitation forward & decrease rendering work and memory usage on the GPU side>double rate fp16 (RPM) to increase rendering throughput
Hm... This could very well be the case. It's hard to tell whether it really does reduce CPU overhead since we don't know exactly how the culling works. Because it might as well put all the strain on the CPU and cull only for the GPU side of things, meaning the GPU load would be reduced but the CPU overhead would stay the same. But... If we assume the fill rate is what becomes the issue at 4k, then the performance we're seeing actually makes sense at the lower resolutions, if the primitive discard is working, as long as it also alleviates the CPU.
Edited by NightAntilli - 10/5/17 at 11:26am
post #249 of 497
Quote:
Originally Posted by NightAntilli View Post

Was that at 1080p or 4k? What is the improvement on the opposite?

That was 1440p. Much smaller at 1080p, like 3%.

Yeah, best case without all of these features is probably slightly below 1080 Ti. There's no doubt the Ti is a pixel pusher.
Quote:
Originally Posted by NightAntilli View Post

Hm... This could very well be the case. It's hard to tell whether it really does reduce CPU overhead since we don't know exactly how the culling works. Because it might as well put all the strain on the CPU and cull only for the GPU side of things, meaning the GPU load would be reduced but the CPU overhead would stay the same. But... If we assume the fill rate is what becomes the issue at 4k, then the performance we're seeing actually makes sense at the lower resolutions, if the primitive discard is working, as long as it also alleviates the CPU.

It should, the more you cull the less geometry. Even if it doesn't have much of an effect on the CPU side there's no doubt it will on the GPU side.

Removing draw calls is what I'm really interested in. Does the game calling for something alone qualify as a draw call or is it only when it enters the rendering pipeline?
How soon is this culling happening? I guess that's the question.

Doesn't really matter via Vulkan and DX12. Good use of it should render draw call limits completely irrelevant.
Edited by pengs - 10/5/17 at 12:02pm
post #250 of 497
Quote:
Originally Posted by pengs View Post

That was 1440p. Much smaller at 1080p, like 3%.

Yeah, best case without all of these features is probably slightly below 1080 Ti. There's no doubt the Ti is a pixel pusher.
It should, the more you cull the less geometry. Even if it doesn't have much of an effect on the CPU side there's no doubt it will on the GPU side.

Removing draw calls is what I'm really interested in. Does the game calling for something alone qualify as a draw call or is it only when it enters the rendering pipeline?
How soon is this culling happening? I guess that's the question.

Draw calls are issued by the cpu. There's no way AMD is actually any better at handling it than nvidia, it's all driver side. Culling should happen early in the pipe in vertex/geometry shaders. If it reaches CUs and it turns out to be hidden geometry (obscured or too small to matter) you wasted the CU's time because it won't be seen but it took up cache and cycles. Iirc the was a DICE presentation talking about exactly this problem. Improving drawcall throughput doesn't stop the fact that the gpu gets stuck with tons of wasted work because of slow or improper culling.
New and Shiny
(18 items)
 
  
CPUMotherboardGraphicsRAM
Ryzen 7 1700 ASUS ROG Crosshair VI Hero Sapphire RX Vega 56 G.SKILL TridentZ F4-3200C14D 
Hard DriveHard DriveHard DriveHard Drive
Kingston HyperX 3K Crucial MX300 Western Digital Black Western Digital Green 
Hard DriveCoolingOSMonitor
Western Digital Red Noctua NH-U14S Windows 10 Pro Dell U2414H 
MonitorKeyboardPowerCase
Dell P2414H Ducky One Corsair RM650x NZXT H440 White 
MouseAudio
Logitech G502 Proteus Spectrum Schiit Modi2U & Magni2U 
  hide details  
Reply
New and Shiny
(18 items)
 
  
CPUMotherboardGraphicsRAM
Ryzen 7 1700 ASUS ROG Crosshair VI Hero Sapphire RX Vega 56 G.SKILL TridentZ F4-3200C14D 
Hard DriveHard DriveHard DriveHard Drive
Kingston HyperX 3K Crucial MX300 Western Digital Black Western Digital Green 
Hard DriveCoolingOSMonitor
Western Digital Red Noctua NH-U14S Windows 10 Pro Dell U2414H 
MonitorKeyboardPowerCase
Dell P2414H Ducky One Corsair RM650x NZXT H440 White 
MouseAudio
Logitech G502 Proteus Spectrum Schiit Modi2U & Magni2U 
  hide details  
Reply
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Hardware News
Overclock.net › Forums › Industry News › Hardware News › [wccftech] AMD RX Vega 64 Outperforms NVIDIA GTX 1080 Ti By Up To 23% In DX12 Forza 7