• We should now be fully online following an overnight outage. Apologies for any inconvenience, we do not expect there to be any further issues.

GCN... Direct3D/DirectX 12?

Page 3 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Enigmoid

Platinum Member
Sep 27, 2012
2,907
31
91
The article didn't say MASSIVE ...

That's an overstatement on your part ...

There is an overhead but not to the degree you imply since voxelization (which enforces conservative rasterization by default in it's pipeline) was easily achieved in The Tomorrow Children on the PS4 ...

Sure, but to what extent?

Voxelization is being used solely on static objects with relatively large voxels. You, can, as Virge said, program pretty much anything. Its just that the overhead may be quite large.

When you look at the code for some of these approaches you can easily see how much overhead there will be compared to hardware.

https://developer.nvidia.com/gpugems/GPUGems2/gpugems2_chapter42.html

An example would be PCSS which shows minimal drops on Maxwell and the 285 but large drops on other cards.
 

RussianSensation

Elite Member
Sep 5, 2003
19,458
765
126
Now I'm just trying to figure out which will be fully compatible in OS X and net a good perf/$ ratio.

Here is my advice if the choice is difficult for you.

1) Go look at the games you play and see how well CF/SLI works in them. For example, if you play a lot of GW titles, you are very likely to get better SLI support on Day 1. If FC4 and AC Unity are on your list, then go with 970s for sure.

2) If custom display resolutions matter, NV is also a great option.

3) Do you want the ability to have the cards turn off its fans off at idle or in light gaming? An R9 295X2 cannot do that.

4) Can you live without PhysX? Some gamers who love Batman Arkham series or Borderlands don't enjoy those games without PhysX.

5) Do you see yourself buying a 4K HDMI 2.0 monitor in the living room and using a PC with it in the next 2 years?

6) Side-step and warranty. EVGA has good warranty and step-up program. You get a minimum of 3 years vs. I believe just 2 on the XFX 295X2. EVGA FTW with 1317mhz Boost is $335 on Newegg.

7) Resale value and ability to resell as easily. I think it's easier to sell 2 decoupled GPUs than a single one with 2 GPUs onboard. To be able to resell the 295X2, you'll really want to find someone who wants to deal with CF and a power hungry card. I think it'll be easier to offload each 970 for $175 in 2 years than to sell an R9 295X2 for $350.

The more difficult choice are after-market R9 290s for $480-500 vs. GTX970 SLI for $660. Now, that one is a much harder comparison. Stock vs. stock performance is 94-95% as good and you have $150+ for Steam games, an SSD bump. If you are looking for max overclocked performance, in your small case 970 SLI > R9 290s. Something like MSI 970 Gaming would still overclock well and run quieter than after-market max overclocked 290s. Not sure if this is worth $150+ extra to you though.

Anyway, to help your decision, imo DX12 is less important than the other factors I outlined.
 
Last edited:

ThatBuzzkiller

Golden Member
Nov 14, 2014
1,120
260
136
Sure, but to what extent?

Voxelization is being used solely on static objects with relatively large voxels. You, can, as Virge said, program pretty much anything. Its just that the overhead may be quite large.

When you look at the code for some of these approaches you can easily see how much overhead there will be compared to hardware.

https://developer.nvidia.com/gpugems/GPUGems2/gpugems2_chapter42.html

An example would be PCSS which shows minimal drops on Maxwell and the 285 but large drops on other cards.

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

Voxelization works with dynamic objects just as easily with static objects as shown in the video ...

Rendering with voxels MUST be done using a more coarse resolution since there's a lot of memory overhead associated with voxels and the 2nd gen Nvidia Maxwell isn't an exception to this ...

As I had asserted earlier, the overhead of conservative rasterization is relatively moderate. At it's worst it'll only tank the performance by 13% on a R9 270X, on even higher end hardware like the R9 290X it's no more than 7% to get that feature enabled ...

Hardware support for conservative rasterization makes most sense on lower end hardware like iGPUs ...

Just keep using those shaders and ROPs. No need to add more functionality to the rasterizers. The issue will be gone before you even know it ...
 

Enigmoid

Platinum Member
Sep 27, 2012
2,907
31
91
https://www.youtube.com/watch?v=fAsg_xNzhcQ

Voxelization works with dynamic objects just as easily with static objects as shown in the video ...

Rendering with voxels MUST be done using a more coarse resolution since there's a lot of memory overhead associated with voxels and the 2nd gen Nvidia Maxwell isn't an exception to this ...

As I had asserted earlier, the overhead of conservative rasterization is relatively moderate. At it's worst it'll only tank the performance by 13% on a R9 270X, on even higher end hardware like the R9 290X it's no more than 7% to get that feature enabled ...

Hardware support for conservative rasterization makes most sense on lower end hardware like iGPUs ...

Just keep using those shaders and ROPs. No need to add more functionality to the rasterizers. The issue will be gone before you even know it ...

Sure it can work. The point is its only being used for static objects because it is quite demanding.

I have no idea where you are pulling those numbers from.
 

ThatBuzzkiller

Golden Member
Nov 14, 2014
1,120
260
136
Sure it can work. The point is its only being used for static objects because it is quite demanding.

I have no idea where you are pulling those numbers from.

http://soconne.blogspot.ca/2013/04/single-pass-gpu-voxelization-demo.html

It's not the voxelization process that's intensive according to Sean O'Connell ...

He was using a GTX 660 which doesn't have hardware support for either viewport multicast or conservative rasterization ...

Voxelization is INDEPENDENT from geometry. It only describes a way to represent represent geometry as a 3D texture data ...

It's the acceleration structure associated with the voxelization that's a big impact on performance ...

http://www.seas.upenn.edu/~pcozzi/OpenGLInsights/OpenGLInsights-SparseVoxelization.pdf

If you take a look at OpenGL Insights on chapter 22 under the results for compositing of voxel fragments, it takes about 4ms to do conservative rasterization on a GTX 480 with a voxel size of 128^3 (which is enough for GI purposes) using an RG16 format ...

4ms is hardly anything to cry over about ...
 
Last edited: