No DX12 Software is Suitable for Benchmarking *spawn*

I'm guessing the reason why the GTX 970 wasn't shown in the Async Compute part on Anandtech was that it showed no gains (IE - performance didn't drop when it was disabled)?
You are correct sir!

I did quickly test GTX 970 and there are no gains. Not that you'd expect to see any given that NVIDIA has yet to enable concurrency on Maxwell 2, and likely never will. (It's too dang hard to get any perf benefits out of it with static scheduling)
 
Guru's results are currently kinda worthless because the CPU score is inflating the overall results. They should post the GPU and CPU scores separately like the other websites. At least we can see in the screenshot that the 1070 has a Graphics score of 5654 which is in line with Anandtech's and Computer Base's results.
I know, but that does not help explain then why the 480 has the same overall performance on all 3 sites :)
So it cannot be said "kinda worthless", furthermore both Guru3d and pcworld use the 5960x.
For the 480:
Guru3D, 4,286
Anandtech, 4,218
PCWorld, 4,299.

So while not ideal, the fact the 1070FE is actually at least 5% faster on Guru3d is still a valid question and point.
Knowing drivers used by all sites (at least we have this info from Guru3d) is probably as useful as knowing the GPU score on Guru3d.
If it is drivers, then this means when both AMD and Nvidia release a driver it may improve the benchmark, a consideration if using these scores as an absolute this early.
Cheers
 
Last edited:
I wonder how Guru3d ended up with at least 5% better score for their 1070, their AMD 480 seems to have near identical scores to the other sites.
Driver (it is a reference 1070)?
Guru3d only one I think to mention their drivers.
Having to use overall scores here, but still seems pretty aligned for 480 between all of them including Guru3d, only the 1070 seems to have some noteable divergence.
Possibly because Guru3D used the latest Nvidia drivers released today (WHQL Geforce 368.81). ComputerBase used the previous Geforce 368.69.
 
Scored around 500 on an intel 580.

Aside Im curious to see doom benches for apus a la 6770hq vs 7850k.

Another aside, wow is intel's driver situation not good. Only the latest beta driver supports doom and a previous beta driver to support doom vulkan...just not the latest.
 
I think the basic edition doesn't let you turn off async, still interesting to see Pascal actually showing an improvement with async. It's also interesting to see the 480 having less of an improvement in comparison to the 390/fury (11% vs 15%).
Must be lacking DX12 and low-level API support in general on the RX 480 then. *SCNR* Or maybe RX 480 is able to better utilize it's ressources in the first place. [Removed unnecessary sarcastic remark]

I did quickly test GTX 970 and there are no gains. Not that you'd expect to see any given that NVIDIA has yet to enable concurrency on Maxwell 2, and likely never will. (It's too dang hard to get any perf benefits out of it with static scheduling)
If I had to place a bet, I'd say if anything, 3DMark is a valid target from a corporate perspective for such an endeavor. Except of course when you want to highlight your own new products' superiority. :D
 
Last edited:
Guru's results are currently kinda worthless because the CPU score is inflating the overall results. They should post the GPU and CPU scores separately like the other websites. At least we can see in the screenshot that the 1070 has a Graphics score of 5654 which is in line with Anandtech's and Computer Base's results.

Mhi7m5A.jpg


Oddly enough my Fury X (stock clock) is faster than theirs (5067 vs 5342):

mbAm80m.jpg

GUru3D have use the graphic score only on the comparaison with async off on.. Dont worry its the holliday, the article will be updated continousl. ( theres even some error in his chart that need to be corrected. ) ( Fury X was missing etc )

FuryX on guru3D.. 5460 points ( a bit faster than your )..
 
Last edited:
From the (surprisingly detailed) 3dmark technical guide.

Before the main illumination passes, asynchronous compute shaders are used to cull lights, evaluate illumination from prebaked environment reflections, compute screen-space ambient occlusion, and calculate unshadowed surface illumination. These tasks are started right after G-buffer rendering has finished and are executed alongside shadow rendering.

Particles are simulated on the GPU using asynchronous compute queue. Simulation work is submitted to the asynchronous queue while G-buffer and shadow map rendering commands are submitted to the main command queue.
 
Oddly enough Time Spy uses DirectX 12 feature level 11_0. (shouldn't make any big difference in the longer run..but surprising)
 
IIRC It's the lowest common denominator that includes GCN, that still doesn't support conservative raster and raster order views.

It would have been nice to see those new two features being used in this benchmark :(
 
IIRC It's the lowest common denominator that includes GCN, that still doesn't support conservative raster and raster order views.

It would have been nice to see those new two features being used in this benchmark :(


Well kinda invalidates a DX12 benchmark if some cards can do something an others can't so probably why it was left out.
 
Well kinda invalidates a DX12 benchmark if some cards can do something an others can't so probably why it was left out.

It would be nice though. API overhead test includes Mantle which is obviously AMD only.

At least it would give us insight if there is any performance benefit for Nvidia.
 
It would be interesting to see how a ROV OIT implementation would compare to Futuremark's A-buffer approach.

Both nVidia and AMD have been really slacking on the demo front lately. The moon landing VXGI demo was pretty lame.
 
From the (surprisingly detailed) 3dmark technical guide.
Futuremark's technical guides have consistently been good, but personally I think they really hit it out of the park with Time Spy. Virtually everything I wanted to know was in the guide, and a lot more stuff I wouldn't have thought to ask about or wouldn't have expected is in there as well. I wish we could get this much detailed information out of other benchmarks and games as well.
 
Last edited:
Futuremark's technical guides have consistently been good, but personally I think they really hit it out of the park with Time Spy. Virtually everything I wanted to know was in the guide, and a lot more stuff I wouldn't have thought to ask about or wouldn't have expected. I wish we could get this much detailed information out of other benchmarks and games as well.

That's the reason I paid for my 3D Mark yesterday. Support job well done :)
 
Oddly enough Time Spy uses DirectX 12 feature level 11_0. (shouldn't make any big difference in the longer run..but surprising)

I imagine that's there to support anything older than Maxwell (or was it Keplar) for Nvidia, and anything older than GCN based cards (I believe) for AMD.
IIRC It's the lowest common denominator that includes GCN, that still doesn't support conservative raster and raster order views.

It would have been nice to see those new two features being used in this benchmark :(

Interesting that you left out Fermi, Kepler and Maxwell 1 which also don't support conservative rasterization.

If Fermi, Kepler and Maxwell 1 are able to run this, that would likely also be part of why feature 11_0 was used.

Regards,
SB
 
Back
Top