PCPer Tests 3DMark Timespy: Pascal Does have Async Compute

[Briggsy]

http://www.pcper.com/reviews/Graphics-Cards/3DMark-Time-Spy-Looking-DX12-Asynchronous-Compute-Performance

 

PCPer has published an article on the new 3DMark DX12 Time Spy Benchmark, and tested a number of graphic cards from both AMD and Nvidia. Based on the results, it confirms that Pascal does have support for Async Compute, unlike its predecessor Maxwell which clearly does not show any performance gains (ahem** still waiting for async compute driver).

 

While it is apparent that the FuryX is getting as much as twice the relative performance gain over Pascal with Async Compute enabled, This should at least confirm that Nvidia have (in part) successfully circumvented the need for power hungry componentry, while still achieving noteworthy gains in async compute performance, with the 1080 showing gains similar to the R9 Nano, and the 1070 just trailing the 480.

 

How this translates to real world performance mostly remains to be seen, as it depends on the actual adoption of the new API's, and competent implementation from developers. It should also be noted that large % gains with async compute should not necessarily be applauded, or compared between cards. As we can see just below, the 480 barely makes up any ground against the 1070 in this test. I personally feel perf/watt still remains the best metric in determining what approach is the best, and not just for environmental reasons.

 

Actual performance increase per Card (highest to lowest):

FuryX: 652 points 

Nano: 496 points 

GTX 1080: 471 points

RX 480: 335 points

GTX 1070: 305 points

1080 SLI: 164 points

GTX 970: ?

GTX 980: ?

 

[Curufinwe_wins]

Maxwell technically does have support for async it just does it so poorly that there are almost no benefits and often reduces performance due to increased overhead. 

 

Thus the negligible performance change. 

 

I don't know if they tested a Hawaii gpu, but that is what I'd like to see in this. Remove some of those bottlenecks. 

[laminutederire]

That's because they always had it in software implementation while AMD is the only one having an hardware implementation . It's at least what I understood 

[Megahurt]

9 minutes ago, Curufinwe_wins said:

Maxwell technically does have support for async it just does it so poorly that there are almost no benefits and sometimes reduces performance due to increased overhead. 

 

Thus the negligible performance change. 

 

I don't know if they tested a Hawaii gpu, but that is what I'd like to see in this. Remove some of those bottlenecks. 

While I do believe maxwell might have some sort of async compute, we've never seen it, cause it has never been enabled in drivers, nvidia said so. But if it does have it, it's probably broken to the point where it causes problems or reduces performance.

[AresKrieger]

15 minutes ago, Curufinwe_wins said:

Maxwell technically does have support for async it just does it so poorly that there are almost no benefits and sometimes reduces performance due to increased overhead. 

 

Thus the negligible performance change. 

 

I don't know if they tested a Hawaii gpu, but that is what I'd like to see in this. Remove some of those bottlenecks. 

As far as I know the 980ti is the only maxwell gpu with async support of any kind, and it is limited to a queue of 16 instructions, which would be fine if dx12 and vulcan weren't based on mantel which expects more due to amd's architecture, but since it is nvidia has just disabled it up to this point (and likely forever)

[Morgan MLGman]

18 minutes ago, Curufinwe_wins said:

Maxwell technically does have support for async it just does it so poorly that there are almost no benefits and sometimes reduces performance due to increased overhead. 

 

Thus the negligible performance change. 

 

I don't know if they tested a Hawaii gpu, but that is what I'd like to see in this. Remove some of those bottlenecks. 

A 290X would be the perfect candidate to test

[Curufinwe_wins]

7 minutes ago, Megahurt said:

While I do believe maxwell might have some sort of async compute, we've never seen it, cause it has never been enabled in drivers, nvidia said so. But if it does have it, it's probably broken to the point where it causes problems or reduces performance.

 

3 minutes ago, AresKrieger said:

As far as I know the 980ti is the only maxwell gpu with async support of any kind, and it is limited to a queue of 16 instructions, which would be fine if dx12 and vulcan weren't based on mantel which expects more due to amd's architecture, but since it is nvidia has just disabled it up to this point (and likely forever)

It was initially enabled during the AoTS beta where it was tanking perf by a good 5-10%. Like I said, it supports it (it doesn't have the proper native hardware to benefit), but that doesn't mean it's a good feature to enable most of the time with Maxwell.

[AresKrieger]

For reference async compute is only useful in scenarios with numerous non-parallel calculations happening frequently or continously (such as ai in a rts game), in other scenarios it would be easier and more beneficial to simply optimize the game for cpu use than to implement the feature.

[Prysin]

3 minutes ago, Curufinwe_wins said:

 

It was initially enabled during the AoTS beta where it was tanking perf by a good 5-10%. Like I said, it supports it (it doesn't have the proper native hardware to benefit), but that doesn't mean it's a good feature to enable most of the time with Maxwell.

Time Spy does NOT like 295x2

crashes my PC every time i try to run it...

[ivan134]

25 minutes ago, Curufinwe_wins said:

Maxwell technically does have support for async it just does it so poorly that there are almost no benefits and often reduces performance due to increased overhead. 

 

Thus the negligible performance change. 

 

I don't know if they tested a Hawaii gpu, but that is what I'd like to see in this. Remove some of those bottlenecks. 

I know PCWorld is the best with their testing methodologies, but it's all I can find.

Spoiler

 

[Curufinwe_wins]

15 minutes ago, Prysin said:

Time Spy does NOT like 295x2

crashes my PC every time i try to run it...

It doesn't like my 980ti SLI either...

 

I run the demo flawlessly, but then it hangs during the first graphics test... It doesn't crash the nvidia driver but it does bug my cards to pinned at top speed (i had to run another 3d application to get them to downclock.)

 

Seems like this program is still flaky. Also really pissed off that I own the advanced edition and they don't include the ability to turn of the f'n demo without an additional purchase.

[Prysin]

1 minute ago, Curufinwe_wins said:

It doesn't like my 980ti SLI either...

 

I run the demo flawlessly, but then it hangs during the first graphics test... It doesn't crash the nvidia driver but it does bug my cards to pinned at top speed (i had to run another 3d application to get them to downclock.)

 

Seems like this program is still flaky. Also really pissed off that I own the advanced edition and they don't include the ability to turn of the f'n demo without an additional purchase.

yeah, Graphics test 1 hangs, then i have to hard reset the machine to get out of it.

 

[Curufinwe_wins]

14 minutes ago, ivan134 said:

I know PCWorld is the best with their testing methodologies, but it's all I can find.

  Hide contents

 

 

ok so looks like the hawaii cards do gain the most relatively speaking (even if both they and the fury cards gain about 600 points.)

[VagabondWraith]

Good to see Asynchronous Compute have a positive effect on nVidia Pascal GPU's. Free performance. Will be interesting to see if Volta has fully implemented ACE's.

[Curufinwe_wins]

2 minutes ago, Prysin said:

yeah, Graphics test 1 hangs, then i have to hard reset the machine to get out of it.

 

Sigh, I really wanted to see the difference between OC'd 980ti sli and stock 1080 sli... Looks like I'll have to wait a while.

[ivan134]

21 minutes ago, AresKrieger said:

For reference async compute is only useful in scenarios with numerous non-parallel calculations happening frequently or continously (such as ai in a rts game), in other scenarios it would be easier and more beneficial to simply optimize the game for cpu use than to implement the feature.

ALL games could benefit from it. With all the other advancements we've had in games, AI is the most lacking, and sometimes it seems like they're getting even dumber. Most games just have AI that does very simple and scripted things. Who wouldn't want an unpredictable AI?

[The Benjamins]

25 minutes ago, Prysin said:

Time Spy does NOT like 295x2

crashes my PC every time i try to run it...

 

9 minutes ago, Curufinwe_wins said:

It doesn't like my 980ti SLI either...

 

I run the demo flawlessly, but then it hangs during the first graphics test... It doesn't crash the nvidia driver but it does bug my cards to pinned at top speed (i had to run another 3d application to get them to downclock.)

 

Seems like this program is still flaky. Also really pissed off that I own the advanced edition and they don't include the ability to turn of the f'n demo without an additional purchase.

Try turning off SLI/Crossfire so it can run in Multi GPU mode. as far as I know DX12 does not support SLI or Crossfire.

 

If Multi GPU works can it run a 980ti and Fury?

[Coaxialgamer]

7 minutes ago, The Benjamins said:

 

Try turning off SLI/Crossfire so it can run in Multi GPU mode. as far as I know DX12 does not support SLI or Crossfire.

 

If Multi GPU works can it run a 980ti and Fury?

disabling SLI only runs one card

[The Benjamins]

4 minutes ago, Coaxialgamer said:

disabling SLI only runs one card

so It doesn't support Multi GPU?

[Matu20]

17 minutes ago, huilun02 said:

Guru3D did a test

 

Shows 290X outperforming the RX 480 and matching the GTX 980

390X finally makes a clear distinction

It isn't surprising since 390x is just a rebrand and the 480 rivals a 390 non x.

[Coaxialgamer]

7 minutes ago, The Benjamins said:

so It doesn't support Multi GPU?

doesn't appear to , at least not on my system .

[AresKrieger]

15 minutes ago, ivan134 said:

ALL games could benefit from it. With all the other advancements we've had in games, AI is the most lacking, and sometimes it seems like they're getting even dumber. Most games just have AI that does very simple and scripted things. Who wouldn't want an unpredictable AI?

The quality isn't the issue being addressed by async it's the quantity more than anything (hordes of independent moving enemies), but I agree adding more complex ai to games in general would be nice but that is for the most part a separate issue (it does take less complex AIs to make it beneficial than dopey ones I'll admit but async isn't an all purpose tool even for ai)

[don_svetlio]

3 minutes ago, Matu20 said:

It isn't surprising since 390x is just a rebrand and the 480 rivals a 390 non x.

Well, 390X did get quite a few improvements - and we can see above that the 480 is on par with a 980 with A-sync. And we also see the Fury X beating the 980 Ti

[The Benjamins]

2 minutes ago, Coaxialgamer said:

doesn't appear to , at least not on my system .

that sucks. It be nice to test Multi GPU vs SLI/Crossfire and a mix of cards

[ivan134]

2 minutes ago, AresKrieger said:

The quality isn't the issue being addressed by async it's the quantity more than anything (hordes of independent moving enemies), but I agree adding more complex ai to games in general would be nice but that is for the most part a separate issue (it does take less complex AIs to make it beneficial than dopey ones I'll admit but async isn't an all purpose tool even for ai)

Yes, it is. Asynchronous compute enables a GPU to do compute and graphical workloads at the same time. It's not just about putting more things on the screen. AI = compute workload.