PCPer Tests 3DMark Timespy: Pascal Does have Async Compute

[xAcid9]

Just tried the test with my 290x at reference speed. 

 

3 930 GPU score with AsyncCompute

3 397 GPU score w/o AsyncCompute.

[Matu20]

7 minutes ago, don_svetlio said:

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

You are quick to throw irrelevant comparisons at me. All I am saying is that 390X is a complete rebrand of 290X with VRAM upgrade + OC and it shouldn't be surprising that the 290X beats a 480 when fully utilized. 

[don_svetlio]

1 minute ago, Matu20 said:

You are quick to throw irrelevant comparisons at me. All I am saying is that 390X is a complete rebrand of 290X with VRAM upgrade + OC and it shouldn't be surprising that the 290X beats a 480 when fully utilized. 

Not exactly - have you wondered why almost every 390(X) achieves 1150-1250 whereas almost no 290(X) could even do 1200? The whole power delivery has been redesigned reducing the overall board draw and improving stability.

[AresKrieger]

13 minutes ago, ivan134 said:

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.

What I meant is that in many cases it isn't a practical solution, you'll see a greater performance gain with many ai with there own set of calculations than one ai with more complexity, especially if they are using it in conjunction with shader calculations (since there are more moving objects to calculate for)

[Matu20]

7 minutes ago, don_svetlio said:

Not exactly - have you wondered why almost every 390(X) achieves 1150-1250 whereas almost no 290(X) could even do 1200? The whole power delivery has been redesigned reducing the overall board draw and improving stability.

I didn't really elaborate enough, but with the VRAM upgrade and clock improvement, redesigning the delivery had to be done (how else could they release the same card, they had to do something to improve it). Honestly this was a cop out and shouldn't have been marketed as "new" cards.

[don_svetlio]

1 minute ago, Matu20 said:

I didn't really elaborate enough, but with the VRAM upgrade and clock improvement, redesigning the delivery had to be done (how else could they release the same card, they had to do something to improve it). Honestly this was a cop out and shouldn't have been marketed as "new" cards.

Also - the price was vastly reduced. Remember the original MSRPs? It's a refresh. Just like the 770 was to the 680.

[Matu20]

1 minute ago, don_svetlio said:

Also - the price was vastly reduced. Remember the original MSRPs? It's a refresh. Just like the 770 was to the 680.

And again you are bringing Nvidia into the conversation. Comparing 300 series to the original MSRP of 200 series launch price is completely irrelevant. The only thing that mattered when 300 series released was the current price of the 200 series and they were way cheaper. 

[don_svetlio]

1 minute ago, Matu20 said:

And again you are bringing Nvidia into the conversation. Comparing 300 series to the original MSRP of 200 series launch price is completely irrelevant. The only thing that mattered when 300 series released was the current price of the 200 series and they were way cheaper. 

So when you make a comparison it's fine but when I do it's not? The cards were lowered a tier down - the same way it has been since 2006. I don't see the problem with it.

[gilang01]

I'm curious on how the 1060 will manage DX12, since the 10 series does "have" async compute. 

[Matu20]

3 minutes ago, don_svetlio said:

So when you make a comparison it's fine but when I do it's not? The cards were lowered a tier down - the same way it has been since 2006. I don't see the problem with it.

I haven't compared any other cards that I didn't bring up in the first post (290x and 390x). And if you don't see problems with rebrands and get the concept, then why would you be surprised that they perform similarly?

[don_svetlio]

3 minutes ago, gilang01 said:

I'm curious on how the 1060 will manage DX12, since the 10 series does "have" async compute. 

Depends on drivers - Nvidia have a lot of overhead in DX12

[samcool55]

Tbh i expected a bigger boost with nvidia cards...

If you look how big the improvement is for AMD cards, even older ones, the improvement on Nvidia cards isn't great...

It's good, don't get me wrong, but i expected better.

 

The boost on the GTX1080 is smaller than on the RX 480.

 

Does anyone have a source for benchmarks of older AMD cards? I've found some numbers from random posts but nothing that's perfectly comparable.

[ivan134]

6 minutes ago, Matu20 said:

And again you are bringing Nvidia into the conversation. Comparing 300 series to the original MSRP of 200 series launch price is completely irrelevant. The only thing that mattered when 300 series released was the current price of the 200 series and they were way cheaper. 

Uhhh, no it doesn't. More people need to pick up an economics text book. NO CORPORATION ever prices a product based on what the sale price of their older product was. The factors that affect those sale prices do not affect their own cost of production. A 290 being cheaper didn't mean it was cheaper for AMD to make a 390. You're asking that they sell at a loss because their older product is on sale. That's not how it works.

[Humbug]

Also id software said they are currently working with Nvidia to enable async compute support in Doom for Nvidia GPUs. They were probably talking about Pascal. Expect to see a driver & game patch which enables that. In that case it's vulkan not dx12 but same concept.

[Matu20]

1 minute ago, samcool55 said:

Tbh i expected a bigger boost with nvidia cards...

If you look how big the improvement is for AMD cards, even older ones, the improvement on Nvidia cards isn't great...

It's good, don't get me wrong, but i expected better.

 

The boost on the GTX1080 is smaller than on the RX 480.

 

Does anyone have a source for benchmarks of older AMD cards? I've found some numbers from random posts but nothing that's perfectly comparable.

That is because Nvidia stopped implementing hardware level Async since Fermi, whereas AMD still does it hardware level (which is always faster than software).

[3DOSH]

10 minutes ago, Matu20 said:

And again you are bringing Nvidia into the conversation. Comparing 300 series to the original MSRP of 200 series launch price is completely irrelevant. The only thing that mattered when 300 series released was the current price of the 200 series and they were way cheaper. 

Give AMD some credit, the 290 and 390 still kick ass where as Kepler is no where to be found. Even with refresh it still goes toe to toe with the 970 and 980 that shows you how much ahead AMD where thinking. 

[samcool55]

1 minute ago, don_svetlio said:

Depends on drivers - Nvidia have a lot of overhead in DX12

Umm doesn't it depend on the fact it uses async or not?

I always thought the point of DX12 was to be able to program at a lower level (async compute) so driver overhead would be eliminated.

 

I'm also suprised about the fact you say Nvidia has a lot of overhead in DX12, i always thought they did a good job and it was AMD that had issues with overhead.

[ivan134]

3 minutes ago, Humbug said:

Also id software said they are currently working with Nvidia to enable async compute support in Doom for Nvidia GPUs. They were probably talking about Pascal. Expect to see a driver & game patch which enables that. In that case it's vulkan not dx12 but same concept.

We already have one person on this forum accusing iD Tech of AMD bias. People still call Oxide (devs for AotS) biased even though they spent more time working with Nvidia than AMD to make sure their performance wasn't absolute garbage and even made a special code path just for them.

[don_svetlio]

2 minutes ago, samcool55 said:

Umm doesn't it depend on the fact it uses async or not?

I always thought the point of DX12 was to be able to program at a lower level (async compute) so driver overhead would be eliminated.

 

I'm also suprised about the fact you say Nvidia has a lot of overhead in DX12, i always thought they did a good job and it was AMD that had issues with overhead.

With DX11 - DX12 requires a lot of driver work from Nvidia since their cards lack the hardware support for A-sync and thus a small driver error will be disastrous

[Curufinwe_wins]

57 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?

Knowing this money grab, you probably need to buy the extra timespy features to be able to test multigpu's.

[samcool55]

3 minutes ago, Matu20 said:

That is because Nvidia stopped implementing hardware level Async since Fermi, whereas AMD still does it hardware level (which is always faster than software).

Why did they even bother with that so long ago? Afaik DX11 didn't have any async support and it's only getting relevant now.

I'm completely lost. I always thought AMD started async stuff with GCN and the 7XXX series and it didn't exist before that O_o

 

[Matu20]

3 minutes ago, ivan134 said:

Uhhh, no it doesn't. More people need to pick up an economics text book. NO CORPORATION ever prices a product based on what the sale price of their older product was. The factors that affect those sale prices do not affect their own cost of production. A 290 being cheaper didn't mean it was cheaper for AMD to make a 390. You're asking that they sell at a loss because their older product is on sale. That's not how it works.

I was viewing it from the consumer point of view. Of course AMD is going to price it higher than the last gen, but since it wasn't a huge upgrade, but rather a sidegrade, it made sense to opt for the cheaper last gen product that performed very similarly. 

 

And the 390 was definitely cheaper to produce at launch than the 290 was when it released. As production and lithography ages, it matures and becomes more efficient, thus cheaper and more reliable.

[samcool55]

3 minutes ago, don_svetlio said:

With DX11 - DX12 requires a lot of driver work from Nvidia since their cards lack the hardware support for A-sync and thus a small driver error will be disastrous

Well yea if the game uses async compute and their cards don't have native support for that it makes sense the driver overhead increases because it has to compensate for it.

 

But when you run a DX12 game without async compute, than there isn't much driver overhead because there isn't any async compute it has to compensate for, right?

Just talking about Nvidia here.

[ivan134]

3 minutes ago, Matu20 said:

I was viewing it from the consumer point of view. Of course AMD is going to price it higher than the last gen, but since it wasn't a huge upgrade, but rather a sidegrade, it made sense to opt for the cheaper last gen product that performed very similarly. 

 

And the 390 was definitely cheaper to produce at launch than the 290 was when it released. As production and lithography ages, it matures and becomes more efficient, thus cheaper and more reliable.

It wasn't price higher than last gen. The sale prices you saw for the 290 at the end of its life were due to market factors, not prices set by AMD. MSRP for R9 290 was $399. MSRP for The R9 390 was $329. Where is the price increase you're talking about?

[don_svetlio]

1 minute ago, samcool55 said:

Well yea if the game uses async compute and their cards don't have native support for that it makes sense the driver overhead increases because it has to compensate for it.

 

But when you run a DX12 game without async compute, than there isn't much driver overhead because there isn't any async compute it has to compensate for, right?

Just talking about Nvidia here.

A-sync is the thing that increases performance though - all the other DX12 effects have no effect or decrease it