Nvidia 9xx asynchronous compute

[Kryptyx]

According to the following article Nvidia has confirmed that 9xx cards will support asynchronous compute but that it is currently disabled in the driver. Once enabled we should expect performance to increase quite a bit. Hopefully this puts to rest all the DX12 claims about them falling so far behind AMD.

 

Source:

http://www.extremetech.com/gaming/223567-amd-clobbers-nvidia-in-updated-ashes-of-the-singularity-directx-12-benchmark

[79wjd]

4 minutes ago, Kryptyx said:

Hopefully this puts to rest all the DX12 claims about them falling so far behind AMD

Can I have some of what you're smoking 

 

As long as there is a bandwagon to jump on, people will jump on one. 

[Kryptyx]

1 minute ago, djdwosk97 said:

Can I have some of what you're smoking 

 

As long as there is a bandwagon to jump on, people will jump on one. 

I guess one can hope...

[2FA]

Async compute is purely hardware based and Maxwell lacks the physical part needed to do Async compute. Driver based Async compute isn't real and is really Synchronous compute.

[79wjd]

3 minutes ago, DeadEyePsycho said:

Async compute is purely hardware based and Maxwell lacks the physical part needed to do Async compute. Driver based Async compute isn't real and is really Synchronous compute.

Unless I'm mistaken, Maxwell supports Async compute, there's just no hardware pipeline for it -- so just the pipeline needs to be implemented in drivers (which is manageable).

[2FA]

9 minutes ago, djdwosk97 said:

Unless I'm mistaken, Maxwell supports Async compute, there's just no hardware pipeline for it -- so just the pipeline needs to be implemented in drivers (which is manageable).

Maxwell has compute shaders which can do Async but they rely heavily on context switching which is delaying a thread to be used later and as a result increases latency. AMD has compute engines that are specifically designed to do async and do so much better than Nvidia. Basically async is supposed to give higher FPS with lower latency which it does on AMD. Nvidia increases FPS but does not decrease latency. 

 

Basically it comes down to designing specifically for async or just tacking it on as an after thought.

[ybriK]

Maxwell has no hardware async compute, Nvidia has to do driver updates for it on the software level which isn't as effective in terms of performance boost if it were on the hardware level.

 

Tinfoil Hat Time: Nvidia purposely got rid of async compute at the hardware level to force Maxwell users to upgrade to Pascal.

 

 

[xAcid9]

If Async Compute is disabled in driver then how did they ran Async Compute benchmark with Nvidia card? Thru voodoo magic?

[Atmos]

34 minutes ago, AlexTheRose said:

funniest shit is, you're not totally wrong.

 

Async support takes quite a bit more than a driver update, the actual core design needs to be re-worked for it to take make a real difference, even if they "implement" it through a software patch, their still lacking the built-in support you find in amd's lineup. Plainly, even when maxwell sees async support the game will still play out with amd getting a much larger boost to performance than the boost nvidia will get.

 

 

[bilalgooner]

will nvidia implement async compute on 900M series as well

[Majestic]

On 25-2-2016 at 8:43 AM, xAcid9 said:

If Async Compute is disabled in driver then how did they ran Async Compute benchmark with Nvidia card? Thru voodoo magic?

Through some sort of software emulation probably. Unsuprising the performance declined by using it, and why there was a visible "wall" that nvidia hit up to a certain performance point. As the 980TI performs the same as the GTX 980. That wouldn't make sense, in any form of shape, unless there's a common denominator.

 

And can I just add that I really don't understand the hype. I still see the 390X and 980 butting heads like they have been doing on low-drawcall / CPU-unintensive games. And may I also remind people that all 390X's are aftermarket units running overclocks, and not running "reference" speeds, since there are none. And that most Nvidia cards in the tests are running reference values. Most aftermarket models run 1329-1400 mhz instead of 1216. 

 

The only real winner is the Fury X, the rest... really uninspiring leap in performance (in relative terms, yes they jump in performance from DX11 because their DX11 drivers are terrible). And the superlatives being thrown around like AMD is making mincemeat out of Maxwell (that is apparently disabling async support) are way out of proportion.

 

I mean it's great AMD's performing atleast up to what the hardware is supposed to do now, and not being limited by CPU speeds, meaning they can atleast compete again with Maxwell and possibly Pascal. But for the love of god people, try to contain your hypetrains.

 

EDIT: and can I also add that the CPU overhead AMD has still isn't really resolved? Only postponed due to the lower overhead from the new API.

 

 

 

The 970 is still saturated much earlier.

[Humbug]

Good but this isn't this the exact same thing they said back in September?

http://www.guru3d.com/news-story/nvidia-will-fully-implement-async-compute-via-driver-support.html

 

Not saying they are lying but it's been about 6 months since AOTs alpha came out with the feature and even longer since DX12 spec came out. Their driver team better get on it... Oxide has been providing them with ashes of the singularity source code since 2014.

[Majestic]

6 minutes ago, Humbug said:

Not saying they are lying but it's been about 6 months since AOTs alpha came out with the feature and even longer since DX12 spec came out. Their driver team better get on it...

To be honest, the stakes are only high because people are making it so. It's one early-access game, should they dump a good portion of their R&D onto a game subject to change? Or wait for some stable build to release and then optimize.

 

Not rationalizing, but from a grand perspective...it seems pretty silly for one developer to force this much investment out of the hardware manuf. It's not even a big budget AAA title that is going to be bundeled or anything. And the game seems pretty mediocre past the new tech. It's not gettin raving reviews on steam either. 

 

This just pretty much seems to boil down to marketing at this point, if it hadn't been hyped noone would've batted an eye about the game. Still, now that it is, they can't slack.

[Fetzie]

17 minutes ago, Majestic said:

To be honest, the stakes are only high because people are making it so. It's one early-access game, should they dump a good portion of their R&D onto a game subject to change? Or wait for some stable build to release and then optimize.

 

Not rationalizing, but from a grand perspective...it seems pretty silly for one developer to force this much investment out of the hardware manuf. It's not even a big budget AAA title that is going to be bundeled or anything. And the game seems pretty mediocre past the new tech. It's not gettin raving reviews on steam either. 

 

This just pretty much seems to boil down to marketing at this point, if it hadn't been hyped noone would've batted an eye about the game. Still, now that it is, they can't slack.

I think at least some of the hype is because it is the first real demo of DX12 in an actual game. It isn't a particularly long stretch to assume that if nVidia doesn't have Async-Compute and is quite far behind AMD's GPUs right now, that if nVidia doesn't implement it at some point then AMD will still wipe the floor with nVidia's cards in other titles that use it.

[Majestic]

18 minutes ago, Fetzie said:

I think at least some of the hype is because it is the first real demo of DX12 in an actual game. It isn't a particularly long stretch to assume that if nVidia doesn't have Async-Compute and is quite far behind AMD's GPUs right now, that if nVidia doesn't implement it at some point then AMD will still wipe the floor with nVidia's cards in other titles that use it.

But it currently isn't "swiping the floor". That sort of superlative would be reserved if they beat them on a substantial margin, like 40-50% or higher. The 380 clearly beats the 960, no question. The 970, once on DX12, clearly loses to the 390. The 980 and 390X are pretty much tied. And the Fury X beats the 980TI. With the added caveat that current tests that are out are between factory overclocked AMD cards, versus tame/reference Nvidia cards (that are supposedly not unlocked, but i'm going to assume this is marketing BS).

 

 

By significant margins, but not by landslides to warrant such superlatives. Nothing unlike we've seen the 980TI do in all the major DX11 titles in the past two years. And i'm going to wait for A. more games with DX12, B. nvidia's claimed "support" for async, C. what vulcan is doing, which I think is more important if we want to finally get rid of being forced to windows, D. how much async is going to be actually implemented by developers not being funded by AMD (c'mon, they've been advertising if for quite some time), before i'm really picking a side on this debate. Right now, the only thing i'm trying to do is provide a dampening factor on atleast some of the hype and superlatives being hurled around. 

 

As for the current online discussions;

All i'm seeing, is people that have been in denial about objective performance now suddently don't have to be in denial anymore, since they have objectivity on their side now (actually being better), yet are being extremely smug and obnoxious about it. They've completely jumped the shark and went from playing the victim card with gameworks, to mocking nvidia customers about Asynchonous Compute. Which is basically the same thing as Gameworks represented for AMD with Tessallation. It ran poorly because AMD lacked the drawcall throughput and Tessellation efficiency, not because it was "black box".

 

But somehow, this is supposed to be good for all people (eventhough 80% of the market has nvidia), as opposed to Tessellation being evil, black box and nefarious of intent. It's incredibly hypocritical if you ask me. And the worst thing is, is that the people being incredibly smug and unapologeticly hyping this, were the same people crying wolf with Gameworks.

 

We've done a full 180 now. And noone sees the irony of it.

[Fetzie]

10 minutes ago, Majestic said:

But it currently isn't "swiping the floor". That sort of superlative would be reserved if they beat them on a substantial margin, like 40-50% or higher. The 380 clearly beats the 960, no question. The 970, once on DX12, clearly loses to the 390. The 980 and 390X are pretty much tied. And the Fury X beats the 980TI. With the added caveat that current tests that are out are between factory overclocked AMD cards, versus tame/reference Nvidia cards (that are supposedly not unlocked, but i'm going to assume this is marketing BS).

 

 

According to that graphic the nVidia cards have GPU Boost turned on (which will which will give them an overclock to something in the region of 1150ish MHz).

[Majestic]

7 minutes ago, Fetzie said:

According to that graphic the nVidia cards have GPU Boost turned on (which will which will give them an overclock to something in the region of 1150ish MHz).

Aftermarket 970's boost from 1329 to 1400mhz, depending on the vendor. Now scaling isn't 100%, so I can't add that percentage to the score. 

[Fetzie]

18 minutes ago, Majestic said:

Aftermarket 970's boost from 1329 to 1400mhz, depending on the vendor. Now scaling isn't 100%, so I can't add that percentage to the score. 

So that's a 30% overclock over "stock"?

[Majestic]

6 minutes ago, Fetzie said:

So that's a 30% overclock over "stock"?

Stock boost is 1178mhz. So ~19%-ish over stock. But 1500mhz isn't an uncommon OC for a 970, which is 27%. Now again, scaling isn't 100%, so the scores won't increase by 27%. But certainly to the degree of 10-15% or 20% for the OC, being very tame in expectations.

 

Especially the 980TI has some pretty substantial scaling, because it has a fairly low baseclock to start with, but can OC to the same hights as the other maxwell chips. 

[Nena Trinity]

This is simply Nvidia forcing the card to do something their card was not originally designed to do which is pure desperation...

[Humbug]

53 minutes ago, Majestic said:

 

. Which is basically the same thing as Gameworks represented for AMD with Tessallation. It ran poorly because AMD lacked the drawcall throughput and Tessellation efficiency, not because it was "black box".

 

But somehow, this is supposed to be good for all people (eventhough 80% of the market has nvidia), as opposed to Tessellation being evil, black box and nefarious of intent. It's incredibly hypocritical if you ask me.

 

thing is game developers were caught using over-tesselation, slowing down performance on all GPUs (including NVIDIA) for no added visual benefit. It's not just a gameworks concern it has happened to game assets which were even not part of gameworks. That is something that everybody should fight against because nobody wins from that. Even Nvidia users are getting short-changed.

 

personally I don't think that all DX12 and Vulkan games will perform like AOTS. Most games don't have so much going on and only some of them will implement asynchronous compute. I think NVIDIA will be competitive. And also asynchronous compute doesn't slow down Nvidia GPUs. It makes AMD faster. I.e. Nobody is getting slowed down by it, no losers.

[Majestic]

Just now, Humbug said:

thing is game developers were caught using over-tesselation, slowing down performance on all GPUs (including NVIDIA) for no added visual benefit. It's not just a gameworks concern it has happened to game assets which were even not part of gameworks. That is something that everybody should fight against because nobody wins from that. Even Nvidia users are getting short-changed.

 

personally I don't think that all DX12 and Vulkan games will perform like AOTS. Most games don't have so much going on and only some of them will implement asynchronous compute. I think NVIDIA will be competitive. And also asynchronous compute doesn't slow down Nvidia GPUs. It makes AMD faster.

You're assumingly talking about The Witcher 3? Which is one isolated example, to which nefarious intent of Nvidia has never been attributed by outset. And to iterate, the 64x tessellation decision fell on CD Project RED to make. I'm not apologizing, I just think it could've just as easily been a mistake. I think it's just a bit too easy to partake in such broadstroking based on one example.

 

That last part I agree with.

[Humbug]

3 minutes ago, Majestic said:

You're assumingly talking about The Witcher 3? Which is one isolated example, to which nefarious intent of Nvidia has never been attributed by outset. And to iterate, the 64x tessellation decision fell on CD Project RED to make.

 

not just witcher 3. Arkham Origins and Crysis 2 come to mind. I too blame game developers for this, they have to take responsibility for their code.

[BrightCandle]

There is a misleading discussion that keeps happening around this issue. Nvidia's cards already support Asynchronous Compute, CUDA is based on it as is the entire DX12 API and its working even in DX11. The cards are already capable of taking asynchronous work and scheduling it there is no question of that at all. The only issue is whether Graphics can run concurrently with compute, rather than being scheduled serially. The point of which is usually to reduce latency but it can also allow more compute time per frame depending on how an engine is written. So its important to understand what is and is not in dispute, its not asynchronous compute its whether compute and graphics can work concurrently given a mixed workload involving both.

 

The official documentation on this from Nvidia says it can do one graphics task in parallel with some compute tasks, a bundle of amateur testers have sort of shown it currently can't but might be able to and Nvidia is still saying it needs to work on the drivers. They might be lying about it actually working concurrently or intentionally misleading people by using the word asynchronous instead of concurrent but we don't really know. What we do know is that Nvidia doesn't think its drivers do it yet so at the very least, the question only really remains if they will ever actually release something that does produce a reasonable performance boost in this game or not. We aren't going to know until Nvidia says its done doing what its doing and we see how it actually works.

 

Its worth knowing for example that Nvidia has hardware schedulers for the jobs that are a lot more sophisiticated than AMDs, but they might also be a problem. They may very well not be designed for doing compute and graphics in parallel and hence resulting in more software from their part than they would like to do, a workaround. But AMD uses a lot less hardware and does a lot more of the scheduling work on the CPU, which is where some of the DX11 overhead is likely coming from because they do a lot more in software than Nvidia does.

 

Its not an important benchmark its not a game I care about but it should at least shake the tree enough and we should be getting to the truth soon, its just a little too early to call it right now.