Anandtech Review Summerize Up Pascal's "Async Compute" Truth

[Deleted member 50521]

Reference from Original Review article:
http://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/9

Regarding Maxwell2 Gen cards(980/970)

Meanwhile not shown in these simple graphical examples is that for async’s concurrent execution abilities to be beneficial at all, there needs to be idle time bubbles to begin with. Throwing compute into the mix doesn’t accomplish anything if the graphics queue can sufficiently saturate the entire GPU. As a result, making async concurrency work on Maxwell 2 is a tall order at best, as you first needed execution bubbles to fill, and even then you’d need to almost perfectly determine your partitions ahead of time.

Regarding Pascal's improvement in terms of Async Compute

Right now I think it’s going to prove significant that while NVIDIA introduced dynamic scheduling in Pascal, they also didn’t make the architecture significantly wider than Maxwell 2. As we discussed earlier in how Pascal has been optimized, it’s a slightly wider but mostly higher clocked successor to Maxwell 2. As a result there’s not too much additional parallelism needed to fill out GP104; relative to GM204, you only need 25% more threads, a relatively small jump for a generation. This means that while NVIDIA has made Pascal far more accommodating to asynchronous concurrent executeion, there’s still no guarantee that any specific game will find bubbles to fill. Thus far there’s little evidence to indicate that NVIDIA’s been struggling to fill out their GPUs with Maxwell 2, and with Pascal only being a bit wider, it may not behave much differently in that regard.

So yes, Nvidia did try to improve the Async compute from Maxwell2's non-existent state to a better condition. However under careful anaylsis it is predicted such a small change won't provide much help during DX12/Vulkan age.

Regarding AMD:

Meanwhile, because this is a question that I’m frequently asked, I will make a very high level comparison to AMD. Ever since the transition to unified shader architectures, AMD has always favored higher ALU counts; Fiji had more ALUs than GM200, mainstream Polaris 10 has nearly as many ALUs as high-end GP104, etc. All other things held equal, this means there are more chances for execution bubbles in AMD’s architectures, and consequently more opportunities to exploit concurrency via async compute. We’re still very early into the Pascal era – the first game supporting async on Pascal, Rise of the Tomb Raider, was just patched in last week – but on the whole I don’t expect NVIDIA to benefit from async by as much as we’ve seen AMD benefit. At least not with well-written code.

In the end it boils down to how soon Nvidia have chosen to make the shift from old ways to unified ALUs.

GPUs continually sit on the fence between being an ultra-fast staticly scheduled array of ALUs and an ultra-flexible somewhat smaller array of ALUs, and GPU vendors get to sit in the middle trying to figure out which side to lean towards in order to deliver the best performance for workloads that are 2-5 years down the line. It is, if you’ll pardon the pun, a careful balancing act for everyone involved.

Then what is pre-emption?

In a nutshell, it’s the ability to interrupt an active task (context switch) on a processor and replace it with another task, with the further ability to later resume where you left off. Historically this is a concept that’s more important for CPUs than GPUs, especially back in the age of single core CPUs, as preemption was part of how single core CPUs managed to multitask in a responsive manner. GPUs, for their part, have supported context switching and basic preemption as well for quite some time, however until the last few years it has not been a priority, as GPUs are meant to maximize throughput in part by rarely switching tasks.

Nvidia tried to cater to current gen and older games by implementing more traditional(that is suitable for DX11) way of GPU design to ensure max performance gain with minimum cost of R&D as well as production. At the same time Nvidia introduces some new workaround for DX12/Vulkan titles that are compute heavy. In this way you get happy customers and great return on revenue for the company. People will buy new cards down the road for DX12/Vuklan titles any way, If I were to work for Nvidia this will be a safe and sound strategy for GPU development for sure.

The generation that comes after Pascal will probably see drastic design change comparing to Maswell-Maxwell2-Pascal. Nvidia will shift heavily towards hardware level async compute and comes out with a super robust design. And maybe by that time the entire Tessellation situation will repeat itself again.

AMD on the other hands have been shifting towards async compute for a long time. I strongly doubt this is related to console development. After all what SONY or MS want is a hardware they can milk on for a long time without worrying the graphic/Visual lacking behind PC too much. DX12/Vulkan helps unlock max potential of hardwares which will great for console development.

My verdict:
Nvidia focus more on PC Consumer experience, which always aims for best performance during the GPU's usefulness life span.

AMD focus more on Console. PC GPU design may just be tagging along. Sometime their assumed "futuristic design" may help in new games. However as soon as Nvidia as shifts towards the direction AMD's design becomes obsolete.

 

[erocker]

I'll be "worried" when there's DX12 games out that I actually want to play. Even then it really makes no difference. If Nvidia needs to change something, the have the resources to do it. Frankly, they probably would already if they had better competition.

The only game that I would be concerned about if I was actually concerned is Deus Ex: Mankind Divided. First DX12 game I want to play. Even then, AMD won't have their higher end card(s) launched and 10XX series will most likely do fine. The thing that matters is the end result.

 

[Deleted member 50521]

I'll be "worried" when there's DX12 games out that I actually want to play. Even then it really makes no difference. If Nvidia needs to change something, the have the resources to do it. Frankly, they probably would already if they had better competition.

The only game that I would be concerned about if I was actually concerned is Deus Ex: Mankind Divided. First DX12 game I want to play. Even then, AMD won't have their higher end card(s) launched and 10XX series will most likely do fine. The thing that matters is the end result.

If you have read through my thread you would see that I totally agree with this. Nvidia can make a great GPU design. Problem is they don't have to for:

1. The 1st gen of these cards won't even be relevant long enough to help actual Async Compute
2. Not good for return on asset.
3. Making a GPU with too much Async Compute emphasis will hurt performance of games just published or about to come out.

Nvidia made the best decision--->gave consumer max current gen performance with a little improvement for the card to be relevent during the initial wave of DX12/Vulkan titles. After all very few people use their GPU for 10 years.

 

[erocker]

Yes, I read through your thread, then commented my opinions on the article (not your post) which roughly share your own opinions.

 

[Deleted member 50521]

And this is NOT bashing on AMD for console first. A small company needs to survive after all.

 

[GC_PaNzerFIN]

Every time some DirectX feature has been touted as groundbreaking and reason to pick the worse performing card because its more "future proof, designed for blabla next gen", the card has been obsolete in terms of performance long before that. I am buying what is best now, been burned so many times by sugary sweet promises.

 

[ViperXTR]

So just like others already mentioned, Maxwell 2 and Pascal in that regard are effectively near maxing their capabilities and that adding async to the mix will make little to no (and even negatively impact) difference and the results are seen on actual tests. GCN on the other hand are finally spreading its wings and finally making full use of their idle cores, and perhaps memory bandwidth as well? Pascal would need other "marketing" performance or graphical enhancements to showoff in order benefit them aside from async compute (GCN has their async compute engines and shader intrinsic derived from consoles)

 

[FR@NK]

I gotta say that this new Vulkan update for doom has made it seem like I upgraded my GPU.

 

[RejZoR]

I'll be "worried" when there's DX12 games out that I actually want to play. Even then it really makes no difference. If Nvidia needs to change something, the have the resources to do it. Frankly, they probably would already if they had better competition.

The only game that I would be concerned about if I was actually concerned is Deus Ex: Mankind Divided. First DX12 game I want to play. Even then, AMD won't have their higher end card(s) launched and 10XX series will most likely do fine. The thing that matters is the end result.

Well, there are quite a lot of people with R9 Fury and Fury X. They'll benefit a lot from all this. And to some extent even R9 390X users. DX12 games are on the rise and while NVIDIA might provide a functional async in the future, AMD already has that since HD7000 series in a limited capacity.

Also, GTX 900 series do have async compute capability, but with very short queues. If game is programmed to utilize only that queue size and no more it can be beneficial. But as soon as you go over it (which is, lets be frank, very quickly), it'll just tank in performance or shall I say the benefits disappear entirely. Where AMD cards can process very long queues in parallel and benefit tons from them in basically any kind of scenarios. That's why Radeons universaly gain performance in DX12 and Vulkan titles.

 

[Assimilator]

So yes, Nvidia did try to improve the Async compute from Maxwell2's non-existent state to a better condition.

Please, show me in the article you're referencing where it's stated that Maxwell's async compute is "non-existent". Or are you outright lying, again?

 

[RejZoR]

Basically it is. Like I said above, it has async capability for very small queues. As soon as it becomes larger than hardware can do it, it just drops in performance where Radeons churn more and more performance as queues become bigger.

 

[the54thvoid]

I can sense this free falling but kudos to @xkm1948 for impartiality.

What is important to take from this thread is what many (mostly Nvidia owners) have been saying for a while. Nvidia doesn't need to deal with Async compute because it's design choice is fast, efficient and profit maximising. And it works.

Imagine how many cards would sell if the GP104 was at the £450 mark? Its almost a Godsend to AMD that Nvidia priced so high.

I thought about a second build with a Fury X but it's still over £500 (though there was an offer recently).

Like @erocker says, Deus Ex is on my radar as hit game for 2016. If AMD choke it with Async it'll work out well for them. Even so, I'm sure I'll be happy with my 980ti's performance.

 

[laszlo]

This "small" company is going to launch 8C/16T Zen processor 95W TDP !

what had to do this with async? if i follow your line i can say they had great launches.....bulldozer...

back to topic; as i see async is quite irrelevant now (and next year maybe); all gains are good but in the end depend on developers if they use the code or not...

 

[little cat]

what had to do this with async? if i follow your line i can say they had great launches.....bulldozer...

back to topic; as i see async is quite irrelevant now (and next year maybe); all gains are good but in the end depend on developers if they use the code or not...

it refers to the word "small"
Also RX 480 is a very good card for $200

 

[RejZoR]

RX 480 gaining what, 30% in Doom (Vulkan). I wouldn't call that "small".

 

[the54thvoid]

RX 480 gaining what, 30% in Doom (Vulkan). I wouldn't call that "small".

See, this is the free fall. The OP is nothing to do with that. It's about why Nvidia aren't doing Async compute right now and why AMD are.
Also, missing the point, one mainstream game properly supports Vulkan. That is NOT a reason to schedule a chip design.
It's not even a DX12 environment yet so DX11 is still absolutely viable, ergo, Nvidia don't have to do anything.

Maybe in 2 years it will matter but we'll have Volta and maybe even Navi?

 

[R-T-B]

Please, show me in the article you're referencing where it's stated that Maxwell's async compute is "non-existent". Or are you outright lying, again?

It's not nonexistant, but it might as well be. It doesn't provide any tangible benefits because the architecture's static partitioning is that good at utilizing what it has.

it refers to the word "small"
Also RX 480 is a very good card for $200

It doesn't matter if they make good stuff or not. They are still, economically speaking, way smaller than Intel or NVIDIA.

Maybe in 2 years it will matter but we'll have Volta and maybe even Navi?

Aye. In 2 years I'll be looking at my next big upgrade for certain, if not way sooner.

 

[little cat]

It's not nonexistant, but it might as well be. It doesn't provide any tangible benefits because the architecture's static partitioning is that good at utilizing what it has.



It doesn't matter if they make good stuff or not. They are still, economically speaking, way smaller than Intel or NVIDIA.



Aye. In 2 years I'll be looking at my next big upgrade for certain, if not way sooner.

$4.30B market value is not so small

 

[GC_PaNzerFIN]

$4.30B market value is not so small

AMD has been reducing and out sourcing their employees for a long time now since the golden days have passed.
Now they are doing all the CPU & GPU stuff and R&D with same amount of people as NVIDIA is doing just GPUs.
AMD debt is now over half market value, stock value can't go much lower. In contrast, NVIDIA is pretty much debt free now.

I might not call it small, but shrinking and smaller than main competition.

 

[R-T-B]

I might not call it small, but shrinking and smaller than main competition.

That's what I meant. "Small" in the Silicon Valley sense of the word.

 

[RejZoR]

See, this is the free fall. The OP is nothing to do with that. It's about why Nvidia aren't doing Async compute right now and why AMD are.
Also, missing the point, one mainstream game properly supports Vulkan. That is NOT a reason to schedule a chip design.
It's not even a DX12 environment yet so DX11 is still absolutely viable, ergo, Nvidia don't have to do anything.

Maybe in 2 years it will matter but we'll have Volta and maybe even Navi?

Vulkan used to be OpenGL. Do you know what all runs on OpenGL as well? Prety much everything in the "serious" industry, medicine, engineering. If they want to support that, they HAVE TO support Vulkan. It's as simple as that. Sure, you still have OpenGL, but when you run business, it's not just % of framerate anymore... And AMD designed async tech years ago. Meaning most of their userbase already supports it, even if you have an "aging" R9 290X... It's a value and a reason why user of such graphic card will consider AMD instead of NVIDIA in the future. You can't just look at "now", you have to look long term and how business will give you income across several years, not just this moment.

 

[yogurt_21]

Back in the Day then Ati chose not to go all in on SM3.0 on the X800 series cards. They felt there weren't enough games out to warrant the move. Nvidia Chose to go all in on it and their 6000 series line supported it. (at least 6600 on up)
http://www.theinquirer.net/inquirer/news/1047414/nvidia-games-shader-model

some reviewers went all for green teams choice and slanted their reviews against the X800 and X850 series due to this. Others just ran the benches with the games that were available at the time and mentioned the difference in the conclusion.

How much did that difference matter now, looking back? Well the 6800's launched in april 2004 and the biggest difference I remember came in the ability to run the full suite of tests in 3dmark06 which came out in january 2006.

By that time both the X1800 series and X1900 series were out, both of which supported sm3.0
Also the 7800 series had been out for a bit with the limited edition 512MB variant which would become the 7900 GTX having been released to counter the performance of the amd offerings.

every other of the 13 games mentioned in the article above, played just fine on the X800 series cards. Sure you were missing a feature, but unless you were actively comparing, you wouldn't notice (think xbox one vs ps4).

while the async issue isn't about a graphics feature so much as performance the same applies here. Waiting until a long list of games supports the feature didn't hurt then Ati and it won't hurt nvidia now.

or are you all really just going to play 2 games, ashes and doom, until more titles come out?

 

[little cat]

Back in the Day then Ati chose not to go all in on SM3.0 on the X800 series cards. They felt there weren't enough games out to warrant the move. Nvidia Chose to go all in on it and their 6000 series line supported it. (at least 6600 on up)
http://www.theinquirer.net/inquirer/news/1047414/nvidia-games-shader-model

some reviewers went all for green teams choice and slanted their reviews against the X800 and X850 series due to this. Others just ran the benches with the games that were available at the time and mentioned the difference in the conclusion.

How much did that difference matter now, looking back? Well the 6800's launched in april 2004 and the biggest difference I remember came in the ability to run the full suite of tests in 3dmark06 which came out in january 2006.

By that time both the X1800 series and X1900 series were out, both of which supported sm3.0
Also the 7800 series had been out for a bit with the limited edition 512MB variant which would become the 7900 GTX having been released to counter the performance of the amd offerings.

every other of the 13 games mentioned in the article above, played just fine on the X800 series cards. Sure you were missing a feature, but unless you were actively comparing, you wouldn't notice (think xbox one vs ps4).

while the async issue isn't about a graphics feature so much as performance the same applies here. Waiting until a long list of games supports the feature didn't hurt then Ati and it won't hurt nvidia now.

or are you all really just going to play 2 games, ashes and doom, until more titles come out?

not "buzzed" now

not "buzzed" now

That thread is going to be a hit

 

[FordGT90Concept]

Nvidia can make a great GPU design.

I think you're forgetting how much NVIDIA struggled with DX10/11 in Kepler and sons. NVIDIA has to play catch up to AMD and AMD already has a mature architecture for Vulkan and DX12. They could design a "great GPU" but I'm skeptical seeing how long it took them to do DX11 (Maxwell).

 

[the54thvoid]

Vulkan used to be OpenGL. Do you know what all runs on OpenGL as well? Prety much everything in the "serious" industry, medicine, engineering.

Tesla.

http://www.nvidia.com/object/gpu-applications-domain.html