Only AMD has true Async Compute - Doom Devs

[D.P.]

News to me, where is this benchmark compared with Maxwell?



So you would admit Mantle is in DX12?

On Nvidia website.

No. AMD contributed to DX11 with ideas from Manlte, Nvidia also contibuted, as dis ibtel and dozens of others. DX12 was in development team long before Manlte and some of the key concepts were in discusion for some time because they are just onvious. Of course the API needs to be multithreaded etc.

 

[Ferrari100]

taken from:
http://www.extremetech.com/extreme/...ading-amd-nvidia-and-dx12-what-we-know-so-far


“There were claims originally, that Nvidia GPUs wouldn’t even be able to execute async compute shaders in an async fashion at all, this myth was quickly debunked. What become clear, however, is that Nvidia GPUs preferred a much lighter load than AMD cards. At small loads, Nvidia GPUs would run circles around AMD cards. At high load, well, quite the opposite, up to the point where Nvidia GPUs took such a long time to process the workload that they triggered safeguards in Windows. Which caused Windows to pull the trigger and kill the driver, assuming that it got stuck.


Adding a working pre-emption model to prioritise more important compute tasks alleviates the problem of crashing on Nvidia GPU's but does not improve the Asynchronous nature of offloading compute tasks in parallel, at will - unless Nvidia prove otherwise.

 

[humbug]

No. AMD contributed to DX11 with ideas from Manlte, Nvidia also contibuted, as dis ibtel and dozens of others. DX12 was in development team long before Manlte and some of the key concepts were in discusion for some time because they are just onvious. Of course the API needs to be multithreaded etc.

There is nothing like that on Nvidia's site.

No. AMD contributed to DX11 with ideas from Manlte, Nvidia also contibuted, as dis ibtel and dozens of others. DX12 was in development team long before Manlte and some of the key concepts were in discusion for some time because they are just onvious. Of course the API needs to be multithreaded etc.

Yeah this is Nvidia's marketing, and yet its AMD who benefit from DX12, not Nvidia.

I believe facts over Nvidia's marketing.

 

[drunkenmaster]

There is an earlier segment on non-VR performance (which is plainly non-VR gaming) where (albeit the graphs are kind of funky) but its plainly stated as being around 1.6-1.7X the performance of a 980 and around 1.2-1.3x the performance of a TX and then moves onto a VR segment - it would be kind of weird to having spent some time promoting its normal gaming performance as one thing then to start claiming it was another later on - especially in what is plainly marked as a VR segment.

Yup, earlier he claimed the difference was smaller but with a graph that made a bigger number between the 1080 and Titan X by putting the baseline what, below a 950 or something, it's been a while since I watched it.

However not long after saying it was twice as fast in VR, he started just calling it twice as fast without qualifying it with in VR and he repeated it over and over.

 

[drunkenmaster]

On Nvidia website.

No. AMD contributed to DX11 with ideas from Manlte, Nvidia also contibuted, as dis ibtel and dozens of others. DX12 was in development team long before Manlte and some of the key concepts were in discusion for some time because they are just onvious. Of course the API needs to be multithreaded etc.

Yup, DX12 being a low level API which was the point of Mantle and all the main DX12 features being there in Mantle, and Vulkan which used Mantle as the starting point(as well), but Mantle contributed to DX11, and DX12 was in development for donkeys years.... yet uses from what I know, all the same calls marginally renamed as Mantle.

On top of that Nvidia working with MS for 5 years(supposedly) managed to not address several major features of DX12 and managed to lie about Fermi DX12 support and lied for a year(and kinda still are) about hardware async support coming for Maxwell. Architectures available after Nvidia claim they started working on DX12 with MS are not available in Nvidia hardware, but AMD who started with Mantle first, and support all the most important features of DX12 and have hardware async also have this in their older 7970 architecture.

Here is a hint, when someone is really involved with building the API from the ground up with 5 years notice... they don't generally lack support for fundamental features they supposedly came up with.

 

[flopper]

On Nvidia website.

No. AMD contributed to DX11 with ideas from Manlte, Nvidia also contibuted, as dis ibtel and dozens of others. DX12 was in development team long before Manlte and some of the key concepts were in discusion for some time because they are just onvious. Of course the API needs to be multithreaded etc.

AMD has true async compute, Nvidia dont.
isnt that funny how good AMD is and how they support true async compute vs some half hack attempt at it?

TRUE Async compute from AMD
No one else.....

 

[brawl3r]

Can we put this to rest, with maybe 3d mark making a Async compute benchmark.

 

[D.P.]

AMD has true async compute, Nvidia dont.
isnt that funny how good AMD is and how they support true async compute vs some half hack attempt at it?

TRUE Async compute from AMD
No one else.....

Someone else who needs to educate themselves on async compute instead of spurting out PR pieces without the slightest inkling of what you are saying

 

[Rroff]

The fact that people on here and in real life were going on about Twice the Perf tells me it had the effect it was meant to. Why else would he keep repeating it other than trying to drum this fact in to people heads.

Yup, earlier he claimed the difference was smaller but with a graph that made a bigger number between the 1080 and Titan X by putting the baseline what, below a 950 or something, it's been a while since I watched it.

However not long after saying it was twice as fast in VR, he started just calling it twice as fast without qualifying it with in VR and he repeated it over and over.

Unlike the whole overclockers dream though in the presentation is does explicitly state multiple times where the card stacks up in non-VR performance sure there might be a bit of salesmanship to big it up and leave the VR performance as the last thing people remember - though its fairly obvious as he mentions both the 2x/3x thing that he is repeating the last VR slide if you pay attention.

i was talking about async not preemption, and i still dont believe preemption and async compute are the same thing.
to me preemption is a tool used to prioritize queues and submitting 1 at a time, and async allows you to submit multiple workload queued simultaneously getting rid of bubbles that waste space that could have been used for more instructions, while async allow a more efficient way of doing it.
at least thats my understanding of it, and keep in mind i am not a developer, i just read articles like everyone and try to understand correctly what is it about.
if i have everything wrong feel free to correct me.

Pre-emption doesn't limit to serialising the whole workload - it tries to predict the most efficient way to despatch multiple commands if appropriate so that the GPU doesn't get hung up on a long task that can't be stopped until the next drawcall when there is more work waiting to be done and try and load up the tasks to make as efficient use of the time as possible. The problem is AMD have slapped ASync over a wide range of both parallel compute functionality (which Pascal and Maxwell absolutely do have) and ASync capabilities in the scheduling which Maxwell falls down on and Pascal has to rely on tricks to do the equivalent thing so you end up with people arguing at cross purposes as different sides assume the other is arguing about something slightly different.

EDIT: Made some changes to wording to more accurately reflect what I'm trying to say.

 

[Ferrari100]

Pre emption:

https://en.wikipedia.org/wiki/Preemption_(computing)

Asynchronous communication:

https://en.wikipedia.org/wiki/Asynchronous_communication


Put the TWO together and you have AMD's Asynchronous Compute Engine Model.

 

[Lokken86]

Looks like he has had to overcome a lot to get up there and speak publicly none the less I assume English is a second language for him.

Isn't he Indian? They speak English.

 

[fs123]

Someone else who needs to educate themselves on async compute instead of spurting out PR pieces without the slightest inkling of what you are saying

I'm pretty sure the Doom devs said AMD had true Async Compute in the video. Are they talking BS too?

 

[Tute]

I can't say i'll be going for a 480 over a 1070 as coming from a Nano I was hoping to move to at least Fury X/980Ti performance.

I might have been interested in a 480X mind, I guess we'll have to wait and see about that though.

[Edit] Possibly wrong thread, sowwy.

 

[Rroff]

Isn't he Indian? They speak English.

He doesn't speak it as if its his first language though.

Pre emption:

https://en.wikipedia.org/wiki/Preemption_(computing)

Asynchronous communication:

https://en.wikipedia.org/wiki/Asynchronous_communication


Put the TWO together and you have AMD's Asynchronous Compute Engine Model.

2 random wiki links that talk generally about 2 elements of the subject and don't really have much baring on comparing the implementation of the 2 without a lot more further reading.

 

[TheRealDeal]

Unlike the whole overclockers dream though in the presentation is does explicitly state multiple times where the card stacks up in non-VR performance sure there might be a bit of salesmanship to big it up and leave the VR performance as the last thing people remember - though its fairly obvious as he mentions both the 2x/3x thing that he is repeating the last VR slide if you pay attention.

I agree with what you are saying. For people like us it was easy to understand but if you go onto Youtube and watch parts without a good understanding then i can see why people might be taken in. It was good marketing from my stand point of view as i seen it work for real. Go back and check the thread and you will see a few on here who fell for it.

 

[D.P.]

I'm pretty sure the Doom devs said AMD had true Async Compute in the video. Are they talking BS too?

No one said AMD didn't.

The BS comes from people saying Pascal doesn't do async compute, or only does it in software, or can't do it simultaneously. Anyone who says anything Ali g those lines are either ignorant or trolls.

 

[stooeh]

Booyakasha?

 

[Dave2150]

There is nothing like that on Nvidia's site.



Yeah this is Nvidia's marketing, and yet its AMD who benefit from DX12, not Nvidia.

I believe facts over Nvidia's marketing.

Same here, plus no point arguing with an individual financially integrated with NVIDIA, it's his only aim on this forum - sell more NVIDIA cards.

 

[fs123]

No one said AMD didn't.

The BS comes from people saying Pascal doesn't do async compute, or only does it in software, or can't do it simultaneously. Anyone who says anything Ali g those lines are either ignorant or trolls.

Booyakasha?

Respek Lol.
I think DP is working overtime. Damage control is hurting his fingers or his keyboard..

 

[brawl3r]

Booyakasha?

Hahaha I spat my drink out after reading that.
Ali g = along those lines, ROFL