Probability over the next 4+ years that DX12 Feature level 12_1 wont deliver tangible benefits?

[pmc25]

They're not the important part of DX12. Lower overhead, native multi-GPU support and asynchrous shaders are the important bits. In terms of hardware feature support asynchrous shaders is the most important bit ... and NVIDIA can't do that until Pascal.

 

[layte]

They're not the important part of DX12. Lower overhead, native multi-GPU support and asynchrous shaders are the important bits. In terms of hardware feature support asynchrous shaders is the most important bit ... and NVIDIA can't do that until Pascal.

Nvidia have similar but differently named functionality in Maxwell v2.

 

[pmc25]

Nvidia have similar but differently named functionality in Maxwell v2.

They don't. They emulate it in software which is very slow.

 

[layte]

They don't. They emulate it in software which is very slow.

No, you are wrong. Hyper-Q in Maxwell v2 allows for 32 simultaneous compute queues in graphics mode. Whilst that is half what GCN offers, it is a massive step up from the single queue currently on offer in DX11.

 

[pmc25]

No, you are wrong. Hyper-Q in Maxwell v2 allows for 32 simultaneous compute queues in graphics mode. Whilst that is half what GCN offers, it is a massive step up from the single queue currently on offer in DX11.

Are you sure that's not emulation? I think it is ....

 

[humbug]

Im considering the long term (4+yr) value of a cheaper card as opposed to maxwell 2 purchase in relation dx12 feature level support. (Fury/GCN 1.2 is 12_0 it would seem?)

Is it reasonable to predict any of

• that 12_1 feature levels wont be adopted in games in a significant way because consoles dont* support these feature levels (or another reason)?
• that if 12_1 feature levels were adopted, that they could could be emulated to no significant detriment on a high end CPU?
• that supporting these features levels is of no benefit in general or to the current class of hardware (eg dont have the horsepower to use)?

Im wondering how much value to ascribe to Feature level 12_1 support over the long term....

*(assumption based on the age/class of gpu).

thanks

DX12.1 at this point appears to be something made up by Nvidia, there is not reference to it from MS.

 

[Dygaza]

People are confusing 12_1 feature level with 12.1.

Just like people were confusing DX11 FT 11_0 and 11_1 with DX12 11_0 and 11_1.

Can't really blame people, as it's ms's own fault for giving them so missleading names.

Feature level A, B, C, D would have been way more better.

 

[David Bisset]

DX12.1 at this point appears to be something made up by Nvidia, there is not reference to it from MS.

I thought I'd seen a reference before for feature level 12_1 but can't find it now, only the Tiers stuff ( https://msdn.microsoft.com/en-us/library/dn899127.aspx )

Are you sure that's not emulation? I think it is ....

Looks like it's hardware according to http://www.nvidia.co.uk/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf
'...by allowing 32 simultaneous, hardware-managed connections...'

 

[humbug]

I thought I'd seen a reference before for feature level 12_1 but can't find it now, only the Tiers stuff ( https://msdn.microsoft.com/en-us/library/dn899127.aspx )

Yes i know about those, they are DX11 feature sets. its likely they along with the rest of DX11 may be pulled over to conglomerate DX12.

I just haven't seen anything Microsoft specifically calls DX12.1

 

[ubersonic]

No not that long.

It was two years after Direct 3D 10 launched before games started properly using it and three before they were widely using it, so 2-3 years for D3D 12 sounds about right

 

[pmc25]

I thought I'd seen a reference before for feature level 12_1 but can't find it now, only the Tiers stuff ( https://msdn.microsoft.com/en-us/library/dn899127.aspx )



Looks like it's hardware according to http://www.nvidia.co.uk/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf
'...by allowing 32 simultaneous, hardware-managed connections...'

99% sure that's their emulation of asynchrous shaders. It may be tied to certain hardware - i.e. Kepler / Maxwell / Maxwellv2, but it is not asynchrous shaders.

Pascal is their first architecture which has native, hardware based asynchrous shading, as opposed to this software/middleware halfway house. The so-called "HyperQ" if used might actually work well in the old APIs (or better than unused or unoptimised hardware based asynch) ... in DX12 / Vulkan / Mantle which support asynchrous shading natively it's going to be very slow compared to the latest GCN cards, both due to being inherently slower and having half the paths / connections.

 

[D.P.]

They don't. They emulate it in software which is very slow.

Wrong, Asynchronous shades is just an AMD term, your argument is like saying that AMD cards don't have geforce experience software. Nvidia just call ti hyper Q or some such. And it really isn't that exciting, it i like hyperthreading on Intel CPus, you can get a slight performance increase and efficiency gain but it isn't a miracle.

 

[D.P.]

Im considering the long term (4+yr) value of a cheaper card as opposed to maxwell 2 purchase in relation dx12 feature level support. (Fury/GCN 1.2 is 12_0 it would seem?)

Is it reasonable to predict any of

• that 12_1 feature levels wont be adopted in games in a significant way because consoles dont* support these feature levels (or another reason)?
  
• that if 12_1 feature levels were adopted, that they could could be emulated to no significant detriment on a high end CPU?
  
• that supporting these features levels is of no benefit in general or to the current class of hardware (eg dont have the horsepower to use)?

Im wondering how much value to ascribe to Feature level 12_1 support over the long term....

*(assumption based on the age/class of gpu).

thanks

As others have said, order independent transparency is a vey big deal and should increase performance.. Most modern game engines employ differed rendering which is much more efficient than forward rendering since only screen pixels end up being processed in pixel shades as opposed to all the hidden fragments. The problem is differed rendering makes it impossible to do transparency so to do that an additional processing pass is required that renders the transparent surfaces on top of the opaque background. OIT will allow developers to send transparent surfaces to the GPU out of order and have the GPU render the transparency in the correct order within a single pass.



The main thing that I would look out for going forwards is vram. with DX12 games wont be limited by the number of draw calls, therefore the number of unique objects rendered on the screen can increase dramatically. At the movement the DX11 API is very inefficient so only a certain number of render instructions can be sent form the CPU to the GPU, and therefore there can only be a few thousand objects at most, form mountains, plants, trees, buildings, cards, sign posts, planes, people, animals, everything. TTo make things like a convincing forest it isn't possible to have that number of unique trees therefore developers use a technique called instancing so the same draw call can end up essentially copying the same object in multiple locations. With DX12 hundreds of unique different tree types could be rendered, along with hundred of different types of grass blades, plants, flowers etc.
This of course means instead of 1 or 2 models getting duplicated hundreds of times there are now hundred of different models with unique assets like texture maps, normal maps, geometry. This will push vram usage up dramaticly.

The other driving force will be high res textures for 4K use, we already see this in some games where 4GB cram is surpassed and causes stuttering on cards with insufficient vram.

 

[Calin Banc]

No not that long.

It was two years after Direct 3D 10 launched before games started properly using it and three before they were widely using it, so 2-3 years for D3D 12 sounds about right

Don't forget that the older consoles were dx9 only, while all current gen are dx12 ready.

 

[humbug]

As others have said, order independent transparency is a vey big deal and should increase performance.. Most modern game engines employ differed rendering which is much more efficient than forward rendering since only screen pixels end up being processed in pixel shades as opposed to all the hidden fragments. The problem is differed rendering makes it impossible to do transparency so to do that an additional processing pass is required that renders the transparent surfaces on top of the opaque background. OIT will allow developers to send transparent surfaces to the GPU out of order and have the GPU render the transparency in the correct order within a single pass.



The main thing that I would look out for going forwards is vram. with DX12 games wont be limited by the number of draw calls, therefore the number of unique objects rendered on the screen can increase dramatically. At the movement the DX11 API is very inefficient so only a certain number of render instructions can be sent form the CPU to the GPU, and therefore there can only be a few thousand objects at most, form mountains, plants, trees, buildings, cards, sign posts, planes, people, animals, everything. TTo make things like a convincing forest it isn't possible to have that number of unique trees therefore developers use a technique called instancing so the same draw call can end up essentially copying the same object in multiple locations. With DX12 hundreds of unique different tree types could be rendered, along with hundred of different types of grass blades, plants, flowers etc.
This of course means instead of 1 or 2 models getting duplicated hundreds of times there are now hundred of different models with unique assets like texture maps, normal maps, geometry. This will push vram usage up dramaticly.

The other driving force will be high res textures for 4K use, we already see this in some games where 4GB cram is surpassed and causes stuttering on cards with insufficient vram.

Even then there is a limit to how many times you can run multiple cloned objects on the same Drawcall, and if your using multiple source dynamic shadow maps the whole thing pretty much falls apart.

I could go on whinging about the limitations of DX11 all day, they are plentiful and extreme, with it we haven't really moved forward from where we were 10 years ago, aside from what are basically hacks to get around those limitations, all of which are easily exhausted.
DX12, Mantle, Vulkan..... they are a very long time over due, DX12 will have a massive impact.

Its like 10 years worth of Development coming all at once.

 

[pmc25]

Wrong, Asynchronous shades is just an AMD term, your argument is like saying that AMD cards don't have geforce experience software. Nvidia just call ti hyper Q or some such. And it really isn't that exciting, it i like hyperthreading on Intel CPus, you can get a slight performance increase and efficiency gain but it isn't a miracle.

It's what MS and graphics professionals generally are referring to it by, not just AMD - it's not an AMD trademark or marketing name (HyperQ is a trademark and marketing term). I've seen a couple of feature tables erroneously claiming that Maxwellv2 support asynch, but if Maxwellv2 supported it then so would Kepler & Maxwellv1 (both have HyperQ). There's no support except emulation and HyperQ is emulation.

I've never seen a single developer or person whose opinion I'd trust on graphics claim that NVIDIA supports asynchrous shaders at any level other than emulation.

Also, given that next gen engines will use it as a major feature in their rendering pipelines (as both consoles feature it and it should help them considerably), there's likely to be a large gulf on PC versions of the same titles between GCN and pre-Pascal NVIDIA cards.

 

[ginjaninja]

No not that long.

It was two years after Direct 3D 10 launched before games started properly using it and three before they were widely using it, so 2-3 years for D3D 12 sounds about right

Whilst there is always going to be a lag, M$ seem to be billing DX12 as something quite different, akin to when cpus went from 1 to multiple cores.

in GDC video MS alluded to goal/prediction of 66% steam users on DX12 compatible hardware at launch..IF FL 12.1 was a prerequisite this number would be significantly lower.

IF the difference between pre dx12 and dx12 is 'night and day' and the user base has compatible hardware, and theres synergy with xbone environment, then maybe the takeup will be quicker than before.

It make sense that developers who have the resources to rework engines for clear performance/visual advantages/platform synergy? will want to do so asap.

 

[D.P.]

Even then there is a limit to how many times you can run multiple cloned objects on the same Drawcall, and if your using multiple source dynamic shadow maps the whole thing pretty much falls apart.

I could go on whinging about the limitations of DX11 all day, they are plentiful and extreme, with it we haven't really moved forward from where we were 10 years ago, aside from what are basically hacks to get around those limitations, all of which are easily exhausted.
DX12, Mantle, Vulkan..... they are a very long time over due, DX12 will have a massive impact.

Its like 10 years worth of Development coming all at once.

Microsoft, ATI and nvidia were basically planning DX12 back in 2005 apparently. ATI and Nvidia wanted to get lower level acess and a proper multithreaded API. Lord know why it took so long, points finger at M$.

 

[D.P.]

It's what MS and graphics professionals generally are referring to it by, not just AMD - it's not an AMD trademark or marketing name (HyperQ is a trademark and marketing term). I've seen a couple of feature tables erroneously claiming that Maxwellv2 support asynch, but if Maxwellv2 supported it then so would Kepler & Maxwellv1 (both have HyperQ). There's no support except emulation and HyperQ is emulation.

I've never seen a single developer or person whose opinion I'd trust on graphics claim that NVIDIA supports asynchrous shaders at any level other than emulation.

Also, given that next gen engines will use it as a major feature in their rendering pipelines (as both consoles feature it and it should help them considerably), there's likely to be a large gulf on PC versions of the same titles between GCN and pre-Pascal NVIDIA cards.

I've seen developers on other forums state that Maxwell V2 has very similar aync computer performance to GCN. I will take their word for it over yours, since they are professionals that actually work with such features.

 

[ginjaninja]

As others have said, order independent transparency is a vey big deal and should increase performance.. Most modern game engines employ differed rendering which is much more efficient than forward rendering since only screen pixels end up being processed in pixel shades as opposed to all the hidden fragments. The problem is differed rendering makes it impossible to do transparency so to do that an additional processing pass is required that renders the transparent surfaces on top of the opaque background. OIT will allow developers to send transparent surfaces to the GPU out of order and have the GPU render the transparency in the correct order within a single pass.



The main thing that I would look out for going forwards is vram. with DX12 games wont be limited by the number of draw calls, therefore the number of unique objects rendered on the screen can increase dramatically. At the movement the DX11 API is very inefficient so only a certain number of render instructions can be sent form the CPU to the GPU, and therefore there can only be a few thousand objects at most, form mountains, plants, trees, buildings, cards, sign posts, planes, people, animals, everything. TTo make things like a convincing forest it isn't possible to have that number of unique trees therefore developers use a technique called instancing so the same draw call can end up essentially copying the same object in multiple locations. With DX12 hundreds of unique different tree types could be rendered, along with hundred of different types of grass blades, plants, flowers etc.
This of course means instead of 1 or 2 models getting duplicated hundreds of times there are now hundred of different models with unique assets like texture maps, normal maps, geometry. This will push vram usage up dramaticly.

The other driving force will be high res textures for 4K use, we already see this in some games where 4GB cram is surpassed and causes stuttering on cards with insufficient vram.

This makes sense...Although its like divining the tea leaves, im getting the 'feeling' from the ocuk forumers.... that on balance the greater memory is going to be more valuable than FL12.1 support...probably in part because i wont pay £600 for a 980ti...and am thinking an 8GB 290/390 maybe a sweeter spot for 1080p 30fps dx12 (and dx11) gaming over 5 years than a 970/980...