AMD’s DirectX 12 Advantage Explained – GCN Architecture More Friendly To Parallelism Than Maxwell

Howling · 26 Aug 2015 at 16:18

andybird123 said:
Yes, he is still saying that Ashes is specifically coded to take advantage of GCN, probably to the detriment of Maxwell

how other games approach this remains to be seen

basically he is highlighting what we already know - with DX12 optimisation is in the hands of developers

Looks to me like it's coded to be efficient (ie Parallel) rather than catering to a specific architecture. The current nVidia line of cards are efficient in serial but not in parallel, this is highlighted in Ashes. nVidia's next GPU is likely to be more focused on parallel processing and will probably be as good, if not better, than AMD's current cards - much larger R&D budget ya know.

Harlequin · 26 Aug 2015 at 16:18

andybird123 said:
Yes, he is still saying that Ashes is specifically coded to take advantage of GCN, probably to the detriment of Maxwell

how other games approach this remains to be seen

basically he is highlighting what we already know - with DX12 optimisation is in the hands of developers

no he`s not - he`s saying ashes is written to make use of the biggest feature of DX12 - Asynchronus shading.

humbug · 26 Aug 2015 at 16:18

andybird123 said:
Yes, he is still saying that Ashes is specifically coded to take advantage of GCN, probably to the detriment of Maxwell

how other games approach this remains to be seen

basically he is highlighting what we already know - with DX12 optimisation is in the hands of developers

Its a case of taking advantage of something in GCN 1.1/1.2 which it does better than Maxwell.

Its like Maxwell is a 2 core CPU while GCN 1.1/1.2 is an 8 core CPU, Oxide have coded to use all 8 cores so thats the one that wins.

andybird123 · 26 Aug 2015 at 16:21

Harlequin said:
no he`s not - he`s saying ashes is written to make use of the biggest feature of DX12 - Asynchronus shading.

yes, but he's also saying that ONLY using AS and not having a separate code path for NVidia is what is causing the decrease in performance from DX11

again, at the very least DX12 should be no slower than DX11, unless the developer flat refuses to optimise for NV hardware

there are also other aspects of DX12 that are NV specific, which the developer seems not to have made use of

Harlequin · 26 Aug 2015 at 16:25

andybird123 said:
yes, but he's also saying that ONLY using AS and not having a separate code path for NVidia is what is causing the decrease in performance from DX11

again, at the very least DX12 should be no slower than DX11, unless the developer flat refuses to optimise for NV hardware

there are also other aspects of DX12 that are NV specific, which the developer seems not to have made use of

so instead of using vendor specific coding = for NVidia , and coding for DirectX 12 path he is wrong?

really? no - he is showing what DX12 can do. And NO , DX12 has rendering tiers - to qualify for tier 3 you need a certain set of abilities - which all parties who qualify would use.

vendor specific is not in the DX12 spec.

andybird123 · 26 Aug 2015 at 16:29

Harlequin said:
so instead of using vendor specific coding = for NVidia , and coding for DirectX 12 path he is wrong?

really? no - he is showing what DX12 can do. And NO , DX12 has rendering tiers - to qualify for tier 3 you need a certain set of abilities - which all parties who qualify would use.

vendor specific is not in the DX12 spec.

yes, but the fact remains that currently AMD and NVidia support a different set of capabilities, so using one of the AMD only ones (and sticking the AMD logo on your homepage) is a pretty clear statement of something - there are other currently Nvidia only features in DX12 that could be improving performance on NVidia hardware albeit in a different way

I'm not saying anyone is "wrong", I'm saying this game has been primarily coded for AMD hardware, there's nothing wrong with that, but it doesn't mean that all DX12 games will reflect this

Harlequin · 26 Aug 2015 at 16:31

andybird123 said:
yes, but the fact remains that currently AMD and NVidia support a different set of capabilities, so using one of the AMD only ones (and sticking the AMD logo on your homepage) is a pretty clear statement of something

no - he coded for DX12 spec , not for anything vendor specific. Its not his fault that NVidia made hardware that was great in serial rendering eg DX10 and 11 , but not so great with DX12 ; its the way AMD went especially with Mantle.

so now games will either have to be DX12 or Nvidia specific until Nv catch up - and with the delay to HBM2 (Samsung will be H1 for it , so could be march next year before pascal at the earliest)

edit

seems you have ninja edited - no there are NO NVidia specific features in DX12 - an open api doesn't work like that. Nvidia only would be GPU PhysX , specific api would GLide for example. DX12 was written by Microsoft , its down to the IHV`s at what feature level they comply with.

Howling · 26 Aug 2015 at 16:33

andybird123 said:
yes, but the fact remains that currently AMD and NVidia support a different set of capabilities, so using one of the AMD only ones (and sticking the AMD logo on your homepage) is a pretty clear statement of something - there are other Nvidia only features in DX12 that could be improving performance on NVidia hardware albeit in a different way

I'm not saying anyone is "wrong", I'm saying this game has been primarily coded for AMD hardware, there's nothing wrong with that, but it doesn't mean that all DX12 games will reflect this

Sensible comment. I've said since the Ashes benchmark release that we should never look at just one benchmark to draw conclusions. All the million (green) people calling it a broken or flawed benchmark are just butthurt retards imho. It's a totally valid benchmark, one of many to come I'd hope and if the next benchmark gives nVidia better results, great -- another totally valid result to add to the pool.

humbug · 26 Aug 2015 at 16:37

Its quite simple, GCN 1.1/1.2 has 8 ASynchronous Shader Engines (ACE Units)

Maxwell V2 has 2

So AMD can process 8 engine threads simultaneously, Nvidia can only process 2, as a result Nvidia suffer from a lot more task queuing, that adds Latency because if 2 tasks are running other tasks have to queue, that results in lower performance. its why in heavy scheduling tests the 390 beats the 980TI.

Its not necessarily a problem with Oxide not having a separate code path for Nvidia, its that the AMD GPU does this at the hardware level, it 'may' be possible to make Nvidia more efficient at this but it would require the Developer to engineer it into the software, its a lot more work, its also unlikely to end up anywhere near as efficient as having 8 ASynchronous Shader Engines in the hardware.

8 Core CPU vs 2 Core CPU.

humbug · 26 Aug 2015 at 16:44

One other thing, this might explain why Maxwell is so much more power efficient than GCN and Kepler.

Harlequin · 26 Aug 2015 at 16:49

I do think we might not see any real DX12 drivers for Fermi from Nv - its feature tier is level 1 , same as the 5xxx series

Harlequin · 26 Aug 2015 at 16:55

Conservative rasterization can be performed in the geometry shader so there's no worries there ...

Raster ordered views is just a fancy way to rename Intel's pixelsync and the feature's name in OpenGL is Intel_fragment_shader_ordering which by the way GCN already supports so ROVs are already covered with GCN ...

so Fury is already DX12 tier 3....

humbug · 26 Aug 2015 at 16:56

Harlequin said:
I do think we might not see any real DX12 drivers for Fermi from Nv - its feature tier is level 1 , same as the 5xxx series

Would not do much good on any architecture below GCN/Kepler, those architectures are not advanced enough to take advantage of DX12 on the GPU side tho it may help them on the CPU side.

My suspicion is Kepler may also do better in some situations than Maxwell.

bru · 26 Aug 2015 at 16:56

So 2 ACE units vs 8 ACE units in AMD's favour, I'm surprised the FuryX isn't much further ahead. Even if this is a best/worst case scenario, the top cads are still not miles apart.
Lets see what the next benchmark out brings.

Harlequin · 26 Aug 2015 at 16:58

humbug said:
Would not do much good on any architecture below GCN/Kepler, those architectures are not advanced enough to take advantage of DX12 on the GPU side tho it may help them on the CPU side.

My suspicion is Kepler may also do better in some situations than Maxwell.

I'm not sure sure as Kepler is absed upon Fermi - I`ve sen links to say Kepler is also Tier 1.

if anyone is interested have a read:

https://software.intel.com/en-us/articles/deep-shading-buffers-on-commodity-gpus

a good read and one that lends weight to Gtris being the bottleneck with Fury

humbug · 26 Aug 2015 at 17:03

bru said:
So 2 ACE units vs 8 ACE units in AMD's favour, I'm surprised the FuryX isn't much further ahead. Even if this is a best/worst case scenario, the top cads are still not miles apart.
Lets see what the next benchmark out brings.

Something is still holding Fiji Back, the thing has a colossal amount of raw power! 8.6TFlops.

in time we may or may not see the true performance of the thing, i just don't know what is holding it back.

But Hawaii's true power is looking clear now.

Harlequin said:
I'm not sure sure as Kepler is absed upon Fermi - I`ve sen links to say Kepler is also Tier 1.

if anyone is interested have a read:

https://software.intel.com/en-us/articles/deep-shading-buffers-on-commodity-gpus

a good read and one that lends weight to Gtris being the bottleneck with Fury

i will read that later thanks, maybe the answer is in there.

queamin · 26 Aug 2015 at 17:15

There will be games where maxwell 2 win and are stronger than fury/390, very high tessellation for one the amd cards will hit a limit sooner and if Nvidia's cards don't get to bottlenecked this I can see them winning a lot of titles, it is swings and roundabouts really.

NirK · 26 Aug 2015 at 17:18

Am I missing something here, I just did a search for ashes of singularity benchmarks of 980ti vs fury x and fury x only wins by a very small amount at 4k, not exactly "8 cores vs 2 cores" as was said in this thread. Fury x wins by a small amount at 4k in some DX11 games as well.

queamin · 26 Aug 2015 at 17:24

Because it is in Alpha and with any card there is bottlenecks, you overcome one you get another, amd cards are better at some things and Nvidia are better at others.

Lokken86 · 26 Aug 2015 at 17:30

queamin said:
Because it is in Alpha and with any card there is bottlenecks, you overcome one you get another, amd cards are better at some things and Nvidia are better at others.

That didn't make an awful lot of sense. Surely then nvidia will also be hindered by it being in alpha.

What doesn't add up to me is there are some benches showing the furyx beating the 980ti by a frame or 2 but yet others are showing the 290x almost neck and neck with it and leaving out the fury results.