AMD discrete GPU market share eroded to less than 20 per cent

[Corky34]

Wouldn't Physx want to do compute alongside other rendering tasks?

Ideally at least, it obviously doesn't have to.

How it handles simultaneous compute and rendering tasks is beyond my understanding I'm afraid, if I had to guess I would say the drivers under DX11 dealt with the scheduling of the pipeline to reduce the performance impact, something that seems to be backed up with someone, i can't remember who, saying how Nvidia drivers were written to use spare threads to swap out shaders.

Not read it myself yet but this post by Mahigan on the Overclock.net forums speaking of "driver interventionism" at a guess that maybe what Nvidia's DX11 drivers are doing when having to do compute and rendering tasks.

Corky - are you in the industry?

No, just an interest in CPU/GPU design (no matter who makes them) with far to much time on his hands.

[Jimbo75]

If you read that without your rose tinted glasses you would see that's not what he says at all, what he says is that console guys are getting 30% GPU performance by using Async Compute, something that contrary to what the Oxide representative seems to think is supported by Maxwell, and this PDF (page 31) explains why they got such a performance hit when using it.

Basically Kepler & Maxwell has 1 pipeline which can handle a lot of compute or 1 graphics queues, but it can't do them at the same time without a performance penalty, such a design works fine for DX11 because prior to DX12 there was no way for rendering to occur simultaneously with compute, so there was no need for parallel pipeline/engines like AMD did with GCN.

TBH I would've expected a developer that's working with an API that gives them greater control of how the hardware deals with their code to have known this, it just goes to show how little (imho) Oxide and yourself understand about the subject your discussing, rather than being objective you and Oxide prefer to be sensationalist.

No I know fine well about Kepler's and Maxwell's extreme drawbacks in async compute. Been telling you and everybody else for the past 12 pages and that's just in this thread.

In fact I'm pretty sure it was me who linked that GameWorks VR presentation and that very page in any other thread about VR first.

But let me (once again) help you out. Kollock has only made 2 posts on OCN and his 2nd one was this -

AFAIK, Maxwell doesn't support Async Compute, at least not natively. We disabled it at the request of Nvidia, as it was much slower to try to use it then to not.

Weather or not Async Compute is better or not is subjective, but it definitely does buy some performance on AMD's hardware. Whether it is the right architectural decision for Maxwell, or is even relevant to it's scheduler is hard to say.

Nvidia has some kind of thing that *they* are calling async compute but they have a strange version of it that makes performance so bad it's better off being disabled. But hey, at least having something (no matter how bad) lets them get DX12_1 certification out of their DX11 cards. I wonder how many mugs will fall for that one.

[Corky34]

No I know fine well about Kepler's and Maxwell's extreme drawbacks in async compute. Been telling you and everybody else for the past 12 pages and that's just in this thread.

Since when?

IIRC all you've been doing is spouting off about how crap Nvidia are, and how wonderful AMD are, in between that you've been insulting anyone who doesn't agree with your twisted view of the world, a twisted view that seems to ignore the fact that until recently the main graphics API (DX11) ONLY OPERATED IN PARALLEL, that by the time there's any significant change over to DX12, if that even happens, the fact that Maxwell can only operate in parallel won't matter as we'll be 2-3 generations down the line.

IIRC I've already said if the serial nature of DX12 was introduced when GCN first hit the market things would have panned out very differently, as it is AMD has had a GPU that could operate in serial but with no means to do that for the last 3-4 years, it's why they were forced into developing Mantle.

If I had to guess I would say AMD expected Windows 8 and DX11 to be exactly what DX12 is now (serial in nature), when things didn't pan out as they'd hoped they were stuck with a GPU but no API to take advantage of it's strengths.

For you to define Kepler's and Maxwell's as having extreme drawbacks in async compute just goes to show why you've lost all objectivity, until recently that didn't matter as there was NO API AVAILABLE that could do that, and by the time the recently released API that enables async compute is in wide spread use the fact that older cards are serial in nature wont make the slightest difference.

In fact I'm pretty sure it was me who linked that GameWorks VR presentation and that very page in any other thread about VR first.

Would you mean this thread, where once again you appear to have a very biased view, that as far as I can tell contains no reference from you to Nvidia.

But let me (once again) help you out. Kollock has only made 2 posts on OCN and his 2nd one was this -

Again touting not only your lack of understanding, but also the lack of understanding that a representative working for the company has about the hardware their working with, is it any wonder that Nvidia saw a regression when going from DX11 to DX12 when the company doesn't even have a basic understanding of the hardware and has to be told what to do.

Nvidia has some kind of thing that *they* are calling async compute but they have a strange version of it that makes performance so bad it's better off being disabled. But hey, at least having something (no matter how bad) lets them get DX12_1 certification out of their DX11 cards. I wonder how many mugs will fall for that one.

Do you even know what async compute means? Or for that matter what async means?

[Jimbo75]

Since when?

IIRC all you've been doing is spouting off about how crap Nvidia are, and how wonderful AMD are, in between that you've been insulting anyone who doesn't agree with your twisted view of the world, a twisted view that seems to ignore the fact that until recently the main graphics API (DX11) ONLY OPERATED IN PARALLEL, that by the time there's any significant change over to DX12, if that even happens, the fact that Maxwell can only operate in parallel won't matter as we'll be 2-3 generations down the line.

IIRC I've already said if the serial nature of DX12 was introduced when GCN first hit the market things would have panned out very differently, as it is AMD has had a GPU that could operate in serial but with no means to do that for the last 3-4 years, it's why they were forced into developing Mantle.

If I had to guess I would say AMD expected Windows 8 and DX11 to be exactly what DX12 is now (serial in nature), when things didn't pan out as they'd hoped they were stuck with a GPU but no API to take advantage of it's strengths.

For you to define Kepler's and Maxwell's as having extreme drawbacks in async compute just goes to show why you've lost all objectivity, until recently that didn't matter as there was NO API AVAILABLE that could do that, and by the time the recently released API that enables async compute is in wide spread use the fact that older cards are serial in nature wont make the slightest difference.

Holy crap that was a bizarre and long-winded way of admitting Kepler and Maxwell suck at DX12.

Again touting not only your lack of understanding, but also the lack of understanding that a representative working for the company has about the hardware their working with, is it any wonder that Nvidia saw a regression when going from DX11 to DX12 when the company doesn't even have a basic understanding of the hardware and has to be told what to do.

Right so you're saying that even though this "representative" has pointed out - clearly - that Nvidia suffered a regression in performance that was so severe that Nvidia themselves advised they turn it off?

How *exactly* do you spin that? When Nvidia themselves told them to turn off their "async compute".

Do you even know what async compute means? Or for that matter what async means?

Yes it means not synchronised. Imagine synchronised swimmers. That aren't. They'll still finish the dance but not at the same time.

Now imagine that Nvidia sponsored them. The problem is they picked a team that can't swim, let alone asynchronously, so they all drowned.

Nvidia's "async compute" attempt simply ends with them being drowned under context switches. That's what happens when you try to hack it instead of having the hardware, the actual ACE's. Well, AMD sure holds all the ACE's lol.

[HalloweenJack]

No, just an interest in CPU/GPU design (no matter who makes them) with far to much time on his hands.

so therefore as much information as anyone else who can use google. and an opinion of that information as well.

[Corky34]

so therefore as much information as anyone else who can use google. and an opinion of that information as well.

Indeed and your welcome to form your own opinion on all the available information as well, that's generally how discussions work when the people involved aren't fanbois with their own agenda to push on people.

[CAT-THE-FIFTH]

AMD chimes in on the matter now(thanks Calin Banc on OcUK forums):

Oxide effectively summarized my thoughts on the matter. NVIDIA claims "full support" for DX12, but conveniently ignores that Maxwell is utterly incapable of performing asynchronous compute without heavy reliance on slow context switching.

GCN has supported async shading since its inception, and it did so because we hoped and expected that gaming would lean into these workloads heavily. Mantle, Vulkan and DX12 all do. The consoles do (with gusto). PC games are chock full of compute-driven effects.

If memory serves, GCN has higher FLOPS/mm2 than any other architecture, and GCN is once again showing its prowess when utilized with common-sense workloads that are appropriate for the design of the architecture.

http://forums.overclockers.co.uk/sho...&postcount=317

Also:

http://forums.anandtech.com/showpost...&postcount=560

I do think that early adopters of the aftermarket R9 290 cards,will get a long run from their cards.

[Corky34]

Isn't Maxwell 2 different though?
From my understanding it can do 1 graphics + 32 compute, unlike its predecessors (Kepler, Maxwell 1) neither of which can process both a Graphics and a Compute command in parallel.

The CGN is crossed out as apparently Anandtech got those number wrong.

It's interesting that what the AMD rep posted confirms what I suspected, that AMD expected an API that supported parallel workload queues sooner.

[CAT-THE-FIFTH]

It could be why Mantle appeared too if you come to thing of it.

[HalloweenJack]

At this time it`ll be 8 months before Nv get unified memory ; and whilst both GCn and Maxwel are DX12 - they support slightly different feature set =

AMD support Resource binding 3 and typed UAV formnats 2 ; both `behind the scenes` features , and ones which all the consoles will use as they all want 1080P @60fps and if those 2 features mean they can do well they jolly will use them.


whereas

NVidia Maxwell supports raster ordered views and conservative rasterization 2 both are great for making games look `pretty` (intel supports ROV 2 as well) ; its also the reason why , IMO , we wont see Fermi getting proper DX12 support - it barely gets in under the door for any of the feature levels - its a FL 11 card just. Nv should just say ` thanks GTX 480 - your a DX11 card only` instead of breaking drivers to get it to work.

[CAT-THE-FIFTH]

Apparently asynchronous shading is going to be used in three major DX12 games:

http://www.overclock.net/t/1569897/v...#post_24365329

Isn't Mirror's Edge Catalyst an Nvidia sponsored title though?? Unless the dev is going the GTAV route and optimising for both companies.

[DanceswithUnix]

Isn't Mirror's Edge Catalyst an Nvidia sponsored title though?? Unless the dev is going the GTAV route and optimising for both companies.

The 970 seems to have sold very well, I can't see why Nvidia would be unhappy at a feature in that chip being used.

[CAT-THE-FIFTH]

The 970 seems to have sold very well, I can't see why Nvidia would be unhappy at a feature in that chip being used.

Not if it runs a bit meh!!

[DanceswithUnix]

Not if it runs a bit meh!!

lol, well they have something that they call async compute in there, they just have to beat on the driver until it runs OK. That might need custom code paths for every game that comes out using the feature, but Nvidia are good at that sort of thing

[CAT-THE-FIFTH]

lol, well they have something that they call async compute in there, they just have to beat on the driver until it runs OK. That might need custom code paths for every game that comes out using the feature, but Nvidia are good at that sort of thing

AFAIK,there is noise it is emulated in software which might be the problem,but we will wait and see.

But the thing is Mirrors Edge is going to be on consoles,which is why asynchronous shaders are being used. It will be interesting to see if its easier for Nvidia to just gimp that functionality on the PC version.

[HalloweenJack]