AMD's Richard Huddy on the state of PC graphics, Mantle 2 and APUs

[sontin]

What you've said here doesn't make any sense.

DX12 doesn't support anything yet. It's not out. How can what DX12 is going to have influence nVidia's DX11 drivers vs AMD's DX11 drivers?

Huh? Microsoft said a few things about DX12. DX12 even supports an advanced version of DCL. :hmm:

nVidia is now able to create "driver" command buffers for DX11 engine beside the reduction of the DX11 overhead.

Mantle is only able of reducing the overhead of the graphics API. But there is impact on the multithreading side.

 

[dacostafilipe]

DX12 supports full Multithreading. Mantle is still missing this in the current release.

Ouch ...

The moment when you start to understand how a command queue works is going to be priceless!

:thumbsup:

 

[blackened23]

So, how does DirectX 12, which tries to solve the same issues as Mantle and also adds close to the metal control, manage to work on a wide range of hardware?

Is MS lying when they claim DX 12 is closer to the metal?

Glad you asked. The answer is MS is implementing features for intel, AMD, and nvidia. Yes, MS is implementing "to the metal" features for all graphics vendors, not just one. NV has worked hand-in-hand with MS for NV specific features, as has AMD. When NV demonstrated DX12 beta running Forza on their hardware, they stated that MS was implementing features into DX12 which would benefit to the metal access for nvidia hardware. Intel and AMD will do the same.

AMD's Mantle on the other hand, implements NOTHING except GCN features and therefore will never benefit anything except AMD hardware. Even in the mythical world where Mantle *did* work on everything, there would be zero percent performance improvement when AMD will not implement features for Intel or NV architectures. Aside from this, AMD informed PCPer that Mantle would be closed due to DX12 being in development which was mentioned on a prior podcast. Now suddenly AMD's marketing is going full force with the good guy BS nonsense saying that Mantle works on everything. Please.

Mantle will never implement intel or NV specific features. DX12 will and is. NV proved it with their beta DX12 drivers which were designed for DX12. If DX12 were a Mantle template, but AMD wasn't chosen to demonstrate DX12 running Forza? That's interesting isn't it? Fact of the matter is, Mantle is AMD's for AMD hardware and nothing else. It will never implement NV or Intel GPU architecture features. Period. And that is fine. I have no issue with this, this is what AMD should have been doing YEARS ago. AMD is doing what they should be doing: creating value adds to make their cards more appealing. Mantle is an appealing feature for AMD users. But just call Mantle what it is and not what it isn't, it is particularly ridiculous to see Huddy trying to grab headlines by calling Mantle what it isn't. It isn't the best API for everything as it does not implement intel or NV features. It is the best API for AMD GCN. Come on. AMD's marketing continually pushing half truths to create a good guy image, as I said, is just getting long in the tooth at this point.

 

[BrightCandle]

DX12 is developed by Microsoft but in concert with the IHVs. The IHV's basically form a committee with Microsoft to evolve DirectX going forward. Some things Microsoft dictates the IHVs implement and somethings the IHVs say are hardware they intend to add in the years to come so features they want exposed in the API. I wouldn't be surprised if the IHVs have a small number of developers who spend a lot of time with the MS developers on the driver interfaces and implementations. Its not open source, its Microsoft proprietry but all the information we have about how the process works tells us its developed like a standard.

Contrast that to AMD who intends to have sole control over the contents of the API, the features it exposes and its future direction. Sure you can implement its driver specification if your hardware can do it, but you can't be as efficient as they are and you don't get to input into it. Its made doubly worse when Intel has specifically asked for access to start developing a driver and AMD wont let them yet, its in Beta and this is actually pretty late point to be adding an additional IHV as that IHV might very well want changes to the API. That closing down of Intel is very telling, it shows us AMDs intentions matches Huddy's words on the topic - this API is controlled exclusively by AMD for AMDs purposes and they have no intention of sharing the development of it. You can implement it freely but you have no say on where it goes. This is a bad deal for the IHVs and much much worse than DirectX's process.

 

[Noctifer616]

Glad you asked. The answer is MS is implementing features for intel, AMD, and nvidia. Yes, MS is implementing "to the metal" features for all graphics vendors, not just one. NV has worked hand-in-hand with MS for NV specific features, as has AMD. When NV demonstrated DX12 beta running Forza on their hardware, they stated that MS was implementing features into DX12 which would benefit to the metal access for nvidia hardware. Intel and AMD will do the same.

But they all have the same feature support, don't they? And why isn't the 5000 and 6000 series supported by DX 12?

You are saying that DX 12 knows how to talk to AMD, Intel and NVidia hardware. But that's not the case. If that were to be true, then what is the point of the driver? Couldn't you just make the application (game/engine) talk directly to the hardware?

 

[dacostafilipe]

As long as AMD releases the code, that's all they need.

nVidia/Intel could then grad the code and implement the API in the driver without having to wait for AMD. This alone would make it possible to run BF4 with Mantle on nVidia/Intel hardware.

Now the important information that's missing, and that's a question that we should ask our "friend" Richard, is if AMD is willing to allow pull requests on the code or if it's only a dump zip/tar you can download.

If it's the latter, what would stop a dev (ex: @repi) to fork Mantle and make it the "default" one? Nothing!

PS: For the guys that do now believe that AMD releases code for anything, they just released an HSA driver for Linux here : http://www.phoronix.com/scan.php?page=news_item&px=MTczOTY

 

[BrightCandle]

Its not sufficient for AMD to release the code, unless what we have is games implementing AMD Mantle and Nvidia Mantle and Intel Mantle separately. Because each vendor would need to change the feature set to suit their cards and remove features only AMD has. So we would not have a different low level API from each which was incompatible, and presumably called different things. Its one way to go about the problem but I don't think games developers really wanting to be implementing their engines 3 times especially with the differences in features present in each card.

 

[Noctifer616]

Its not sufficient for AMD to release the code, unless what we have is games implementing AMD Mantle and Nvidia Mantle and Intel Mantle separately. Because each vendor would need to change the feature set to suit their cards and remove features only AMD has.

Isn't that why there is an abstraction layer?

 

[dacostafilipe]

Its not sufficient for AMD to release the code, unless what we have is games implementing AMD Mantle and Nvidia Mantle and Intel Mantle separately. Because each vendor would need to change the feature set to suit their cards and remove features only AMD has. So we would not have a different low level API from each which was incompatible, and presumably called different things.

Mantle does not talk directly the the hardware, there is still a driver.

If you look at the list of methods inside the Mantle DLL (here), you'll see that most of them a really "simple" stuff.

With the code, nVidia/Intel can develop a driver compatible with Mantle. How the driver processes stuff on the hardware itself is up to nVidia/Intel.

... don't think games developers really wanting to be implementing their engines 3 times especially with the differences in features present in each card.

They are. Johan said he would do it if he needs to. The guys behind Oxyde also said the same thing.

 

[Topweasel]

AMD's Mantle on the other hand, implements NOTHING except GCN features and therefore will never benefit anything except AMD hardware.

Except this has been the way since the beginning of time. DX implements features that chip manufacturers have to design their cards around. OpenGL adds features that the chip manufacturers have to design their cards around. As for AMD getting a jump on it because they are developing the spec and therefore can release it for everyone after they have already implemented in hardware is the same thing that Intel does with every USB release and Thunderbolt.

What it does do is give them a hand up in the future if Mantle is used widely is that they can always be year ahead. Having their newest cards support the newer API as it's released. But in reality that doesn't matter it'll take awhile for games to come out that actually use the latest Mantle and therefore by the time it matters Nvidia and Intel can have a couple GPUs using that version of Mantle.

This doesn't mean that they have to design GCN cards. Just as they have with DX and OpenGL before it include features that utilize the code in Mantle. People are assuming that means make a GCN chip that's older than the one AMD is creating, but that wouldn't actually be the case. On top of that if Nvidia can make a decent enough wrapper in their driver they wouldn't even need to change the design of the GPU as long as key features are supported in one way or another.

 

[BrightCandle]

We really need full api details to be sure. Seeing that list of functions raises a lot of questions for me about potential specifics of gcn and the actual paramters those functions accept. From that list Nvidia and Intel's hardware could be 100% compatible, it could also never possibly be compatible, the difference is not in the functional capabilities but the actual contracts of the functions themselves. Point remains AMD controls the APIs direction, no one else does. They have the ability to hurt their competitors if others implement it, that is why I suspect Nvidia won't.

 

[Noctifer616]

We really need full api details to be sure. Seeing that list of functions raises a lot of questions for me about potential specifics of gcn and the actual paramters those functions accept. From that list Nvidia and Intel's hardware could be 100% compatible, it could also never possibly be compatible, the difference is not in the functional capabilities but the actual contracts of the functions themselves.

Isn't that the drivers job? It takes the command from the API and translates it into something the hardware understands?

 

[Topweasel]

We really need full api details to be sure. Seeing that list of functions raises a lot of questions for me about potential specifics of gcn and the actual paramters those functions accept. From that list Nvidia and Intel's hardware could be 100% compatible, it could also never possibly be compatible, the difference is not in the functional capabilities but the actual contracts of the fun times themselves. Point remains AMD controls the APIs direction, no one else does. They have the ability to hurt their competitors if others implement it, that is why I suspect Nvidia won't.

They would be able to tell better than us. But if AMD was smart they would keep it as unlimiting as possible to help adoption and just use the fact that they can have a Mantle 4 supported card before Mantle 4 even comes out as a selling point. On the Nvidia end they are a much bigger fan of being as proprietary as they can so they might never adopt Mantle just on principal.

AMD I have a feeling would do the exact opposite. If Gsync was free they would probably implement it as soon as they could. Same thing SLI, if it didn't require the stupid Nvidia PLX chip, their boards would probably already support it. Part of that was due to the necessity of adopting MMX and SSE in the past. But no one accept maybe MS and Intel knows more about how important adoption rate is and the quickest way to do that is make it as open and easy to implement as possible.

 

[DeathReborn]

Shocking MS would claim their API better than the competition

At least Microsoft can afford to hire as many people as it takes to make it happen, AMD robably has to take staff away from other important things just for this "vanity" project.

ps: Vanity is appropriate considering the amount of PR & so called "Gaming Scientists" they've got pushing it.

 

[Topweasel]

Isn't that the drivers job? It takes the command from the API and translates it into something the hardware understands?

The point being what if the function is something that the card is incapable of without being designed in a specific way.

 

[caswow]

At least Microsoft can afford to hire as many people as it takes to make it happen, AMD robably has to take staff away from other important things just for this "vanity" project.

ps: Vanity is appropriate considering the amount of PR & so called "Gaming Scientists" they've got pushing it.

yea. almost everything ms touched in the last years failed.

 

[Noctifer616]

The point being what if the function is something that the card is incapable of without being designed in a specific way.

http://forums.overclockers.co.uk/showthread.php?t=18600973&page=4

1. Mantle is not specifically tied to Graphics Core Next, though it's obviously optimized in this direction.

2. Mantle has a meaningful, though thin, layer of abstraction that would permit it to be compatible with any modern, programmable (e.g. DX11-11.2) graphics architecture.

http://forums.overclockers.co.uk/showthread.php?t=18600973&page=5

Graphics chips are more programmable now than ever. What we could consider "full" compute and programmability (e.g. unified memory, SIMD, multi-way vector processing, dynamic scheduling, thread-level parallelism, compiler "independence") are not found on older DX11 µarches like the 6000 Series TeraScale. These shortcomings make VLIW a poor or unworkable fit for expanding hardware capabilities through new languages or APIs like Mantle. Mantle is not compatible with HD 6000 for this reason: the architecture is not programmable enough.

But Mantle would be compatible with architectures that function like GCN. As Dan Baker noted in his blog: GCN is one of the most generally programmable architectures out there. We're very proud of that, and Mantle makes good use of it.

 

[Topweasel]

I didn't say it was the case just telling you what the other guy suggested. Even in that bottom quote it does suggest that a supporting GPU has to function like GCN. That doesn't mean that it has to basically be designed as a GCN part, but where does the likeness end, how similar does it need to be? Those questions matter.

 

[BrightCandle]

One of the things Mantle does is it gets rid of the checks about the order and state of calls, DirectX is unnecessarily restrictive and Mantle is very lenient. So even if its possible to implement a particular function there is still a whole bundle of contractual rules about when it can be called, what state the parameters that called with are in etc etc. This is going to differ between the hardware, even if its stupidly simple things like the types of compression available on the textures or the support antialiasing modes.

Take Antialiasing as a simple example. AMD has its own list of AA modes available, so quite rightly they will not only expose the usual MSAA/MLAA in the interface but also their special ones. Those differ from Nvidia's, will they implement the parameters in such a way that Nvidia's options are in the list as well, or will Nvidia cards have to error when passed an AMD Antialiasing option. We can say generally the API can be supported, the problem is always the specifics. Until AMD releases the documentation I can't say whether there is a problem or not directly with support and if they have thought about the abstraction layer to the extent that it really is implementable on Nvidia hardware efficiently or not. But without information we have to be wary. I had the same query about the API back in October, it feels like a really long time with no extra information and a lot of market speak. They have the API already just release the damn thing so devs can look at it already.

Even if the API today is supported it still leaves AMD the option in the future to add something to that API which would utterly screw their competitors. We could argue AMD wouldn't, but it could. Every company when it gets in control of a market misbehaves, giving this amount of control to one company would be bad for everybody. I certainly wouldn't implement a competitors API in my field given the conditions and things they have said. I might even consider legal action for the defamation that we were not implementing it out of pride, where the issue is clearly unfair market monopoly situations caused by the lack of a standards process.

I am not saying Mantle is bad, I am saying its bad for the IHVs. Its not in their best interests to implement it. It might be in ours in the short term, but in the longer term it could be really bad for the industry.

 

[VulgarDisplay]

Glad you asked. The answer is MS is implementing features for intel, AMD, and nvidia. Yes, MS is implementing "to the metal" features for all graphics vendors, not just one. NV has worked hand-in-hand with MS for NV specific features, as has AMD. When NV demonstrated DX12 beta running Forza on their hardware, they stated that MS was implementing features into DX12 which would benefit to the metal access for nvidia hardware. Intel and AMD will do the same.

AMD's Mantle on the other hand, implements NOTHING except GCN features and therefore will never benefit anything except AMD hardware. Even in the mythical world where Mantle *did* work on everything, there would be zero percent performance improvement when AMD will not implement features for Intel or NV architectures. Aside from this, AMD informed PCPer that Mantle would be closed due to DX12 being in development which was mentioned on a prior podcast. Now suddenly AMD's marketing is going full force with the good guy BS nonsense saying that Mantle works on everything. Please.

Mantle will never implement intel or NV specific features. DX12 will and is. NV proved it with their beta DX12 drivers which were designed for DX12. If DX12 were a Mantle template, but AMD wasn't chosen to demonstrate DX12 running Forza? That's interesting isn't it? Fact of the matter is, Mantle is AMD's for AMD hardware and nothing else. It will never implement NV or Intel GPU architecture features. Period. And that is fine. I have no issue with this, this is what AMD should have been doing YEARS ago. AMD is doing what they should be doing: creating value adds to make their cards more appealing. Mantle is an appealing feature for AMD users. But just call Mantle what it is and not what it isn't, it is particularly ridiculous to see Huddy trying to grab headlines by calling Mantle what it isn't. It isn't the best API for everything as it does not implement intel or NV features. It is the best API for AMD GCN. Come on. AMD's marketing continually pushing half truths to create a good guy image, as I said, is just getting long in the tooth at this point.

I do not understand why your posting behavior has changed so completely from around a year ago. You used to contribute to discussions in a positive and worthwhile manner.

Now the only thing we see is walls of text about amd marketing. Its getting to be a waste of time reading it.

Infraction issued for thread crapping.
-- stahlhart

 

[chimaxi83]

I do not understand why your posting behavior has changed so completely from around a year ago. You used to contribute to discussions in a positive and worthwhile manner.

Now the only thing we see is walls of text about amd marketing. Its getting to be a waste of time reading it.

Yep. Don't know why anyone would waste their time reading that worthless, monotonous drivel.

 

[BFG10K]

I wonder how your facts go together with these facts.

The blog author is the same guy that said AMD's AF was "correct" on the 5xxx series, and he couldn't understand the simple concept of bilinear AF vs trilinear AF.

 

[PPB]

I do not understand why your posting behavior has changed so completely from around a year ago. You used to contribute to discussions in a positive and worthwhile manner.

Now the only thing we see is walls of text about amd marketing. Its getting to be a waste of time reading it.

Infraction issued for thread crapping.
-- stahlhart

It usually involves a squared piece of paper that lets you buy stuff IRL.

Infraction issued for thread crapping.
--stahlhart

 

[antihelten]

The blog author is the same guy that said AMD's AF was "correct" on the 5xxx series, and he couldn't understand the simple concept of bilinear AF vs trilinear AF.

And apparently he still reads the anandtech forums

http://scalibq.wordpress.com/2014/07/12/more-hate-fail/

 

[BFG10K]

And apparently he still reads the anandtech forums

http://scalibq.wordpress.com/2014/07/12/more-hate-fail/

Look what he says on his blog:

I never claimed ATi’s AF is ‘correct’ by the way.

Here's his post: http://alienbabeltech.com/main/ati-5770-image-quality-analysis/

Scali says: May 29, 2010 at 1:09 pm Ofcourse the gray area is correct for the Radeon 5770.

It isn't correct, never was, and never will be. To state otherwise reveals a shocking lack of understanding, especially when reference versions are readily available to compare.

AMD themselves admitted the implementation was flawed and changed it (mentioning it in one of the 6000 series slides), but he's still fighting the good fight on his oh-so-authoritative blog. :awe: