News - AMD Mantle API embraced by Cloud Imperium, Eidos & Oxide

[kalniel]

When people say Intel or Nvidia can implement this open standard, I have to ask if that is possible then why can none of the VLIW4 or VLIW5 cards/APUs I have at home run this?

They could have if there was one unifying architecture common to the range of AMD cards, APUs and consoles. But there wasn't, so the market share for each architecture is much smaller and the gains for the expense of creating such an API much lower. GCN is somewhat equivalent to Intel 'core' architecture - Intel have used it for several generations now, and if AMD do the same for GCN then optimising software for that architecture becomes viable.

[DanceswithUnix]

They could have if there was one unifying architecture common to the range of AMD cards, APUs and consoles. But there wasn't, so the market share for each architecture is much smaller and the gains for the expense of creating such an API much lower. GCN is somewhat equivalent to Intel 'core' architecture - Intel have used it for several generations now, and if AMD do the same for GCN then optimising software for that architecture becomes viable.

Valve claimed a 17% speedup in Left 4 Dead by porting to Linux and OpenGL, despite only having a 32 bit Linux implementation so it should have been faster than that. http://blogs.valvesoftware.com/linux/faster-zombies/

I think that shows that APIs are important, but how much is left to be gained? "Core" architecture is an interesting example, the ideas and basic layout go back to the Pentium Pro which is now a very old device. Just within that one architecture, you have P-Pro, P-II, P3, Pentium Mobile, Core-2 and Core-2 with threading enabled to optimise for. The list of x86/AMD64 targets for GCC makes an interesting read, because if you really care about performance then you have to compile for the exact machine you are going to run on: http://gcc.gnu.org/onlinedocs/gcc/i3...4-Options.html

If you optimise for the wrong target, then your low level interface hurts rather than helps, which is why we have high level interfaces.

Adding a bit for the extra registers you get in 64 bit to the 17% given above, Mantle had better get over 20% improvement or you would be better off just optimising for Linux and staying high level.

[kalniel]

GCC is not optimising a target where there is such an obvious bottleneck though. The limited call lists mean CPUs are being hobbled at the moment and hence you see single threaded performance dictate GPU speed to some extent on PC, compared to consoles. Windows 8.1 and the latest DirectX versions help this a little, nVidia have some features in their cards that also help this a little (with support within DirectX), but even combined they don't allow for anything like the speed up that we're likely to see using Mantle where CPUs were bottlenecking the calls to GPU. This is of huge importance to AMD where they have lots of cores waiting to be used more efficiently (and a relative weak single threaded performance compared to Intel).

[shaithis]

I see it as a step backwards, back to a time when games could use D3D, OpenGL or Glide.....yet for the most part only one of them worked properly.

What we need is Microsoft to rip out some of the APIs and provide better tools for developers so that multiple layers of API are not required.....I do feel that all the APIs they created helped but maybe we are moving into an area where developers do not need that crutch any more.

[DanceswithUnix]

I see it as a step backwards, back to a time when games could use D3D, OpenGL or Glide.....yet for the most part only one of them worked properly.

What we need is Microsoft to rip out some of the APIs and provide better tools for developers so that multiple layers of API are not required.....I do feel that all the APIs they created helped but maybe we are moving into an area where developers do not need that crutch any more.

Graphics APIs (any APIs) aren't a crutch, they are a needed abstraction so that you can optimize your bit and let the graphics people optimize their bit. The overall result won't be the fastest possible, but in a world of changing hardware it should be close enough. Glide was an oddity, created ISTR because OpenGL was too hard to put together in a hurry. It was replaced by Mini-GL, then OpenGL, then Microsoft had to put their oar in and step up back half a decade. That is what gripes me here, DirectX isn't an API it is a means of Microsoft controlling things. A lock in tool. If I thought Mantle could throw off those shackles then I would be more supportive, but my spider senses are telling me that it will be a small boost for a subset of Windows users which Microsoft are going to try and shut down for "security" or some such excuse.

Standards: http://xkcd.com/927/

I seem to be getting less tolerant in my old age

But yeah, it feels like a step backwards.

[Kumagoro]

I would have thought that it would not be worth AMD making pre gcn cards work with mantle. Makes better business sense to save your resources and get people to buy your new stuff.

[crossy]

I see it as a step backwards, back to a time when games could use D3D, OpenGL or Glide.....yet for the most part only one of them worked properly.

What we need is Microsoft to rip out some of the APIs and provide better tools for developers so that multiple layers of API are not required.....I do feel that all the APIs they created helped but maybe we are moving into an area where developers do not need that crutch any more.

I too remember the "shall I use the D3D or OpenGL versions?" question with a lack of enthusiasm. Then again, I also remember getting Volume 1 of the SGI OpenGL developers reference set and being overwhelmed with how unwieldy it all looked. ButI'd come from a Tektronix 3D programming background - where their API was pretty low level, but also correspondingly quick. And if you needed that last erg of performance there was actually a base level API - think graphics assembler code. I guess Mantle is akin to that low-level-but-not-assembler level.

I've always had a soft spot for OpenGL over D3D/DX simply because the former is (supposed to be?) cross platform, whereas the latter is a Windows only gig. And cross-platform ability is pretty important these days of gaming on PC's, consoles, tablets and smartphones. Then again, as someone said to me, "all these developers just use toolkits like Frostbyte etc. So as long as those toolkits support whatever you need then does it matter if you're using DX, OpenGL or monkey poop thrown at the screen!"

Don't think I want to be on the committee that's reviewing the monkey poop standard.

[DanceswithUnix]

I would have thought that it would not be worth AMD making pre gcn cards work with mantle. Makes better business sense to save your resources and get people to buy your new stuff.

Which is all very well, but my daughter's quad core APU *is* new stuff. They still sell them. AMD is the dominant graphics vendor in our house, yet supported hardware is... zero.

I don't expect support for the 4670, that is ancient in computing terms. The 6770 cards are about two years old, so I would grudgingly let them off. The Llano APU and the Brazos APU are similarly getting on a bit, but the Trinity? You can still buy those, it is the same tech as Richland.

I noticed recently that Nvidia have what looks like a stupid cutoff in when they support recording from a graphics card in their windows app and my 460 is not supported despite being almost the same as a 5xx series which are supported. So yeah, I expect some arbitrary cutoff in support as most companies do that and you can't test every combination of hardware, but either AMD just aren't trying here or their fancy new API is so hardware specific that it is basically a broken hack.

[kalniel]

So yeah, I expect some arbitrary cutoff in support as most companies do that and you can't test every combination of hardware, but either AMD just aren't trying here or their fancy new API is so hardware specific that it is basically a broken hack.

It's your usual cost/benefit thing. How many VLIW4 products are they selling to gamers verses GCN? Given consoles add to the GCN count, I think it'd be quite bad business to spend the extra cost on the minority market. Particularly if the gain in performance is coming from people coding against GCN on consoles in the first place.

[kingpotnoodle]

It is great to see this getting support. I would like this to be manufacturer agnostic though so all gamers can benefit from it rather than just AMD customers.

Which is what DirectX and OpenGL are - a manufacturer agnostic API. The complaint is that they create overheads to accomplish this, and that is true, abstraction layers and APIs always will its just that hardware specific ones can be a bit leaner. However I think what AMD are doing is worse - vendor lock-in via proprietary APIs, this is what I thought had disappeared with 3Dfx/Glide and the bad old days before DirectX. IMHO it's a huge backwards step and not to be celebrated in the slightest, a better solution is to collaboratively improve the common standards. Game developers will have to make games to support Mantle to appease AMD owners, DirectX for everyone else, it's like the early days all over. Then you start getting exclusive games and suddenly PCs have degenerated into a console-like fragmented mess.

Whilst we're on the point NVidia do the same thing proprietary thing with CUDA but it's a professional product and not so high profile. Intel with Xeon Phi focussed more on a standards compliant product support OpenCL and Fortran++ etc - which was nice of them, but it's not as fast from what I've seen.

[CAT-THE-FIFTH]

I hope it does well,so it give a kick up the arse of MS,to actually bother with DX. It seems as the years past,they are doing less and less,despite how many devs have complained about the state of DX development and its inefficiencies(I have a GTX660 so are not going to benefit from Mantle ATM). Also,regarding Mantle,AMD has stated they won't stop anyone from modding it for their GPUs. Its the same with all the effects they have done in the last few years,even the recent physics effects in TR,which work fine on my GTX660.
Nvidia OTH has just gone out of their way to block people out on purpose,even if work arounds have been found to run stuff they help develop,which has already lead to fragmentation in many markets. We should have had open standards for game physics and 3D gaming by now. Also,as AMD stated devs talked to them not the other way around,so its funny how people just ignore this conveniently. MS should take note of this.

[scaryjim]

The 6770 cards are about two years old, so I would grudgingly let them off.

But the 6770 was a direct rebrand of 5770, which is now over 4 years old. So hold the grudge about the rebrand, but not about the new feature support

... I expect some arbitrary cutoff in support as most companies do that and you can't test every combination of hardware, but either AMD just aren't trying here or their fancy new API is so hardware specific that it is basically a broken hack.

Erm ... the switch from VLIW to GCN was a pretty major revision to how AMD graphics worked, as I understand it - shifting more towards nvidia's shader architecture IIRC. Sure, Mantle might be so hardware-specific that making it an open API is a shallow gesture on AMDs part ("look how open we are! What, you can't implement it? Shame, but at least we're open!"), but if there's a difference between VLIW and GCN that stops Mantle being backwards compatible, that might demonstrate it's more likely that nvidia will be able to implement it, not less (and perhaps they could even roll it back to older cards, although I doubt they would).

[CAT-THE-FIFTH]

The thing is I can see Mantle being most useful for AMD APUs,as this is the area I think the increased efficiency will be felt the most.

Also,Crossfire should work better with Kaveri too,as the GPU cores of some of the discrete cards will have the same shader design and number of cores as the Kaveri IGP.

[DanceswithUnix]

But the 6770 was a direct rebrand of 5770, which is now over 4 years old. So hold the grudge about the rebrand, but not about the new feature support


Erm ... the switch from VLIW to GCN was a pretty major revision to how AMD graphics worked, as I understand it - shifting more towards nvidia's shader architecture IIRC. Sure, Mantle might be so hardware-specific that making it an open API is a shallow gesture on AMDs part ("look how open we are! What, you can't implement it? Shame, but at least we're open!"), but if there's a difference between VLIW and GCN that stops Mantle being backwards compatible, that might demonstrate it's more likely that nvidia will be able to implement it, not less (and perhaps they could even roll it back to older cards, although I doubt they would).

I have no problem with rebrands as such. You want a level of performance for a certain price, you pay the money, you expect driver support for a few years and then you move on. What a company do under the hood is fairly irrelevant for most people. They were still selling that part a generation on as 76xx too. A high level API makes this possible.

As for my arbitrary cutoff comment, at some point AMD will move on from GCN, just like they moved on from each instruction set they had before it. If Mantle breaks at that point and the new cards are slower than the old cards, how stupid will that look? If this is a hack that depends on GCN shader scripts, then it *will* break.

[scaryjim]

I have no problem with rebrands as such. You want a level of performance for a certain price, you pay the money, you expect driver support for a few years and then you move on. What a company do under the hood is fairly irrelevant for most people. They were still selling that part a generation on as 76xx too. A high level API makes this possible.

true, but then you're judging support for a card on the basis that it's less than 2 years old, when it's really more than 4. It's (IMNSHO, at least) one of the biggest issues with rebrands: unless you're going to be as arbitrary as NV (who seem to ditch support almost at random sometimes), you'll hit a point where it looks like you're ditching support for a relatively new card, because you're actually ditching support for an old architecture that happens to have been used in newer products. As to the high level API, it's still there and people will still use it. I don't think we even know how extensive the Mantle API is: it could be a tiny subset focusing on just those calls that will benefit most from a low-level API.

... at some point AMD will move on from GCN, just like they moved on from each instruction set they had before it. If Mantle breaks at that point and the new cards are slower than the old cards, how stupid will that look? If this is a hack that depends on GCN shader scripts, then it *will* break.

I agree that *IF* all that happens, AMD will look pretty dim. VLIW (in various revisions and at various market points) lasted them the best part of 7 years though, which is coincidentally (iirc) the average console refresh period. So if GCN lasts a similar length of time, AMD will be looking to move on from GCN around the same time that MS and Sony start looking for the next architecture for new consoles. Convenient, n'est pas? Plus if I was the company who created a new API, I'd be pretty determined to make sure that any major changes to my architecture would still allow me to support that API...

Of course, this is all conjecture without knowing how close to bare-metal Mantle is running, and how many graphics API calls it covers. And I suspect it'll be a while before we get the full picture on that.

[tribaljet]

Which is what DirectX and OpenGL are - a manufacturer agnostic API. The complaint is that they create overheads to accomplish this, and that is true, abstraction layers and APIs always will its just that hardware specific ones can be a bit leaner. However I think what AMD are doing is worse - vendor lock-in via proprietary APIs, this is what I thought had disappeared with 3Dfx/Glide and the bad old days before DirectX. IMHO it's a huge backwards step and not to be celebrated in the slightest, a better solution is to collaboratively improve the common standards. Game developers will have to make games to support Mantle to appease AMD owners, DirectX for everyone else, it's like the early days all over. Then you start getting exclusive games and suddenly PCs have degenerated into a console-like fragmented mess.

Whilst we're on the point NVidia do the same thing proprietary thing with CUDA but it's a professional product and not so high profile. Intel with Xeon Phi focussed more on a standards compliant product support OpenCL and Fortran++ etc - which was nice of them, but it's not as fast from what I've seen.

So much truth there. It was great that Glide did end with Nvidia buying 3dfx. For a time there were quite good developments on the API side of things, but unfortunately the forced push to DirectX did win over OpenGL, which is IMHO the superior API, and crossplatform at that.

The only valid use for Mantle is to, as CAT-THE-FIFTH said, give a swift kick to Microsoft in order for it to do some palpable improvements to DirectX. Other than that, it can't die soon enough.