Issues with CUDA

[AnnoyedDragon]

Considering the news Nvidia is going to redesign the PhysX API to use CUDA for physics acceleration on the GPU, why isn’t anyone concerned?

Based on what I understand about CUDA:

A) It will not work on ATI cards.

B) Most of the architectural needs for CUDA only exist in the Geforce 8 series.

So C) it will not work on the vast majority of gaming machines in the market; including any ATI GPU based computer, Nvidia Geforce 7 and other earlier series and none of the next generation console.

So I ask people here, unless I got something wrong, what developer in their right mind would use the GPU based PhysX API? You can say what you want about the PPU, but at least it worked on any gaming computer regardless of your CPU/GPU configuration.

A game that uses Nvidia’s method for GPU game play physics will only run on an Nvidia 8 series card or later, that is a very small audience to limit your game to. It actually makes multi core CPU physics look more desirable, despite being lower end and slower to progress at least it will run on anything.

 

[zehoo]

http://www.xbitlabs.com/articles/editorial/display/ageia-acquisition-interview-2008_4.html#sect0

X-bit labs: What do you actually mean by open platform initiatives here? Does it mean that Nvidia will keep developing PhysX SDK and it will support x86 microprocessors, Microsoft Xbox 360, Nintendo Wii and Sony PlayStation 3? Or does it mean that PhysX SDK will be compatible with all of the above as well as DirectX 10 GPUs in general?

Michael Steele: It means that Nvidia wants to be open in its further development of PhysX software in order to enable as many platforms as possible on both PC and console. Nvidia understands better than anyone the need to enable a large and unrestricted market for game developers. I don’t think anyone will question their unique capabilities to support the development community. Enabling a physics eco-system and then delivering the best hardware for that environment is good for gaming.

 

[Scali2]

Basically they'll just add Cuda support, giving GF8-users incredible physics performance. Nothing changes for the rest.
PhysX never supported ATi, and probably never will now. Then again, ATi is nowhere near where nVidia is with Cuda at this point, so I wonder when we will ever see any kind of working ATi physics solution at all.

I'd just like to see if Cuda can match or even outperform a PhysX PPU, and how much it will affect graphics performance, because I'm not convinced on this whole GPU-physics thing. Firstly, I don't think a GPU design is flexible enough for efficient physics (although GF8 has some nice non-GPU features specifically for Cuda), and secondly, you can only spend your GPU cycles once, so if you do get efficient physics, does this mean your framerate will drop because you are losing out on graphics shader performance?

 

[mrsmith076]

I don't think that right now it's major enough for ATI consumers worry about. The game support wasn't terrific, and it only made it into a couple really good games. With the Nvidia acquisition, I think that it will just enlarge the amount of things that we seriously consider for gaming in different levels of graphics cards. Right now with all the current benchmarks in ATI cards beat Nvidia cards with AA ratcheted all the way up. Maybe Nvidia will try to gain the crowd back by having marginally better physics.

This isn't reason enough to buy a certain brand of graphics card yet. If anything I'm just excited to see a small company get attention by a big player.

 

[Atech]

Basically they'll just add Cuda support, giving GF8-users incredible physics performance. ?

That remains very much to be seen, and in a age of GPU-limited games, GPU-physics makes no sense to me...

 

[Scali2]

That remains very much to be seen, and in a age of GPU-limited games, GPU-physics makes no sense to me...

That was the second part of my post...

 

[fluxion]

people can make games for 8 series GPU's, because they can reasonably expect such hardware to be in place. with something like this in place, they can use advanced physics rather than refraining due to the CPU's lack of performance in these areas. games are gpu-bound, but they're also ppu-bound in the developer context, that just doesnt carry over to us as prevalently because such interactions are just cut.

people cannot however invest in in-game physics that necessitates the use of PPU's, the market for them is too small. this is a better solution in the long-run.

for some developers this reasonable expectation of physics processing capability on a user's system is more important than them being able to run at 100x AA/AF and super-ultra-high-mega-quality.

the underhanded side to it is that this API will undoubtedly be different from ATI's solution, so the above point is moot since no developer would depend solely on nvidia hardware being in place. and writing gp^2 code for both is expensive.

so somethings gotta change. preferably by both adopting a standardized physics api

 

[jbzx86]

Nvidia has already stated they plan on incorporating some of the hardware aspects of the PhysX hardware into their GPUs. Already the 8XXX hardware is also viable as GPGPU. I think in the future, GPUs will be able to handle high-end graphics and physics with no hit to framerate due to the implementation of physics.

 

[imzjustplayin]

people can make games for 8 series GPU's, because they can reasonably expect such hardware to be in place. with something like this in place, they can use advanced physics rather than refraining due to the CPU's lack of performance in these areas. games are gpu-bound, but they're also ppu-bound in the developer context, that just doesnt carry over to us as prevalently because such interactions are just cut.

people cannot however invest in in-game physics that necessitates the use of PPU's, the market for them is too small. this is a better solution in the long-run.

for some developers this reasonable expectation of physics processing capability on a user's system is more important than them being able to run at 100x AA/AF and super-ultra-high-mega-quality.

the underhanded side to it is that this API will undoubtedly be different from ATI's solution, so the above point is moot since no developer would depend solely on nvidia hardware being in place. and writing gp^2 code for both is expensive.

so somethings gotta change. preferably by both adopting a standardized physics api

You have it completely backwards. Ageia was impartial, it didn't care what video card you had. Now that nvidia owns ageia, the only way to get this "physics" is to own an nvidia card which is total bullshit to me. Nvidia bought ageia because they were pissed off that havokFX failed and so instead of competing with ageia, they decided to buy them out, stupid assholes.

I'm sorry but I'm just really angry that nvidia has destroyed what could've been the next big thing for gaming.

 

[InorganicMatter]

A) It will not work on ATI cards.

And? It didn't work on ATI before, why should it work on ATI now? PhysX acceleration is simply being added to NVIDIA cards. It's a free bonus to go with your $300 card, you should be thankful that NVIDIA's not adding it in the future and making you buy a 9000-series card when the time comes.

B) Most of the architectural needs for CUDA only exist in the Geforce 8 series.

So you're OK with wasting $300 (yes, I said it: wasting) on a PPU that only works in a few games instead of spending $300 on the best graphic card series of all time, which will heavily boost your gameplay experience in all games? Come one, be realistic.

So C) it will not work on the vast majority of gaming machines in the market; including any ATI GPU based computer, Nvidia Geforce 7 and other earlier series and none of the next generation console.

Are you going to complain next that the 9800Pro doesn't support DirectX 10?

So I ask people here, unless I got something wrong, what developer in their right mind would use the GPU based PhysX API?

The same developers who have been making games "The Way It's Meant to be Played" since the beginning of time. With the exception of Valve, almost every developer out there makes their games TWIMTBP and prominently displays that logo. I imagine PhysX acceleration will become a new requirement for bearing a TWIMTBP logo. I also imagine ATI will be required to cough up licensing money for PhysX acceleration, or be left behind, just like AMD has to pay through the nose to license x86 from Intel. You also seem to forget that PhysX is simply a software API that's accelerated by hardware. It will still be able to run on ATI computers, it will simply run slower due to no hardware acceleration.

 

[pxc]

With the exception of Valve, almost every developer out there makes their games TWIMTBP

It seems Valve did join TWIMTBP. http://www.steampowered.com/nvidia/ The logo is in the bottom right corner.

 

[mavalpha]

Everyone crying about GeForce7 and below being neglected, you're overlooking a key issue.

The 8000 series is already 1.5 years old. Now that developers have more reason to incorporate PhysX support into their games, expect new games and engines to emerge with this support. But realize that it will take at least six months (very optimistically) for projects started today to come to fruition. By that time, the 8800 family will have been around for almost 2.5 years, and the "budget-friendly" 8800GT a year by itself. Anyone still using a 7900 or below, will be trying to play on what will then be 3-year-old hardware. Feel like trying Crysis on your 6800? If it runs at all, you certainly won't be complaining that it doesn't work with all the eye-candy turned on; PhysX support in GeForce8 will be a perk for systems that can handle it, like 16x AF is today.

 

[Scali2]

Why are people even arguing about the GeForce 7 series?
The whole Cuda thing is one of the key features of the GeForce 8 design. If you bother to look at some presentations at http://developer.nvidia.com about the architecture of the GF8, you'll see that it has some specific buffers and cache forwarding hardware that has no use for graphics, and is strictly for improving Cuda/GPGPU performance.

Cuda cannot run on GeForce 7 because the hardware simply wasn't designed for it. Therefore, GF7 also isn't very interesting for GPU-accelerated physics in general.

I'm not sure where ATi stands in this exactly. ATi obviously does not have a powerful and easy-to-use SDK for their GPUs like Cuda. ATi may or may not have a similar architecture to the GF8, which would improve GPGPU/physics performance, but then again, they could aswell just have designed a 'straight' DX10 architecture, with no special GPGPU features. In which case, you won't be missing out much by not having Cuda and PhysX support.

 

[AnnoyedDragon]

Judging the Valve hardware survey, you do know that off the bat just over 38% of people cannot use CUDA? That includes ATI and the ‘other’ categories. The remainder is then divided again into the percentage of Nvidia users who own 8 series, which is just under 18% if I got the math right.

You can debate how reliable the sample audience of Valves survey is, but they do have a diverse audience.

Add that despite Nvidia talking about an ‘open’ solution that CUDA does not work on any of the consoles, why would developers support it for reasons other than throwing in a few additional pretty ‘effects’ if you can run it?

Would you design a game only 18% of the market can play? If they have the hardware performance that is, only 9.37% owned an 8800, which is slightly better than the number of DX10 ready machines.

 

[Scali2]

You have to separate Cuda from PhysX.
PhysX is a multiplatform physics API which is already widely in use on PC with multithreaded CPU support, their own PhysX PPU, and on various consoles.
nVidia will *add* Cuda support to this API. So it will still work with CPU, PPU and on consoles.
So PhysX can already cater for 100% of the market (PC + consoles), just not everyone will be able to have hardware acceleration on PC. Only people with a PPU or a Cuda-compatible GPU.

 

[imzjustplayin]

You have to separate Cuda from PhysX.
PhysX is a multiplatform physics API which is already widely in use on PC with multithreaded CPU support, their own PhysX PPU, and on various consoles.
nVidia will *add* Cuda support to this API. So it will still work with CPU, PPU and on consoles.
So PhysX can already cater for 100% of the market (PC + consoles), just not everyone will be able to have hardware acceleration on PC. Only people with a PPU or a Cuda-compatible GPU.

Yeah but since nvidia was so eager to add physx support to their GPUs which I'm pretty sure is only going to be good for second order physics, it probably will mean that nvidia will scrap the separate PPU just to tout their stupid video cards.

 

[InorganicMatter]

it probably will mean that nvidia will scrap the separate PPU just to tout their stupid video cards.

I just don't understand the hate and resentment NVIDIA is getting for this. Thier GeForce 8 cards are second to none, and their stream processing architecture is perfect for something like physics calculations. It's too bad ATI didn't get in on this (like they could afford it), but I'm really glad to see NVIDIA doing something with their great stream processing and CUDA architectures.

 

[Nenu]

I just don't understand the hate and resentment NVIDIA is getting for this. Thier GeForce 8 cards are second to none, and their stream processing architecture is perfect for something like physics calculations. It's too bad ATI didn't get in on this (like they could afford it), but I'm really glad to see NVIDIA doing something with their great stream processing and CUDA architectures.

+1

 

[CountChoculaBot]

This is a *very* good thing IMO. PhysX on the 8-series will give it the user base it never had, meaning it actually has a chance of being used by dev's now on their games. I'd gladly take somewhat worse graphics in exchange for physics processing, too.

 

[imzjustplayin]

This is a *very* good thing IMO. PhysX on the 8-series will give it the user base it never had, meaning it actually has a chance of being used by dev's now on their games. I'd gladly take somewhat worse graphics in exchange for physics processing, too.

Yes but now all we'll get is second order physics which adds nothing useful to the game, not to mention video cards are STILL gpu limited in games and using them to do ANOTHER task is only going to make their current job that much more difficult. I'd much rather have an add-in card than have to deal with half assed physics that adds absolutely NOTHING to the game.

great, MORE PARTICLE EFFECTS!!!!!!

 

[AnnoyedDragon]

And? It didn't work on ATI before, why should it work on ATI now? PhysX acceleration is simply being added to NVIDIA cards. It's a free bonus to go with your $300 card, you should be thankful that NVIDIA's not adding it in the future and making you buy a 9000-series card when the time comes.

Umm, it did work on ATI cards, what are you on about? Ageia's PPU solution worked along side any GPU/CPU combo you wanted; while as Nvidia's only on a single series within their brand, greatly locking down what was previously a very open physics solution.

So you're OK with wasting $300 (yes, I said it: wasting) on a PPU that only works in a few games instead of spending $300 on the best graphic card series of all time, which will heavily boost your gameplay experience in all games? Come one, be realistic.

What on earth does any of that have to do with what I said? PPUs haven't cost $300 for a 'very' long time; my comment was referring to CUDA being limited to the 8 series, I didn't say anything on PPU cost Vs GPU costs and its value in gaming. You are also being very optimistic about a yet to be implemented solution blowing away Ageia's previous one, how many people do you know have $300 to throw on a secondary GPU just for physics? And why would developers make games that will only be accessible to these people? You target that performance level for game play and it becomes a mandatory requirement. Whatever Nvidia plans to do with game play physics, it has to target the lowest common denominator, which is not a secondary $300 GPU.

Are you going to complain next that the 9800Pro doesn't support DirectX 10?

... Ok again what relevance does this have with me stating Nvidia's solution will only work on a very limited number of cards? Supporting DX10 in a game doesn't make it unplayable for people who don't have DX10 hardware; DX10 is also available on either Nvidia or ATI cards and is not brand biased. How is a hardware physics solution, which has a game play impact, in any way similar to something that only improves visual effects? Unless of course you consider Nvidia's solution to only offer physics eye candy.

The same developers who have been making games "The Way It's Meant to be Played" since the beginning of time. With the exception of Valve, almost every developer out there makes their games TWIMTBP and prominently displays that logo. I imagine PhysX acceleration will become a new requirement for bearing a TWIMTBP logo. I also imagine ATI will be required to cough up licensing money for PhysX acceleration, or be left behind, just like AMD has to pay through the nose to license x86 from Intel. You also seem to forget that PhysX is simply a software API that's accelerated by hardware. It will still be able to run on ATI computers, it will simply run slower due to no hardware acceleration.

Consumer choice isn't an issue for you is it? What if Nvidia refuses to license CUDA to ATI in an effort to kill them hmm? Or better yet, what if ATI pulls a 360? Purposly not support the technology and become the lowest common denominator in game development, damaging the advantages of a competitor by having their tech unutilised?

Don't expect things to go as smoothly as you hope them to.

 

[TerranUp16]

Judging the Valve hardware survey, you do know that off the bat just over 38% of people cannot use CUDA? That includes ATI and the ‘other’ categories. The remainder is then divided again into the percentage of Nvidia users who own 8 series, which is just under 18% if I got the math right.

You can debate how reliable the sample audience of Valves survey is, but they do have a diverse audience.

Add that despite Nvidia talking about an ‘open’ solution that CUDA does not work on any of the consoles, why would developers support it for reasons other than throwing in a few additional pretty ‘effects’ if you can run it?

Would you design a game only 18% of the market can play? If they have the hardware performance that is, only 9.37% owned an 8800, which is slightly better than the number of DX10 ready machines.

I'm assuming you have Crysis (or, hell, you don't even need Crysis, just some game with this support). Open the options panel. Go to the Advanced panel for graphics. Look at the options. Hm, is that PHYSICS I see? Ahah, it is! And look! You can tune it all the way from Low to Very High!

No, the physics that CUDA deals with are not likely to affect gameplay for quite some time (but then, look at Crysis again- you could see developers putting in modes that cater to those who have the hardware; and even if it's only 18% at first, usually over time that number grows). But physics really is another form of graphics at the moment- they can provide visual immersion. And yes, you can utilize physics that do not affect gameplay but that do add to visual immersion. So, initially, that's largely what you're going to see CUDA cater to- hardware accelerated physics will just be one more graphical setting, but you're still going to have the processor plugging away at the stuff that's vital to gameplay. Eventually, as more and more people have set-ups that support CUDA-accelerated physics, you will see more and more games take advantage of the gameplay potential they allow.

And for that I feel nVidia is in a race with Intel, because Havok is of course the other entity in all of this, and the closer Intel gets to getting 8 and 16 core cpu's (yes, a little ways off, but not too far- this only took ~3 years for the dual cores to do once they were introduced) as "the standard" (much as dual cores are increasingly becoming now- among the min specs for quite a few modern games is a dual core cpu). Now, the other interesting thing in this mix is ray tracing. Intel is plopping quite a bit of R+D into getting ray tracing to work on 8 and 16 core cpu's. I don't see how those cpu's will be able to handle ray tracing, AI, gameplay, and complex physics- something's gotta give there. Therefore, you may well see CUDA come in and allow nVidia to remain relevant despite graphics then being taken care of by the processor (that's an extreme scenario that suggests that nVidia and ATi will be unable to do a better job of adapting ray tracing to graphical hardware acceleration than Intel can adapt it for multi-core cpu's).

As far as the GeForce 7 series and anything below goes... they don't have stream processors, but rather have specialized shader processors and the like. The whole idea of the stream processors was to create a unified graphical architecture that could be highly efficient with graphics (doing ok with pixels but need some more shading power? With a GeForce 7 or below, you were out of luck. With the GeForce 8, it's a simple matter of instructing a few stream processors to change what they're doing) and which also could do more than just graphics (physics was, has been, and is the primary target of this secondary functionality).

In terms of a 6800 in Crysis, my friend used my old eVGA 6800 Ultra and managed to get it to run playably at Medium settings at 1280x1024. I'm assuming that's only for the first half of the game, but not to be for a card that is pretty much the minimum spec.

 

[AnnoyedDragon]

Preaching to the converted TerranUP16, I know this whole GPGPU thing is GPU companies gut reaction to CPU companies entering their territory. Why not? If ray tracing runs better on CPUs than GPUs, even if it is well down the line it is still an issue they will have to encounter at some point. Not forgetting the hybrid processors just around the corner.

I’ve actually used Crysis in a few examples myself; for instance if you set physics to the lowest setting, which is the default one online, you will notice you can now no longer destroy buildings. This offers a game play advantage to gamers with more powerful hardware, which most developers do their best to avoid.

Game play has to target the lowest common denominators; this applies to all platforms even consoles. If you develop a game that targets the PS3 cell processor to it’s maximum it won’t work on 360, but if you fully utilise the 360s tri core it will still work on the PS3, hence the weakest common denominator is targeted. Anything above that cannot impact game play, so they leave it to visual effects.

Pretty much every setting above medium in Crysis will just have a visual impact.

Now as stated this can apply to the case with CUDA, but it isn’t that simple. The problem with CUDA, as I have stated many times, is it is entirely limited to current and future products under the Nvidia brand. Now Nvidia fanboys that think they cannot do wrong may not have a problem with this, but it matters to developers quite a bit.

CUDA is going to remain nothing better than Havok FX as long as the compatibility issues stay in place, even if the Geforce 9 or 10 comes out. Nvidia has two choices; they can either licence CUDA to ATI and partially remedy the problem (they still will only use it for effects in cross platform games); or they can keep it to themselves and try to become a monopolistic bastard, probably killing CUDA physics in the process because no dev in their right mind would limit their game to only one brand of hardware.

 

[TerranUp16]

It will be interesting to see what they do as far as CUDA goes. They probably will, ultimately, extend the offer to ATi, but I'm guessing ATi is going to have to pay quite a bit for it and are going to have to work on the implementation themselves (more $). In the short-term, that likely limits CUDA to just providing another graphical option (or special modes in a few games, such as Crysis's high-end mode and the UT3 Aegia mods), but in the long-term it would allow the format to provide superior gameplay physics.

In terms of ATi rejecting the offer or nVidia not extending it, if the specter of Intel is as great as we surmise it to potentially be, I feel that both nVidia and ATi will realize that unity on this issue will provide the best path (therefore, gpu-accelerated physics can make the headway they need to towards becoming standard and superior, but nVidia and ATi can still battle over who proffers the better solution).

 

[Scali2]

Yes but now all we'll get is second order physics which adds nothing useful to the game, not to mention video cards are STILL gpu limited in games and using them to do ANOTHER task is only going to make their current job that much more difficult. I'd much rather have an add-in card than have to deal with half assed physics that adds absolutely NOTHING to the game.

great, MORE PARTICLE EFFECTS!!!!!!

Excuse me, but don't confuse Havok FX with Cuda. Cuda is a GPGPU framework and can do any kind of physics you want. It's aimed at a newer generation of hardware than Havok FX was.

 

[Scali2]

Now, the other interesting thing in this mix is ray tracing. Intel is plopping quite a bit of R+D into getting ray tracing to work on 8 and 16 core cpu's. I don't see how those cpu's will be able to handle ray tracing, AI, gameplay, and complex physics- something's gotta give there. Therefore, you may well see CUDA come in and allow nVidia to remain relevant despite graphics then being taken care of by the processor (that's an extreme scenario that suggests that nVidia and ATi will be unable to do a better job of adapting ray tracing to graphical hardware acceleration than Intel can adapt it for multi-core cpu's).

It's not likely that raytracing will ever replace rasterizing completely. Raytracing isn't very efficient, nor is it very flexible (things like texture filtering, mesh animations and high-order surfaces are hard to do with a raytracer, need a lot of processing). Even with offline rendering systems such as Pixar's Renderman, raytracing isn't the main method of rendering. They use an alternative rasterizing method by the name of REYES, which basically subdivides geometry until you get micropolygons that are smaller than a pixel. These are then accumulated in a buffer for antialiasing.
Raytracing is only used on certain effects, like refracting glass or so.
In fact, they often don't WANT physically correct effects, because they are harder to control. They will instead place fake lightsources on places where they want certain lights or shadows. Much like how they work on a real movie set.

I just don't know why raytracing is such a buzzword to most people.
The most logical solution currently is to add raytracing features to current shaders... This is already possible in a limited way, and DX11 should make it more powerful still (and also add features for better GPGPU, so it could help ATi with physics).
Then the rasterizers will get better subdivision surface functionality (the geometry shader was a simple first step), and will more or less evolve into REYES renderers over time.
Then you'll have a realtime Renderman-like setup, which would literally give you Hollywood-quality graphics. Something that pure raytracing has never been able to, and never will. Renderman is the benchmark by which the realism of all renderers is measured in the industry.

 

[Nenu]

Excuse me, but don't confuse Havok FX with Cuda. Cuda is a GPGPU framework and can do any kind of physics you want. It's aimed at a newer generation of hardware than Havok FX was.

Yes, and with no 'slower' buses in the signal paths between PPU processors and GPU graphics hardware, many of the limitations of a standalone PPU will be vastly reduced.
ie Once the data is given to the PPU (that part of the GPU set aside), it doesnt need to pass any more data along the PCI or PCI-E buses before handing the results to the GPU for video processing.

This also reduces traffic on the PCI-E (and PCI) bus allowing other hardware to function with less increased latency while the PPU is in use.

The Graphics card solution is significantly better in many ways.

 

[Scali2]

I’ve actually used Crysis in a few examples myself; for instance if you set physics to the lowest setting, which is the default one online, you will notice you can now no longer destroy buildings. This offers a game play advantage to gamers with more powerful hardware, which most developers do their best to avoid.

Well, I believe that Crysis won't allow you to change the physics level in an online game.
There's also the difference between DX9 and DX10-mode, where in DX10-mode there's more physics, like trees that can be destroyed and such.

Now as stated this can apply to the case with CUDA, but it isn’t that simple. The problem with CUDA, as I have stated many times, is it is entirely limited to current and future products under the Nvidia brand. Now Nvidia fanboys that think they cannot do wrong may not have a problem with this, but it matters to developers quite a bit.

CUDA is going to remain nothing better than Havok FX as long as the compatibility issues stay in place, even if the Geforce 9 or 10 comes out. Nvidia has two choices; they can either licence CUDA to ATI and partially remedy the problem (they still will only use it for effects in cross platform games); or they can keep it to themselves and try to become a monopolistic bastard, probably killing CUDA physics in the process because no dev in their right mind would limit their game to only one brand of hardware.

I wouldn't be too sure of that. Firstly, nVidia has a very large marketshare (especially in the game-buying market), and over time this entire marketshare will have GF8+ GPUs that are capable of Cuda-accelerated physics. It could be over 50% of all gamers (62% of all Steam-users have nVidia hardware, only 31% uses ATi... The 3dmark06 stats are even more interesting: http://www.yougamers.com/hardware/stats/3dmark06/alltime/
There is only 1 ATi card in the most popular videocards of all time, most of them are nVidia and quite a few actually already are 8-series).
Secondly, it wouldn't be the first time that games included special features that were only available on one brand of hardware.
I think nVidia has a good chance of pulling this off, they may have a big enough marketshare. Even so, the PPU might remain in production... so they can sell PPUs to people with ATi cards.

But as I said before, currently nobody even knows what the ATi GPUs are capable of in terms of physics or GPGPU in the first place, because ATi never disclosed much of their architecture, and never released an SDK like Cuda. It could well be that ATi doesn't have the hardware capabilities in the first place. They may have put it off until DX11. Cuda is a technology that is above and beyond DX10 specs, and nVidia developed it specifically for GPGPU, like their Tesla cards. And physics was also promoted at the introduction, but they didn't have an API for it... they will now.

 

[Atech]

I just don't understand the hate and resentment NVIDIA is getting for this. Thier GeForce 8 cards are second to none, and their stream processing architecture is perfect for something like physics calculations. It's too bad ATI didn't get in on this (like they could afford it), but I'm really glad to see NVIDIA doing something with their great stream processing and CUDA architectures.

Because they have promoised GPU-physics before...remember?
Fool me once, shame on you..fool me twice...shame on me.

 

[Atech]

Yes, and with no 'slower' buses in the signal paths between PPU processors and GPU graphics hardware, many of the limitations of a standalone PPU will be vastly reduced.
ie Once the data is given to the PPU (that part of the GPU set aside), it doesnt need to pass any more data along the PCI or PCI-E buses before handing the results to the GPU for video processing.

This also reduces traffic on the PCI-E (and PCI) bus allowing other hardware to function with less increased latency while the PPU is in use.

The Graphics card solution is significantly better in many ways.

Physics still have to be corodinated in the CPU...unless you are talking about tier 2 physcis?

 

[imzjustplayin]

Excuse me, but don't confuse Havok FX with Cuda. Cuda is a GPGPU framework and can do any kind of physics you want. It's aimed at a newer generation of hardware than Havok FX was.

The point is, GPUs are GPU limited in games as they currently stand, having dedicated hardware is a necessity in mid-high performance machines.

I see doing physics on a GPU as an excellent option for lower power notebooks and whatnot. Having a GPU+CPU on one die or chip would be great, adding physics to the GPU on that combo would be perfect for low cost, low power solutions. The one thing I do worry however is that ALL laptops become GPU+CPU solutions and there would be no more laptops with dedicated graphics or whatnot.

 

[Scali2]

The point is, GPUs are GPU limited in games as they currently stand, having dedicated hardware is a necessity in mid-high performance machines.

Not all games though.
Especially RTS games tend to be CPU-limited.
Besides, the new trend in videocard-land seems to be CrossFire/SLI-on-a-card. So then you'll have two GPUs. This solution isn't quite as efficient, so perhaps you can use the spare cycles normally wasted because of synchronization issues for physics. Think of it as a form of HyperThreading.

 

[imzjustplayin]

Not all games though.
Especially RTS games tend to be CPU-limited.
Besides, the new trend in videocard-land seems to be CrossFire/SLI-on-a-card. So then you'll have two GPUs. This solution isn't quite as efficient, so perhaps you can use the spare cycles normally wasted because of synchronization issues for physics. Think of it as a form of HyperThreading.

Excuse my language but fuck that! I don't want to have to buy TWO gpus from the same company just to get physics processing, I want an add-in card. This solution much like AMD/ATI's solution with GPU+CPU is great for SFF machines but sucks for everything else.

If I need better picture quality, all I need to do is buy a new video card. If ATI is ahead in the video card arena because they released a "9700pro" and all I've got is a GF4, I don't want to have to buy two "9700pros" which were $400 so that I can continue to have physics support. GPUs aren't designed at all for this operation so they're inherently inefficient much like CPUs, I don't care that you're more "familiar" with GPUs and can accept them. We need the ability/option for separate, dedicated PPU, not this stupid half assed shit from nvidia. Just like with their 7950GX2, all this is meant to do is sell twice as many of their boards which I DON'T like one bit.

Also don't forget that the memory included with video cards is far more than a physics processor (as of now) currently needs and would be a downright wasteful. If you want to get things done faster you need specialized hardware, plain and simple, a GPU is specialized hardware, FOR GRAPHICS. Hence the name, G.P.U. (GRAPHICS PROCESSING UNIT) I want a P.P.U, PHYSICS PROCESSING UNIT.

They have add-in sound cards, why? Because they're far superior. I don't use integrated sound because it's garbage, so why would I want "integrated" physics processing?

 

[Scali2]

Excuse my language but fuck that! I don't want to have to buy TWO gpus from the same company just to get physics processing, I want an add-in card. This solution much like AMD/ATI's solution with GPU+CPU is great for SFF machines but sucks for everything else.

If I need better picture quality, all I need to do is buy a new video card. If ATI is ahead in the video card arena because they released a "9700pro" and all I've got is a GF4, I don't want to have to buy two "9700pros" which were $400 so that I can continue to have physics support. GPUs aren't designed at all for this operation so they're inherently inefficient much like CPUs, I don't care that you're more "familiar" with GPUs and can accept them. We need the ability/option for separate, dedicated PPU, not this stupid half assed shit from nvidia. Just like with their 7950GX2, all this is meant to do is sell twice as many of their boards which I DON'T like one bit.

Also don't forget that the memory included with video cards is far more than a physics processor (as of now) currently needs and would be a downright wasteful. If you want to get things done faster you need specialized hardware, plain and simple, a GPU is specialized hardware, FOR GRAPHICS. Hence the name, G.P.U. (GRAPHICS PROCESSING UNIT) I want a P.P.U, PHYSICS PROCESSING UNIT.

They have add-in sound cards, why? Because they're far superior. I don't use integrated sound because it's garbage, so why would I want "integrated" physics processing?

Please, study Cuda and the GF8-architecture before making such uninformed rants. A GF8 is nowhere near as dedicated as you think it is. Also, you don't seem to understand that we have unified shaders, which can do anything you like at any time... So you don't dedicate an entire GPU to physics, you just use some idle shaders to do physics. Hence my reference to HyperThreading.
You don't *have* to buy two GPUs, it will work with a single GPU aswell... It's just that ATi's latest product is a very popular 2-GPU solution, and nVidia has a product like that planned for the GF9-series. It seems that dual-GPU cards are the future.

 

[Nenu]

Physics still have to be corodinated in the CPU...unless you are talking about tier 2 physcis?

Thats significant because.... ?

 

[Atech]

Thats significant because.... ?

You are kidding right?

 

[imzjustplayin]

Thats significant because.... ?

I believe what he is talking about is second order physics which simply add competely useless particle effects unlike first order physics which can actually have an effect on gameplay.

Unreal tournament 3 "tornado" map uses first order physics, where there is collision of objects with another. Second order physics is using particle effects where it makes "sparkles" such as in City of Villians when casting a spell, things that have no effect on the outcome in a game.

 

[Nenu]

You are kidding right?

lol
I explained that there is an advantage to losing the PCI bus from the equation and replacing the PCI-E connection from PPU to GPU with much faster local bus transfers.
One thing that has remained constant is that the CPU will be talking to the PPU through a PCI-E bus.

I was wondering why you brought it up as a worthy item.

 

[Scali2]

I was wondering why you brought it up as a worthy item.

I think his point is that in a GPU physics solution, the geometry is available to both the physics and graphics processors at the same time, because they share the memory on the videocard. This cuts out transferring modified geometry through the PCI-e interface.

 

[Atech]

lol
I explained that there is an advantage to losing the PCI bus from the equation and replacing the PCI-E connection from PPU to GPU with much faster local bus transfers.
One thing that has remained constant is that the CPU will be talking to the PPU through a PCI-E bus.

I was wondering why you brought it up as a worthy item.

The PPU wouldn't gain any preformance if you went PCI-E with it.
It's not a question about bandwith, we are not talking graphics here...it's a matter of massive parallel interactions...not triangles.

The PPU in current form is far from saturating the PCI bus...and why would I waste GPU power on anything but graphics?

But hey, perhaps you super-dooper-magic-bus suddenly will make GPU-physcis a reality...