GPU Physics Dead

[ShaidarHaran]

X-bit (http://www.xbitlabs.com/news/multimedia/display/20071119065621_GPU_Physics_Dead_for_Now_Says_AMD_s _Developer_Relations_Chief.html)

Thanks Intel! :roll:

[John Reynolds]

Good news. Last thing I wanted was my GPU spending shader resources on physics, especially when games these days are generally GPU-limited (and I'm not a big fan of multi-GPU setups). Multi-core CPUs or add-in boards are the best way, and economics dictate who'll win between those two.

[ShaidarHaran]

I disagree. Yet another card is the last thing PCs need nowadays, unless we're talking about multi-function cards. Adding a second graphics card that would increase visual fidelity (via higher frame rates and/or more graphical effects/resolution/high-res textures/etc.) as well as increase the immersiveness of the game world through enhanced physics would be the lone exception there. Even a single graphics card handling the physics workload would be fine, assuming there's some measure of profiling that would prevent undue frame rate drops.

[Maintank]

If you listened to Nvidia and ATI. They didnt want you to add a second card, they wanted you to add a third card for Physics. This Physics on the GPU dying is no surprise to me and I wont shed a tear either.

The bottom line is MS needs to develope a Physics API. That will go a long ways in getting physics on a level playing field and into games.

[pjbliverpool]

I disagree. Yet another card is the last thing PCs need nowadays, unless we're talking about multi-function cards. Adding a second graphics card that would increase visual fidelity (via higher frame rates and/or more graphical effects/resolution/high-res textures/etc.) as well as increase the immersiveness of the game world through enhanced physics would be the lone exception there. Even a single graphics card handling the physics workload would be fine, assuming there's some measure of profiling that would prevent undue frame rate drops.

I don't think this is heading towards a seperate board. More that its focussing physics as taking place on the CPU which IMO is a very good thing. CPU's are becoming hugely powerful and games just aren't using all of that power effectively at the moment. I would love to see really instensive (and most importantly gameplay effecting) physics on mutilcore CPU's that can scale well with core count. IMO that much better than either using up GPU cycles on non gameplay effecting physics or buysing a seperate add in board.

[ShaidarHaran]

If you listened to Nvidia and ATI. They didnt want you to add a second card, they wanted you to add a third card for Physics. This Physics on the GPU dying is no surprise to me and I wont shed a tear either.

The bottom line is MS needs to develope a Physics API. That will go a long ways in getting physics on a level playing field and into games.

NV may have only mentioned 3+ cards for physics, but I distinctly remember ATi saying just two or even one card were capable of performing physics calculations in addition to the regular graphical workload.

I don't think this is heading towards a seperate board. More that its focussing physics as taking place on the CPU which IMO is a very good thing. CPU's are becoming hugely powerful and games just aren't using all of that power effectively at the moment. I would love to see really instensive (and most importantly gameplay effecting) physics on mutilcore CPU's that can scale well with core count. IMO that much better than either using up GPU cycles on non gameplay effecting physics or buysing a seperate add in board.

I didn't mean to imply that we are currently on the path to yet another discrete solution for physics, I was only responding to John Reynolds' assertion about AIBs.

I agree that the path for physics the industry is currently on is one of utilizing multi-core CPUs, rather than discrete cards (now that GPU physics is dead).

[Maintank]

NV may have only mentioned 3+ cards for physics, but I distinctly remember ATi saying just two or even one card were capable of performing physics calculations in addition to the regular graphical workload.


They both said it in different ways. Do you really believe ATI or Nvidia believed they could deliver physics on a single card? I dont think they even thought it was totally possible on two cards. The bottom line is even with their SLI and xFires solutions, the cards were being pushed to the max. A 3rd card that was taken completely out of the rendering process was the preferable avenue to get this done.

In the end they werent any different than what that one company was doing when selling an add in card.
I guess the one advantage of course is you should have been able to recycle old video cards to handle the physics duty.

[Geo]

http://www.xbitlabs.com/news/multimedia/display/20071119065621_GPU_Physics_Dead_for_Now_Says_AMD_s _Developer_Relations_Chief.html

According to Richard Huddy, who joined AMD when it acquired graphics chip company ATI Technologies last year, Havok FX is unlikely to be released at all or power many video games. While AMD admits that there are some games on the horizon that can compute physics effects on GPUs, it is highly unlikely that there will be a significant number of them, unless comprehensive tools for GPU physics are available.

Therefore, for AMD, which is the second largest provider of x86 central processing units (CPUs) in the world, it makes more sense now to promote physics calculations on its multi-core processors, granted that there are special development tools offered. As a consequence, without Havok FX and with no substantial intention to support it by AMD, GPU physics is unlikely to become popular in the short term future.

I remember back in mid-2006 that Orton was predicting that gaming physics would be moving significant amounts of Radeons by mid-2007.

[ShaidarHaran]

They both said it in different ways. Do you really believe ATI or Nvidia believed they could deliver physics on a single card? I dont think they even thought it was totally possible on two cards. The bottom line is even with their SLI and xFires solutions, the cards were being pushed to the max. A 3rd card that was taken completely out of the rendering process was the preferable avenue to get this done.

In the end they werent any different than what that one company was doing when selling an add in card.
I guess the one advantage of course is you should have been able to recycle old video cards to handle the physics duty.

It's all academic at this point, since it will never come to fruition.

[Silent_Buddha]

And if Havok hadn't been bought out by Intel that might have come true.

As it is, unless Microsoft moves forward with Physics on a GPU, it's probably dead in the water as far as gaming goes.

Which makes me sad as I'll soon have 2 graphics cards that could have been used as physics processors. **sigh**

Regards,
SB

[Sc4freak]

Screw Direct3D, I want DirectPhysics!

[_xxx_]

Finally.

[davefb]

might be a good thing , might not.

but the way it's happened isnt good, it' always amazes me why people talk about m$ being a monopoly, whereas intel blatently abuse a monopoly position all the time and get away with it!!

(okay so its a only monopoly coz they have ooodles of cash,and AMD is trying to stay solvent..)

if this was m$ , m$ would have bought havok and we'd be moaning about them buying another company , but it still would have come out ...... though maybe late,, and not working on xp or something.. but it still would have come out,, and probably would be nicely integrated into dx...

[_xxx_]

MS wouldn't buy Havok, they'd borrow the idea and make their own version ;)

Since it all goes the way of DX anyway, that's the only proper place to implement a "standardized" physics API IMO (although I do hate MS regardless)

[Acert93]

Too bad, as I think there was some potential here, both in performance as well as the market overlap, i.e. gamers who need performance for physics in games almost always need 3D acceleration as well. The concept of using the GPU as a more general resource would be noteworthy in gaming, but could then validate itself outside of the gaming arena. The ability for developers to use extensions in normal desktop apps to call up the computational power of the GPU is nice in theory. But Intel and AMD see physics as being a major driver for justifying 8+ core CPUs and losing that ground to GPUs is unacceptible.

(although I do hate MS regardless)

Could you just make this your sig and be done with it? :lol:

[_xxx_]

GPU Physics + 3D rendering in parallel is a bad idea. Look at the CPU's, it's kinda comparable to running two apps on a single-core CPU vs. multicore. Just too much hassle with overhead, scheduling etc. to pay off in reality. Although it would have been a somewhat nice way for recycling old cards in a multi-card environment. Though most people avoid multiple gfx cards anyway, the enthusiasts (and the would-be's) with SLI/crossfire are maybe 5% of the population, while everyone has or will soon have multi-core CPU's anyway.

[Blazkowicz]

GPU physics was vaporware anyway.
and not only a quad core CPU can take care of physics now, but a few years from now the units found in AMD Fusion and equivalent Intel CPU will do the job.

[ShaidarHaran]

GPU Physics + 3D rendering in parallel is a bad idea. Look at the CPU's, it's kinda comparable to running two apps on a single-core CPU vs. multicore. Just too much hassle with overhead, scheduling etc. to pay off in reality. Although it would have been a somewhat nice way for recycling old cards in a multi-card environment. Though most people avoid multiple gfx cards anyway, the enthusiasts (and the would-be's) with SLI/crossfire are maybe 5% of the population, while everyone has or will soon have multi-core CPU's anyway.

Last time I checked modern GPUs were capable or having far more threads in flight than modern CPUs, so what's the problem with running physics and graphics in parallel on the same card? Proof-of-concept demos showed it could be done.

[Geo]

http://www.custompc.co.uk/news/601680/amd-considers-buying-ageia.html

[digitalwanderer]

$100 million seems like a lot for a company that's about to go tits up. :???:

[Davros]

Would a DSP be good at physics ?
could be a good thing if an physics could run on a soundcard the xfi cpu seems to have a lot of power to spare

[Arun]

The problem is that they had about $65M of venture capital, and I'm not sure VCs will be willing to let go unless they at least make their money back. Anything below $40M or so is a big no-no, and realistically I'd be surprised if anyone could grab it below $70M or so.

However, I disagree with Huddy regarding who might have some interest in Ageia's technology. I really don't think Sony or Nintendo are potential bidders. Microsoft likely wouldn't be interested either, given they don't need the API and they likely won't want to do next-gen hardware in-house.

So that really leaves us NVIDIA and AMD. Honestly, it could fit in amazingly well in either company's strategy. The problem is that their track records regarding that kind of thing is abysmal, and the wrong strategy would simply play into Intel's hands, IMO.

[BRiT]

$100 million seems like a lot for a company that's about to go tits up. :???:

AMD or Ageia ? :-?

[Temporary Name]

GPU physics was vaporware anyway.
and not only a quad core CPU can take care of physics now, but a few years from now the units found in AMD Fusion and equivalent Intel CPU will do the job.
The concepts "a few years from now" and "AMD" aren't necessarily compatible.

[_xxx_]

so what's the problem with running physics and graphics in parallel on the same card? Proof-of-concept demos showed it could be done.

The problem is that the performance sucks badly when you do both in parallel.

[ShaidarHaran]

The problem is that the performance sucks badly when you do both in parallel.

And your proof of this would be.... Where?

[_xxx_]

And the proof of the opposite would be... where?

You read an article about a speciffically tailored app or two which "just worked" and thought it's the be-all-end-all solution? Uhm, yeah. As seen in those presentations, it struggled to do anything useful even with SLI'd/crossfire'd cards and a relatively simple static 3D app, just imagine how hard it would hit in a bit more complicated scenery and using real dynamic calculations. Think of it like running 2 games in parallel on your card.

[silent_guy]

And your proof of this would be.... Where?

Just look at CUDA: not only can you not run graphics and a CUDA kernel in parallel on the same GPU, you can't even run different kernels on different SM's in parallel.

[Arun]

Just look at CUDA: not only can you not run graphics and a CUDA kernel in parallel on the same GPU, you can't even run different kernels on different SM's in parallel.I was under the impression this was fixed with asynchronous queries? Baron, any chance to confirm/deny that?

[ShaidarHaran]

And the proof of the opposite would be... where?

You read an article about a speciffically tailored app or two which "just worked" and thought it's the be-all-end-all solution? Uhm, yeah. As seen in those presentations, it struggled to do anything useful even with SLI'd/crossfire'd cards and a relatively simple static 3D app, just imagine how hard it would hit in a bit more complicated scenery and using real dynamic calculations. Think of it like running 2 games in parallel on your card.

The fact that it was stated to be feasible. I'm more inclined to believe the IHVs with their proof-of-concept demos than I am someone just stating opinion.

[Silent_Buddha]

Except the thing that everyone seems to be ignoring is that...

At least from ATI's POV, they wanted people to usually...

1. Buy a second or third GPU to be used for physics calcs...or

2. Use an outdated/old GPU that a person ALREADY HAS for physics calcs.

I have a drawer full of old and outdated Video cards.

I was actually hoping that Havok would make progress and this time around I could reuse and thus extend the life of an older GPU by having it dedicated to calculating physics.

Much cheaper and much more practical than buying a dedicated physics processor since I'd just be re-using hardware I already have for a new purpose.

And much cheaper and more powerful than upgrading to quad or octo core just for physics processing. Since I'm re-using something I already have, there would be ZERO cost out of pocket for this.

Ah well... I guess it was too good to be true. Unless MS makes some moves in this direction.

Regards,
SB

[Davros]

Allow me to bump my original question would a dsp be any good for physics ?

[ShaidarHaran]

I'm not saying ATi/NV wouldn't seize the opportunity to sell an additional video card, but I doubt anyone here would argue that the added ability of GPU physics processing (even with just one card) would be huge incentive for people to even purchase one card in the first place. People that previously may have only gone for integrated graphics, may give a discrete card a second glance if GPU physics were an advertised feature.

[silent_guy]

I was under the impression this was fixed with asynchronous queries? Baron, any chance to confirm/deny that?

CUDA 1.1 indeed introduces the asynchronous API, but as far as I understand it (and I really haven't tried it myself yet), this allows queuing up different kernels, let them play one-by-one, and allow the CPU to query the status without blocking. The docs don't say anything about running multiple different kernels in parallel.

This question has been asked a number of times on the CUDA forums actually, and there was never an answer or hint that this would be possible.

[silent_guy]

I'm not saying ATi/NV wouldn't seize the opportunity to sell an additional video card, but I doubt anyone here would argue that the added ability of GPU physics processing (even with just one card) would be huge incentive for people to even purchase one card in the first place. People that previously may have only gone for integrated graphics, may give a discrete card a second glance if GPU physics were an advertised feature.

Note that I'm not saying it isn't possible to run both physics and graphics on the same GPU: there's no question that this can be done, as there are a couple of demos on the CUDA SDK that use CUDA to calculate something and display the results with OpenGL.

But they have to be done serially. You can't calculate and render at the same time, which is what you were hinting at by referring to the multi-threaded nature of a GPU. I'm not sure if that really matters that much in real life.

[3dcgi]

You can't calculate and render at the same time, which is what you were hinting at by referring to the multi-threaded nature of a GPU. I'm not sure if that really matters that much in real life.
Since the physics programs will be completely separate from the shaders I'm not sure it matters that they can't be part of the same draw call. Dependent on the architecture physics programs can run in parallel to an extent. A unified architecture could work on graphics and physics in parallel, likely with pipe bubbles when the physics program is kicked off and retired.

Of course, if the argument is that there's enough graphics work that it's not a good idea to try to also do physics then that's an argument that can't be won as it's entirely dependent on the game. AAA titles tend to be graphics limited, but maybe some games might not be.

I have a feeling that now that GPU physics has been shown to be possible it's up to game developers to demand it if they want it.

[Unknown Soldier]

If you listened to Nvidia and ATI. They didnt want you to add a second card, they wanted you to add a third card for Physics. This Physics on the GPU dying is no surprise to me and I wont shed a tear either.

Except that(lets say) the R700 with 4 cores .. one of them could easily do physics for the card.

US

[hoho]

Allow me to bump my original question would a dsp be any good for physics ?From what I understand physics calculations require a whole lot of random memory access. Caches do not seem to help much unless working dataset fits into them. GPUs are good at doing physics because they have loads of memory bandwidth and FP power but that doesn't mean they are particularly effective at it. I bet when one compares similar simulations on GPU and CPU the latter does the same thing but uses way less memory bandwidth and floating point performance thanks to doing more stuff per clock (per ALU) and because they are more optimized for random memory access.

Of course there are some simulations that are quite effective to run on GPUs. Cloth simulation and particle systems colliding with simple objects (heightmap) should have quite good performance.


Now to answer your original question, yes DSP's could be used if you give them good enough memory subsystem.

[Davros]

thank you hoho