Some interesting independant performance numbers for ATI.

[Menelmarar]

Consider if you will, that interactive physics on the cpu and accelerated physics on a PPU as merely a stepping stone to interactive+effects physics all acclerated on a PPU. This is what everyone wants, no doubt about that! Baby-steps!

Step 1: Accelerated Effects Physics.
We can see this stepping stone in games already like HalfLife 2, UT2004, Doom 3, Oblivion, GRAW and even futures titles like UT2007 and Crysis where the interactive physics are done on your cpu. As well as titles like GRAW and future UT2007 are hardware accelerating the effects physics aspect.

Hypothetically, had Havok and ATI/nVidia been ready, instead of paying $299 for a PPU to accelerate these effects. The mainstream gamer could have paid $99 for an X1600 for the same thing! Having a price that was 1/3 of what it is now could really have helped accelerate the market for accelerated physics.

Then consider the improvements HL2 and Oblivion have shown physics will add to gameplay using only the power of CPUs. Crytek studios for their upcoming title Crysis still think that CPUs are still capable of taking interactive physics to the next level AND without costing consumers unncessary money.

Backing up ATI's performance claims a bit. Take a look at Stanford University's folding project. I found in the folding forum today details on their research into implementing their simulation of protein interactions on ATI's GPU. I thought this was critically important to chare with the crowd interested in PPUs.
http://graphics.stanford.edu/~mhouston/public_talks/R520-mhouston.pdf

[IG]http://www.softwaredaddy.com/~menel/benchmark1.jpg[/IMG]
[IG]http://www.softwaredaddy.com/~menel/benchmark2.jpg[/IMG]
[IG]http://www.softwaredaddy.com/~menel/benchmark3.jpg[/IMG]

Impressive to me! Gives a little more validity to ATI's claims of how fast they can be at general purpose computing with parrallel data.

What is really interesting is they seem capable of running these calculations on the GPU and returning data back from the GPU. If they can do it, why can't Havok return physics results for calculations in games? Maybe there are other flaws in the architecture. But that is something that ATI should be able to focus on improving for future generations.

So with this, as it imrpoves. Havok and ATI both have huge market share already and developer support from previous projects. It seems reasonable that theier strong ties could give them strong allies in future physics technologies.


Future Goal: Interactive+Effects Physics Accelerated
The phase where interactive+effects physics are hardware accelerated without bleeding the consumers unncessarily.

With Windows Vista and Direct10 coming up. You can bet ATI and nVidia are already hard at work and have been for some time on new generations of GPUs. Maybe it's to late for the first round of DX10 GPUs, but maybe the following, or the following 6month refresh.

Can you imagine the market advantage it would give ATI over nVidia or vise versa if they could improve the methods for returning data from the GPU. Making their graphics cards capable of interactive physics. With Havok onboard to work up the SDK and implementation side for game developers, they've got alot of experience with this. Can you imagine the possibilities?

What if...Next spring/summer upgrading to a theoretically named ATI X2800 DX10 GPU for your system for $500 and a second $99-$150 X2300/X2600 ATI gpu in your system for interactive+effects physics.

Not only are thse mainstream/budget GPU's be cheaper than a PhysX's single purpose implementation, but they can be recycled. When they start to become obselete and no longer serve you, the enthusiast anymore, they could be passed down to little sister/brother's pc. Or used in a peicemeal web browing PC for a family member. Or your main graphics card fails and has to be sent off for RMA, whoops just bump your PPU/GPU upto the primary x16 slot.

Interesting....

 

[MrNasty]

Well, basically it comes down to this:

Those GPU calculations are basically packaged codes that run on the GPU once and don't have to be altered according to the final state, ie they are absolutes, much like doing 2+2=4, the end result does not affect any of the results obtained by the rest of the code.

Perhaps a better example would be solving a set of complex quadratics - it's a number crunch that running the solution through multiple times can solve easily - iteration in other words. Hence graphics cards - able to perform a huge number of calculations at the same time - can converge on a solution much faster than a CPU can.

Where this starts to become problematic to, say, fluid simulation is that the result of the immediate predecessor affects the current solution. So in real-time this causes problems in that you solve x+yt=4, 2x+3yt=16 and x=3t but then immediately because of t's constant changing x and y do as well, hence the equations are constantly changing. This is a huge simplification - in actual fact it is very easy to perform a 2d non-turbulent/compressible fluid model on a GPU and I've been looking at 3d methods for a while, but because of limitations in the way this sort of direct numerical simulation works if you suddenly can't quite finish a calculation or there is an instantaneous singularity the calculations after this point can be very innacurate or even downright stupid.

This is where communicative parallelism such as exists in the PPU is good: these systems can be simultaneously solved and the results compared by solving them in a different way from the previous case - you can essentially do away with the horribly complex timestep method and use a time-averaged method which is much easier to solve for but requires this kind of architecture. Hence the PPU is much more suited to fluids modelling.

For rigid bodies this architecture is also a bonus, as instantaneous collision prediction can be done, preventing unnecessary work being performed if say a cube hits another one - the unit that is calculating the rotational and velocity vectors of one cube can stop doing that as they collide rather than going on blind if it was the case that the calculation had to be finished before collision could be determined. There are all sorts of optimisations and so on so forgive me for being so simplistic.

So what it really comes down to is this: a GPU may well be faster than the PPU for gromacs or whatever because it has the raw power benefit, but the elegance of solution is not there. Those folding/computational demonstrations all are tailored for the GPU's architecture, which in turn is designed for one thing: applying as many equations to data (implicitely vice versa) as quickly as possible - otherwise known as number crunching.

Where the PPU's power comes through is in simulation work that requires inter-dependant solutions.


Now all that said, it is perfectly possible to apply any problem to any architecture (CPU to GPU, CPU to PPU, GPU to PPU, and PPU to GPU) but the keyword for this is efficiency. Speed is a byproduct of efficiency as most engineers will tell you. That's a gross simplification but at it's heart the goal of the engineering profession is efficiency, pure and simple. And efficiency for physical simulations is something that the PPU most definitely has.

EDIT: Further to this-

What you just proved in a best case is that for certain calculations the GPU can be around 11x faster than a CPU for those calculations. That's great and all well and good. Problem is, according to AGEIA, the PPU is around 20-30x faster than a CPU at rigid body calcs (50,000 objects max, compared to 5,000 objects max). So by that line of reasoning technically the PPU is about 2-3x faster than a GPU. In fact ATi themselves have shown (at computex) their system can handle 20,000.

So things just aren't adding up for ATi's performance claims right now - I want to see some product before I run off saying "wow that's fast"...

And to answer your problem about interactive physics - I really don't know if GPU's will ever be able to do that, but if they can it's beyond the scope of HavokFX, which doesn't offer interactive physics as part of its feature-set, which then begs the question "how long will it take for the API to be made and how long till games use it after that?".

EDIT 2: To see how the X1900 and PPU stack up see here

 

[quadnad]

So things just aren't adding up for ATi's performance claims right now - I want to see some product before I run off saying "wow that's fast"...

I'd like to see a product that actually makes me want a physics accelerator in the first place. Lets just say that horsepower-wise, one is faster than the other. Lots of times faster. Does this change the fact that there is nothing worthwhile or living up to the claims out there utilizing all this power?

I still need a game to arrive that actually shows me that all the demos and videos showing boulders rolling down a hill will actually be utilized in a meaningful way.

I'm still a bit surprised that we haven't started by using the second core on a CPU. I know it isn't nearly as fast as a dedicated PPU or GPU/PPU, but have I even seen a reason why I would need one at all? Can we not start somewhere, and once the demand is realized move on to more complex and expensive dedicated solutions? Dual core CPUs are far more prevalent and less expensive than the equivalent PPU solution, so in my opinion it would be a great place to start.

 

[psychoace]

Even if this is true you got to remember power isn't everything. Betamax was a far superiour product compared to VHS. It had Higher res and everything. What made it the suckatude though was it's propitary stance. Sony wouldn't liscense it's products to anyone else. It was a Sony only product and no one could get in on it. Ageia is kinda running along that same lines. They are the only one's making the chip for there physics scheme and so they limit there potential market place by doing that. When you limit your market game developers see no reason to consider you a partner or a viable concern of financial resources. If having the Ageia Physix engine as part of there game ment there would be more people buying there game then more game companies would produce games with it. Cost wise though it's showing to not be a profitable resource yet. Most games that are going to be using this engine are using it because they don't want to have to pay the cost of building there own engine. Most big developers have contracts with Havoc or someone already so they arn't going to change. If they could get this stuff installed with 0 cost to them (which is impossible because even if it Ageia was giving away liscenes it would still mean the developer would have to pull resources from other departments in order to add this whole new engine into the game. That's a lot of people spending a lot of hours on something that dosn't show much of an option for profit.

 

[MrNasty]

I'd like to see a product that actually makes me want a physics accelerator in the first place. Lets just say that horsepower-wise, one is faster than the other. Lots of times faster. Does this change the fact that there is nothing worthwhile or living up to the claims out there utilizing all this power?

I still need a game to arrive that actually shows me that all the demos and videos showing boulders rolling down a hill will actually be utilized in a meaningful way.

I'm still a bit surprised that we haven't started by using the second core on a CPU. I know it isn't nearly as fast as a dedicated PPU or GPU/PPU, but have I even seen a reason why I would need one at all? Can we not start somewhere, and once the demand is realized move on to more complex and expensive dedicated solutions? Dual core CPUs are far more prevalent and less expensive than the equivalent PPU solution, so in my opinion it would be a great place to start.

Great point. The way I see it going is that PPU's won't be widely adopted by next year, but Ageia's multicore-enabled SDK should allow games to run well in software too. Having a PPU will enable a few cool things like fluids and cloths and it'll gain momentum from there, eventually allowing games which use the PPU heavily to come along.

I think AGEIA's made a solution to a problem which isn't here just yet, but the fact that the solution exists will speed progress along - I think they're banking on that.

This applies equally to Havok and HavokFX, just within a different feature set.

Ultimately only time will tell which is more popular (all that matters. Period.), it's just a lot of people seem to be trying to say that HavokFX is better than PhysX, when technologically that's just not the case.

 

[psychoace]

I could care less if Havoc solution was slower I think most people are complaining about is that your all saying ATI/NVIDIA arn't providing any kind of physics solutions and thus should be called such. It's like you all want Ageia to equal PPU and whoever tries to take over there territory shall die in a blaze.

 

[MrNasty]

Even if this is true you got to remember power isn't everything.

QFT.

But don't think Ageia's forcing a proprietry standard alone. HavokFX is too.

You'll *need* two graphics cards to do physics with HavokFX, and hence this is also pushing technology in the same way as if Skype started telling everyone you'll need two AMD processors to use the conference call feature. Even if it's technology that already exists, requirements are what makes proprietism possible.

In one sense neither are proprietry - both work on your CPU already - and many developers hold no alliegance to either because these are both (to varying degrees) paid-for products, unlike nVidia's TWIMTBP branding which developers get paid for (or at any rate bought lots of nice dinners ). There is no compunction to "stay loyal" in this particular situation, as Ubisoft/GRAW proved by including last minute PhysX support in GRAW.

What you're saying is that somehow devs are already tied into using Havok... Well tbh since 2000 only 38 PC games have used the system. Havok does not hold a monopoly on the market as many devs prefer to develope their own physics engines (farcry for example).

Another point: AGEIA's API is basically non-commercial opensource - no-one needs permission to use it for non-commercial purposes, this compared to Havok's bulldog licensing rules (read their website) which closes their software even to the devs using it. If you're being a "defender of the free open standard" you're arguing for the wrong side there.

 

[SuperGee]

Not so interesting and independant doesn't matter. because what is benched? Some shynthic full lab test university stuff that use the wole thing with a lot more units then Havok only sM3.0 pixelshaders. On a more widely usage. The test isn't much Game Code Physics related. And the slide only let you now how much they want make GPGPU usage easyer by implemeting Write back option and a whole lot more wishes.
Because retail games would uses HavokFX. instead that GPGPU stuf. A Havok FX game demo benches says a lot more but not all saying then this. It only said some GPGPU codemixes can be done that much better by a specific GPU vs CPU setup.
It's indirectly related to Ageia ATI havok batle.

A bench demo Like CellFactor wich have the Software and PhysX and HavokFX Where a lot of different Physics Features can be enabled/disable. Would be nice. But expensive to do because Havok FX cost a lot extra on top of Havok for dev's

Also there are mentions a lot of old and new CPU Physics enable games.
Well Physics is old stuff it implemented in a lot of games for a long time already. But the point is. The Physics Load. Its all about that. CPU can get in time, more power for Physics. More cycles avaible. This is the normal scale-ing and Crisis would be the latest of them to push CPU physics limit to a higeher level used in a game.
How ever Physics is old an PPU and Havok FX give this level of Physics a large Bump in load what can be handle in hardware. Agains what a CPU can deliver.

So There is a steep load and Physics feature mix difference in all those games.
A PPU can do more and a GPU can do a lot more eye candy Physics.
Imangine what could been done if Crisis supported Ageia and HAvokFX.
PPU for interactive Physics and Havok FX for eye candy.
The game would need a CF render setup to render what those PPU and Physics dedicated GPU would generate to render.

Of course devs must come up with a gameplay mechanic that could put this Physics power to gooed gameplay use or eyecandy use.