Difference between Physics on a GPU and Physics on a PPU?

[SuperGee]

SM4.0 shaders are not 'general purpose' they are just able to act as either vertex or pixel shaders instead of having a constant set amount.

No, because teh new shaders are becoming so General purpouse, nV feels the need of a new name for them. "Stream Proscesors.

If it where just unified shaders and you only could do VS, PS, GS stuff. It would be only usable for GFX stuff.
Now sinds SM3.0 Wich have also some Genral purpouse use and Branching capability They can be use general purpouse.
So folding home
Physics
Video filtering
Can be done.

Writing back in SM4.0 is almost the same as in 3.0 other than fetch can be a bit faster due to the DX10 architecture without having to swap user/kernel mode. Also note that even if it was possible to do full gameplay physics on a GPU (which it's not atm) nobody in their right mind would do this because you would alienate 99.9% of customers that do not have the highest end GPUs or SLI.

Not sure if that true. But for folding home there must eb two way traffic.
Wich does the job.
So it i possible. It could be more efficient. But Keep in mind AIT and nV are now aware of the Physics hardware market and there next or maybe there currently new nextgen are enough optimised to doe the Physics job,
Wich means those Shader are General purpouse enough to do that job.

That is the reason that we will either continue to see physics other than fluff done on the CPU, unless they start building the physics solutions on all boards as part of the chipset, which I highly doubt.

Gameplay Physics have large impact on Gamedesign & produktion & Mechanics
Pathing AI and game complexity, Netload sychonisation load.
That keeps Dev's away from it.
Effect Physics impact and contrains and risk are minor. It's just more work

 

[Scali2]

No, because teh new shaders are becoming so General purpouse, nV feels the need of a new name for them. "Stream Proscesors.

That's purely marketing though...
The whole 'stream processing' thing existed for a while, and earlier GPUs can also be classed as stream processors... It's not a term specific to GPUs either:
http://en.wikipedia.org/wiki/Stream_processing

So it i possible. It could be more efficient. But Keep in mind AIT and nV are now aware of the Physics hardware market and there next or maybe there currently new nextgen are enough optimised to doe the Physics job,
Wich means those Shader are General purpouse enough to do that job.

The GPU can do (some) physics calculations yes... and it can do them quite quickly (not sure how it compares to CPU, but at least the advantage would be that the CPU is free for other tasks). However, I think it's nowhere near the performance-levels or flexibility of a real PPU, because physics aren't really all that stream-processing-like.
The key feature of the PhysX chip is an efficient network that connects the physics processors ('shaders' if you like) so that you can get the data from A to B quickly. This is something that a GPU cannot do. A GPU is far more limited in its data addressing, so you'll see a more straightforward, iterative bruteforce method.

The main disadvantage that the PPU has is that it's so much faster that the physics are in a class of their own... On one side that means that no current games will be able to use the PPU to its full potential... and on the other side it means that if new games want to take full advantage, they have to have a 'watered-down' version that will be able to run on CPU or GPU.
So it's a bit of a catch-22... PPUs are great, but making the step from CPU/GPU to PPU may be too large for them to ever take off.

 

[SuperGee]

Marketing sure. The marketing would jump on that feature and exploid it. But that doesn'nt mean it don't excist.
First in GPU enable encoding a GPU beats CPU by a large margin.
So they are powerfull,also for other thing then shading.

But keep in mind. CPU are serial Proscesing. 2 to 4 threads.
While a G80 as a example is already something like a 128 Core CoProcsesor.
Task that are very suitable for Parralel computing like Graphics and Physics and encoding. Scale better on GPU's Till CPU cores grows in number and some of them are specialised. number Copro's on die, but that the future.

Don't forget PPU is a Lightweight 130nm 125mil Transistor solution.
It must be very well specific optimised like a factor of 5 to 10 time to beat the oposition.
Wich have a large part of near 700mil transistors it's 128 Stream proscesors for Physics.
Lets Say halve of it are for shaders 350mil so yoy get halve of it to share for Physics 175
But on higherclock due to dieshrink and dubble pumped.
We don't know excactly how they perform.
But if you look at nV Physics demo the are using a more fine grain computing.
While PhysX use a large grain solution.
That say something about computing power.
Or could be a memory limit thing.
In my opinion Ageia needs a PPU2 already now on a modern up to date die shrink.
130nm was a big step back and that must go on for 2years. While its competition cranck out high power GPU in a way faster cycle with a lot more transistors and more power to consume on the latest dieshrinks.

And PPU have a 128MB limit to store Local Physics Data.
While GPU shared can use the halve of 512 to 1GB
Or Dedicated from 256 to 1GB full dedicated for Physics.

One day there will be a Physics bench suit who give us more insight in performance difference.

So PPU2 needs at least 500Mil transistors. and 65nm and 1GB MEm
and come in 3 flavors. And SLI support

I think Ageia didn't had calculated in such competition so soon. Before the have build up a new stable market.

 

[Scali2]

Don't forget PPU is a Lightweight 130nm 125mil Transistor solution.
It must be very well specific optimised like a factor of 5 to 10 time to beat the oposition.
Wich have a large part of near 700mil transistors it's 128 Stream proscesors for Physics.
Lets Say halve of it are for shaders 350mil so yoy get halve of it to share for Physics 175
But on higherclock due to dieshrink and dubble pumped.

As I say, the real strength of the PPU is in its communication network, and that's where a GPU is very poor.
The actual processing power in terms of flops is not all that important.
I have yet to see generic tests such as the NovodeX SDK samples on a GPU. Not sure if it's even possible.

 

[SuperGee]

At the moment not, but they will use HAvokFX for it.
Benching CPU vs PPU is easy the API support both.
But PPU vs GPU. Depends on different middleware API.
But are full features API's. So they can do both code paths to support both hardware.

Like Graw uses Havok for CPU and PhysX for PPU.

It could be HavokFX and PhysX in one game and let the game detect and choose the code path in runtime depending on detected hardware or setup option choices.
DirecXPhysX would be nice but still far and away for now.