totally newbie question

[notAFanB]

hi, I am wondering do current generation of GFX cards still keeps animantion and physics routines on the CPU side? are there any plans to move to the GPU domain or not?

thanks.

 

[Nick]

Animation is partially done by graphics cards in their vertex shaders, e.g. waving flags. Physics is much more complex and the calculations are not so parallel so it can't be accelerated by current graphics chip designs.

But there is also no need for it. CPUs are still clocked much higher than GPUs and are very good at executing that sort of algorithms. Dedicated physics engines use lots of assembly optimizations to work very efficiently.

 

[Dio]

But there is also no need for it.

Yet

A VPU is more similar in architecture to a classic supercomputer (Cray etc.) than the CPU is. These vector computers are, traditionally, better at physics calculations than CPU's are. So in the long term VPU's may be much better than CPU's at physics.

But right now, as Nick says, no need at all.

 

[KimB]

hi, I am wondering do current generation of GFX cards still keeps animantion and physics routines on the CPU side? are there any plans to move to the GPU domain or not?

thanks.

The entire idea of GPU's was to offload the graphics so that the CPU would be left to have more time for things like physics and AI. Some animation has been making its way to the GPU, but not all (obviously, since physics simulation for games is animation!).

That is to say, I doubt they'll ever be accelerated by the GPU. The main goal is to continue to make the graphics depend less and less on the CPU, leaving the CPU to do these things.

 

[notAFanB]

thanks for the replies guys. it's been informative

 

[Grall]

I assume the CPU would read back the numbers the physics calculations spits out from the GPU, because if there isn't some kind of feedback from the video card if it's supposed to handle this also, how could the CPU know what's happening on-screen?

*G*

 

[Myrmecophagavir]

There is work being done to move some physics stuff to GPUs, though not really in the sense of "physics" in games at the moment (rigid body dynamics)... check out Mark Harris' page - he's around here isn't he?

http://www.cs.unc.edu/~harrism/cml/index.html

 

[Dave H]

But there is also no need for it.

Yet

A VPU is more similar in architecture to a classic supercomputer (Cray etc.) than the CPU is. These vector computers are, traditionally, better at physics calculations than CPU's are. So in the long term VPU's may be much better than CPU's at physics.

But right now, as Nick says, no need at all.

OTOH, general-purpose CPUs have been evolving in the general direction of supercomputers as well: beefy SIMD units, high-bandwidth high-latency (relative to cycle time) memory accesses (and prefetch to deal with it), etc. Certainly as GPUs gain more capable vertex shader pipelines many calculations will be offloaded to them. But large-scale physics and the like should still be more efficient on the CPU.

 

[Mintmaster]

That is to say, I doubt they'll ever be accelerated by the GPU. The main goal is to continue to make the graphics depend less and less on the CPU, leaving the CPU to do these things.

I wouldn't say this at all. Even for the GF3, NVidia had a water simulation demo running on the GPU to handle the waves caused by moving objects, which would take much longer on the CPU. Once texture access gets into the vertex shader, there will probably be a lot more physics done by the GPU, such as cloth and fluid simulation. Of course, the core physics will be done by the CPU, and act as a backbone for feeding the physics to the GPU and tying everthing together.

I think physics on the GPU will be very important in contributing to realism within a few years.