General purpose for physics

[DarkRage]

I think it is an interesting subject, but it was linked in a non-related thread. So let's just create a new thread. ;-)

AGEIA PhysX Hardware benchmarked:

http://anandtech.com/video/showdoc.aspx?i=2751

Dissapointing in my opinion. It has been discussed in this forum that physic calculations are not so massively vectorial-nothing-else calculations like AGEIA or Sony have tried us to believe. It looks like both could be wrong.

Benchmarks for GRAW are hard to interpret, as they are showing a different complexity level, but framerate drops are too big for a 300$ investment, IMO. But test aplication, with no GPU dependency and loads of simple calculations is a big dissapointment. Differences with a single core are pretty small (15%). I can imagine the comparison against the upcoming dual and quad Conroes with improvements on SSE units, or even a dual AMD core.

It is safe to expect an improvement on both drivers and PPU usage, but I was expecting much more from a dedicated chip like this. It is also possible to expect bigger differences with things like fluids simulations.

It makes you wonder too if Xenon could be better suited for physics than expected, even betterthan Cell, as benchmarks are suggesting than a general purpose processor is competing extremely well against multiple simple vectorial units...

 

[Titanio]

Framerate isn't a good way to gauge relative performance with something like PhysX. From talk before, a PhysX card could constrain framerate because of the bandwidth to it (I think Carmack alluded to this), but the more important quanitifier is how much is being done per frame.

The GRAW comparison obviously is apples to oranges in that regard. In the first "tech demo" comparison they concede that it's difficult to know how performance would scale with more boxes (more work per frame) versus the two.

Also, this ain't really relevant to the consoles at all. But "I see what you did thar".