SunnyD
Belgian Waffler
Originally posted by: thilan29
Originally posted by: zagood
For anyone looking at uber-cheap nV cards to pair up for physx, look at some benchmarks. I assume they'll be similar using an ATi card as your main one.
http://www.pcgameshardware.com...nchmark_review/?page=2
http://www.pcgameshardware.com...ative_to_Nvidia_Physx/
Hmm...8400 seems really slow...maybe a 8600GT or maybe even 9600GT is the minimum? I'm surprised it's that slow though...I have my 8800GT underclocked to 400/1200/800, it made no difference from when it was at 650/1600/950 and it doesn't heat up at all while doing any PhysX so I figured the PhysX calculations required very little horsepower.
Given that Physics are inherently parallel calculations that are well suited to things such as GPUs... let's take a look at what's happening shall we?
8400GS - 64-bit memory interface, 450MHz core clock, 400MHz memory clock, 900MHz shader clock, 16 shaders - 19fps
8600GT - 128-bit memory interface, 540MHz core clock, 700MHz memory clock, 1180MHz shader clock, 32 shaders - 51fps
9600GT - 256-bit memory interface, 650MHz core clock, 900MHz memory clock, 1625MHz shader clock, 64 shaders - 59fps
9800GTX+ - 256-bit memory interface, 738MHz core clock, 1100MHz memory clock, 1836MHz shader clock, 128 shaders - 64fps
It looks to me judging by the pcgh review that PhysX requires 32 shaders to run properly, and from that point on it's a function of speed with diminishing returns (probably due to bus traffic, etc).
Again, if we look at the 8400GS versus the 8600GT, this appears to account for the 63% difference in speed - 50% comes from having the requisite number of shaders, the remaining difference is from the increased shader clock. Once you get past 32 shaders, it appears to become purely a function of shader clock speed.
Interesting to wonder - what would happen if nvidia were to truly allow an entire GPU to be utilized as a PPU, basically allowing multiple PhysX calculation groups in parallel. Technically a 9800GTX should be a bit more than 4 times faster than an 8600GT if that were the case. But odds are that it won't make much when given the fact that data must be pushed between CPU, PhysX GPU, and the Video GPU. We already see this when PhysX is calculated onboard the same GPU - there's a performance hit because PhysX data is pushed from the CPU to the GPU for calculation, then back to the CPU and once more back to the GPU for rendering.
Originally posted by: evolucion8
The PhysX card proved to be even faster than the 9600GT, so I guess that the minimum should be a 8800GT or better. Probably the new single slot 9800GT can fit nicely there.
I would say probably the 9600GSO, probably even (or especially, due to the faster shader clock) the newer 512 version will fit the roll for PhysX and be the best price-performance you can get - 48 shaders, 1625MHz shader core, 256-bit memory bus.
