PhysX - Which dedicated card is powerful enough?

[newtekie1]

I've seen a few outragous claims about what is required for a dedicated PhysX card. So I decided to do a quick test with what I had laying around to see.

The two biggest things I've seen are:

• A 128-bit makes the card too slow.
• A G92 w/ 112 shaders is needed.

Ok, this is just a quick test, using Batman:AA, because it is the holiday weekend, and I didn't want to spend hours pulling machines apart and benchmarking every combination I could.

Test setup:

Xeon X3370@3.6GHz
RAM: 8GB PC2-800 4-4-4-12
eVGA 780i
Primary GPU: GTX470@800/1600/900(c/s/m)
PhysX GPU: 9600GT
PSU: Corsair HX850

| Min FPS | AVG FPS | Max FPS
PhysX Disabled | 74 | 128 | 186
CPU PhysX | 11 | 17 | 26
GTX470 | 32 | 61 | 102
GTX470+9600GT | 32 | 61 | 102
GTC470+9600GT UC *| 27 | 58 | 102

Some interesting results. There was no change by adding the 9600GT, kind of odd, I expected better performance. So that by right there tells me the GTX470 is powerful enough to do PhysX in Batman itself. I'd try again with Metro, but I can't find an accurate way to benchmark it easily. To me this confirms also that when PhysX is enabled, at least with Batman, performance is purposefully limitted, probably to meet a target framerate to ensure smooth gameplay.

Now, in the last test, I underclocked the 9600GT as low as possible, both the core, shaders, and memory. I did this to see how a much weaker card, with far less memory bandwidth would perform. There was almost no difference, but it did finally show a little performance difference.

There are also other users that report great results with 128-bit cards. Erocker uses a GT240 with his HD5850s in crossfire and reports good results.

Now, what is my conclusion? You don't need a powerful GPU for PhysX. PhysX relies more on the shader power of the GPU then the memory bandwidth. Remember, the original PPU only had a 128-bit bus.

I think the minimum I would recommend for a dedicated PhysX card is a GT220, or 48 shaders. I would not recommend anyone with a decently powerful nVidia card already to waste any money on a dedicated PhysX card. Anything from the GTX200 or GTX400 series is enough, don't waste your money. For anyone with an ATi card wanting a dedicated PhysX card, do not spend a large amount of money, and don't worry about getting a high performance card, go by the minimum I set above. For the small number of games that would even use it, it isn't worth spending more then $50 on a card.

 

[erocker]

I completely agree with you. GT 220's and 240's make great PhysX cards. They are generally smaller cards that won't restrict airflow to your main GPU's and they don't require an external power connector. If anything, having the PhysX card in my machine has dropped the temp of my 2nd GPU due to the PPU card cooling the backside of the GPU. Power consumption at idle went up a mere 12 watts.

 

[Loosenut]

Thank you for the brief review Newtekie, you may have saved me a few bucks

EDIT: nice rig you got there Erocker

 

[Lionheart]

Since that News about using a Nvidia card with ATi GPU's, ima grab me a GT240, feels like the perfect card for a dedicated PhysX card

 

[$immond$]

Even if Nvidia decides that they want a monopoly on the physx technology, I am sure we will find away to modify the drivers.

 

[devguy]

My friends that have done this ATI + nVidia combo all agree that the perfect PhysX card is the 8800GS (96 stream processor 9600 gso). It does take an additional pcie connector though. I'm intrigued that a GT220 is good enough. Good info!

 

[$immond$]

I would only settle for a 9800 GT

 

[erocker]

Min: 49 fps
Max: 112 fps
Avg: 92 fps

All settings maxed, PhysX on high

Watching the bench I don't even notice any slowdown, everything is smooth as butter. Considering it doesn't even use any more than 50% of the physics card, a GT 220, maybe even lower would be fine.

 

[KainXS]

I have heard that people use underclocked 9600GSO's that have the external power disabled, anyone know if thats really true

 

[TheMailMan78]

Thanks for the info newtek. I've been debating for a while now to get a Physx card and this helps.

 

[ShRoOmAlIsTiC]

Min: 49 fps
Max: 112 fps
Avg: 92 fps

All settings maxed, PhysX on high

Watching the bench I don't even notice any slowdown, everything is smooth as butter. Considering it doesn't even use any more than 50% of the physics card, a GT 220, maybe even lower would be fine.

compairing mine to erocker, I have a gtx260 for physx with a 5850 and get
Min 46
Max 128
avg 87

 

[erocker]

compairing mine to erocker, I have a gtx260 for physx with a 5850 and get
Min 46
Max 128
avg 87

With a single 5850? CrossFire works great with this game.:shadedshu

 

[ShRoOmAlIsTiC]

With a single 5850? CrossFire works great with this game.:shadedshu

my bad I forgot you had crossfire. Anychance you could retest it with crossfire off?

 

[erocker]

CrossFire off I lose 5 fps avg. The PPU maxes out at 65%. This is one of those games that I don't expect CF to work very well anyways. But what is strange is when the bench is running with CrossFire it seems to go faster. Meh.

 

[newtekie1]

Min: 49 fps
Max: 112 fps
Avg: 92 fps

All settings maxed, PhysX on high

Watching the bench I don't even notice any slowdown, everything is smooth as butter. Considering it doesn't even use any more than 50% of the physics card, a GT 220, maybe even lower would be fine.

I would really think even a 8600GT or 9500GT would be enough, those would be the minimum though(32 Shaders). I think if I was buying today, I'd get a little buffer in cause PhysX gets a little more use, and the minimums go up. I'd say at least something with 48 Shaders(crappy version of the 9600GSO or GT220). Memory bandwidth definitely doesn't matter though, so the $50 8800GT 128-bit newegg has as a shell shocker yesterday would have been killer as a PhysX card.