Seperate Cards for PhysX?

[BenSkywalker]

3DMark dropping points when using a dedicated PhysX card makes complete sense with the setups you are comparing, invert the setup and it will work the other way.

What is going to be faster handling physics? A 8800GTX or a 280? Of course the 280 will be, that is a no brainer. If you hook up your 8800GTX as a primary board and use your 280 as the PhysX card then your physics score for 3DMark will be the highest it can be with your hardware. Your overall score would go down enormously of course.

There are many reasons this makes complete sense, the easiest way to explain it is the Physics test is just that, a physics test. A 280 is going to beat a 8800 in a physics test every time if they are running identical code. That said, if the 280 is doing everything it can to render a scene, it isn't going to have spare resources to handle physics. This is why the offloading to a 8800 can cause a dramatic improvement in framerates in game, while showing a decrease in a synthetic bench.

I was speculating about this with Apoppin the other day. It is possible, that PhysX steals memory, memory bandwidth, and memory controller cycles when run on the primary card.

Not only that but it is going to be consuming internal cache for the GPU, which may cause a considerably larger performance impact then anything else. If you are exceeding the limitations of the on die cache by adding in the strain of the PhysX processing on the same GPU then you are going to create stalls in the rendering pipe which is catastrophic for any GPU's performance. That isn't to say all of the reasons you listed won't be a factor, just there are reasons where the performance could scale in a way that may seem to defy all logic(like a 8600 doubling the performance of a 295 setup).

 

[Keysplayr]

Originally posted by: cmdrdredd

Originally posted by: apoppin

Originally posted by: Keysplayr

Originally posted by: cmdrdredd
I just installed a 9600gt as a physx card. here's my problem or my question. How come I went from 45fps average in cryostasis with just a GTX295 to 77fps average, yet I lose 4k points on 3Dmark and when looking at the results side by side I noticed the physx portion of the test was almost 1/2 what it was before. Am I to assume that 3D Mark is simply forcing multiple calculations on the card and not doing any real 3D work to simply get a max number of calculations per second, thus the GTX295 being set as the physx card would be superior in this test while a game that is rendering textures, lighting and geometry etc would benefit from offloading the Physx work?

I was speculating about this with Apoppin the other day. It is possible, that PhysX steals memory, memory bandwidth, and memory controller cycles when run on the primary card.

Offloading PhysX to another card that has it's own resources to draw from seems to be the best method for the most performance in PhysX games. While a single higher end card can handle PhysX pretty well, the reason for the rather large jump in performance "could" be as stated above. That could be why a 16sp 8400GS offered so much performance gain. At the same time, we are freeing up valuable resources on the primary card, and granted even more in the form of a dedicated PhysX card. So, it doesn't appear that PhysX is all that shader hungry, but it does "appear" to be bandwidth (memory) hungry. Thus the 256MB minimum requirement for any given PhysX capable card.

This is just a theory, I really have no info yet. I was going to configure some sort of testing on this. Just have to figure out how, and what to do.

i'd be glad to help; it is an easy test to set up, but your PMs are set to ignore me


ANYway, the 2nd card does wonders for FPS in Cryostasis demo; i am using GTX 280 + 8800-GTX

in testing that configuration in Vantage, the score only drops a little bit (GeForce 186.18):
- with a single GTX 280 = 13046
- with GTX 280 + 8800 GTX = 12867

The FPS are the same in the mini-games

I went from ~22k with a GTX295 only to ~19k with a 9600GT doing physx in 3dmark vantage. Games show vast improvements ranging from 10-40fps depending on game and scene.

Well, I think your improvements are where they are supposed to be. In games. Wouldn't worry too much about 3DMark.

 

[Keysplayr]

Originally posted by: apoppin

i'd be glad to help; it is an easy test to set up, but your PMs are set to ignore me

Sorry about that. Did that a long time ago when we were snarling at each other.
PM's unblocked. I'd appreciate the help.

Although I'm pretty sure that what I said before is exactly what is happening. As Ben had mentioned, there are probably a bunch of factors providing reasons why even a simple 8400GS could manage such a performance boost as a dedicated PhysX card.

1: Relieves the primary renderer (In my case and cmdrdredd's, a GTX295) and frees up all resources that would be used to run PhysX. That would include any sp's used for processing, shader cache memory as Ben suggested, Memory controller stalls when accessing the cards Physical memory.

2: Moves all of these functions to another card that has all of it's own resources. The GTX295 has to share nothing. PhysX is completely taken over by a card that does not have to access sp's, cache, memory of the GTX295 leaving it completely free to do what it needs to do.

To test the 8400GS again, I'll need to revert my drivers to the latest 180 series for it to work. After the 190s, anything less than 32 shaders will not work and will automatically use the GTX295 for PhysX processing. There will be no option to select the 8400GS in NV control panel with drivers 190 and beyond. I asked Nvidia about this when the 190s came out. I told them I couldn't use the 8400GS any longer for PhysX. They asked me if I had found that the 8400GS really helped in PhysX performance. I referred them to my Cryostasis benches Using the 295 and the 8400GS. So maybe in a driver down the road, less than 32 sp's will be allowed to function. Up to them of course, i can only provide suggestions.

I was actually considering picking up one of these for a PhysX sweet spot, if I can determine that there is any substantial increase in PhysX performance going from a 16 shader 8400GS to a 32 shader 8600GT, to a 96 shader 8800GTS 640 to a 128 shader 8800GTS 512. I wish I had a 64 shader 9600GT on hand to test in this mix as well. But I think I have enough cards throughout the spectrum to come up with a semi-educated guess.

 

[Keysplayr]

Starting a new PhysX test thread.

 

[Z3RiN]

I actually just ran into this - I had offloaded to the built in display on my nForce motherboard but it wasn't up to the task. I got lower framerates in Batman (v1.1 retail is an excellent PhysX title, BTW) until I assigned PhysX to run from my main cards (2 x 285GTX2G)

I posted the whole process of testing on my website:
http://www.synthetic-x.com/200...ance-using-batman.html

I also have some FRAPs video of Batman running in Normal and High settings. It's just one quick scene with leaves right now but you can definitely see a difference.

 

[cmdrdredd]

Originally posted by: nOOky
You did the right thing. A dedicated second card will probably never, ever be noticeable in any game. One good main card will be able to handle PhysX and the rest. It will eat up more wattage and create more heat in your case, as well as be more a burden to the power supply. Probably the best bet is to oc the main card a bit, sell the other one and buy a copy of Windows 7 or something with the proceeds

there's a thread that shows a pretty sizeable difference in some games such as Batman: Arkham asylum. Others, not so much but drivers can be better and I'm sure over time they will be.

Generally no, I would say a GTX295 would be good enough alone, but if you have the card already, maybe try.