Recommended PhysX card for 5xxx series? [Is vRAM relevant?]

[Dr. Dro]

You can't pair Kepler with Blackwell, so it's moot point.

Are there any iGPUs that can run it ?
I know ION can't, because 32 Stream Processors (Tesla generation) are required as minimum for GPU to be PhysX capable (aside from 256MB of VRAM).

GeForce 320M makes that cut, barely (256 MB, 48 CUDA cores)

 

[Veseleil]

GeForce 320M makes that cut, barely (256 MB, 48 CUDA cores)

Your sig is making me anxious.

 

[LabRat 891]

Yes, it has always been enough. The problem is when Nvidia drops 10 series support, which is any day now.

Sorry if already answered...
What about Quadro cards?

Lots of single slot and low-profile Pascal and newer Quadros out there, cheap.

edit:

Where are you getting that info? Asking because I am currently running two different cards(Geforce and Quadro) in one of my compute systems. The Geforce does GFX, the Quadro does compute. Granted, My Geforce is Ada and the Quadro is Maxwell, but still.

I see. I'm considering putting an RX6500XT8GB into my X570 x4, for 0-cost lossless scaling, multi-monitor, and to fiddle w/ LLM stuff on. (free up the 'big' card).
'Wondering if a Quadro might be a better choice? PhysX was never a 'big deal' for me, but there are 1 or 2 titles I really loved that used it.

Also, can't you pass frames to the nV card for CUDA/DLSS things?

 

[lexluthermiester]

'Wondering if a Quadro might be a better choice? PhysX was never a 'big deal' for me, but there are 1 or 2 titles I really loved that used it.

That is an interesting question. Quadro's are good number crunchers for sure. For AI work, you'd need something fairly recent, Turing or better is what I would suggest. However, from what I understand anything RDNA and better in the AMD camp would be a solid choice as well. So for AI, you're not limited to NVidia. Intel Arc cards are very good for that from what I understand.

Also, can't you pass frames to the nV card for CUDA/DLSS things?

I haven't played around with it too much. I've been mostly science stuff with my setup. Maybe? In line with the topic of this thread, I'm going to be playing with the setting in the NV control panel that let a user assign the PhysX card specifically. The CUDA settings are selectable as well. I'll chime in with the results.

 

[agent_x007]

Where are you getting that info? Asking because I am currently running two different cards(Geforce and Quadro) in one of my compute systems. The Geforce does GFX, the Quadro does compute. Granted, My Geforce is Ada and the Quadro is Maxwell, but still.

Myself. Using driver that doesn't support all gen of cards used, always failed in some way when I tried.
Note : Different cards marketing name isn't enough. You have to be outside of generation support for both in drivers to see this problem. It is VERY limited issue.
R340 branch of NV driver supports Tesla and up to Kepler.
R391 branch of NV driver supports Fermi and up to GTX 10-series.
R47x branch can support Kepler and up to RTX 30.
Current driver still get's you GM107 and up to RTX 50 support.

Issue I wrote about earlier should occur only with :
1) GTX 900/10-series and newer being paired with Tesla-class GPUs,
2) RTX 20/30 and newer being paired with Fermi/Tesla,
3) RTX 40/50 being paired with Tesla/Fermi/Kepler class cards.

 

[LabRat 891]

Myself. Using driver that doesn't support all gen of cards used, always failed in some way when I tried.
Note : Different cards marketing name isn't enough. You have to be outside of generation support for both in drivers to see this problem. It is VERY limited issue.
R340 branch of NV driver supports Tesla and up to Kepler.
R391 branch of NV driver supports Fermi and up to GTX 10-series.
R47x branch can support Kepler and up to RTX 30.
Current driver still get's you GM107 and up to RTX 50 support.

Issue I wrote about earlier should occur only with :
1) GTX 900/10-series and newer being paired with Tesla-class GPUs,
2) RTX 20/30 and newer being paired with Fermi/Tesla,
3) RTX 40/50 being paired with Tesla/Fermi/Kepler class cards.

Interesting. I wonder if OS is the issue?
Was that all attempted on 10? I assume Lex is on Win11 (for windows OS), and I've seen Win11 'plays nicer' with multiple GPUs than 10

 

[agent_x007]

This was under Windows 7 and 10, I don't use 11.
Note : I didn't tested all likely failed configuration options on both OS'es.

 

[lexluthermiester]

Myself. Using driver that doesn't support all gen of cards used, always failed in some way when I tried.
Note : Different cards marketing name isn't enough. You have to be outside of generation support for both in drivers to see this problem. It is VERY limited issue.
R340 branch of NV driver supports Tesla and up to Kepler.
R391 branch of NV driver supports Fermi and up to GTX 10-series.
R47x branch can support Kepler and up to RTX 30.
Current driver still get's you GM107 and up to RTX 50 support.

Issue I wrote about earlier should occur only with :
1) GTX 900/10-series and newer being paired with Tesla-class GPUs,
2) RTX 20/30 and newer being paired with Fermi/Tesla,
3) RTX 40/50 being paired with Tesla/Fermi/Kepler class cards.

I'm current doing it. So whatever. Granted, I'm not running Blackwell, but by your statement that wouldn't matter.

This was under mostly under Windows 7 and 10, I don't use 11.

I'm running 11.

On that note, this actually works. I installed a Quadro K2200 into my gaming rig, configured the settings and ran a game I know uses PhysX(Witcher3) and the game was using the Quadro for that functionality.

Qualifier though, it didn't improve my frame rate much, margin of error kind of thing.

However, this seems a very doable thing. Does anyone with an RTX5000 card want to give this a go?

EDIT: Also just tried XCom Declassified and it seems to work there too. The K2200 is being utilized. Also no framerate improvement, but it ran.

 

[Dr. Dro]

Sorry if already answered...
What about Quadro cards?

Lots of single slot and low-profile Pascal and newer Quadros out there, cheap.

edit:

I see. I'm considering putting an RX6500XT8GB into my X570 x4, for 0-cost lossless scaling, multi-monitor, and to fiddle w/ LLM stuff on. (free up the 'big' card).
'Wondering if a Quadro might be a better choice? PhysX was never a 'big deal' for me, but there are 1 or 2 titles I really loved that used it.

Also, can't you pass frames to the nV card for CUDA/DLSS things?

In terms of relative time it might last a little longer, but that's because Quadro branches are longer lived. They'll be the same driver, just supported a while longer. Best thing to do is buy something with the Turing or Ampere architecture. As for the 6500 XT... been there. I was going to buy one to hold me over until the 5090 arrived, and then I found this RTX A2000 I'm using right now at more or less the same price. It's very much a cut above, and also runs on slot power. Shop around and you may just find something that interests you.

As for the last question, most games disable all CUDA-related functionality if the primary adapter's vendor is not Nvidia. Very long time ago, when @Regeneration still ran the NGOHQ forum, they had come up with a tool that would enable dedicated PhysX with a Radeon primary GPU, but it required a very specific set of drivers. I recall using it with my HD 5970 back in the day. Not sure if anything like it still exists but they might know.

Myself. Using driver that doesn't support all gen of cards used, always failed in some way when I tried.
Note : Different cards marketing name isn't enough. You have to be outside of generation support for both in drivers to see this problem. It is VERY limited issue.
R340 branch of NV driver supports Tesla and up to Kepler.
R391 branch of NV driver supports Fermi and up to GTX 10-series.
R47x branch can support Kepler and up to RTX 30.
Current driver still get's you GM107 and up to RTX 50 support.

Issue I wrote about earlier should occur only with :
1) GTX 900/10-series and newer being paired with Tesla-class GPUs,
2) RTX 20/30 and newer being paired with Fermi/Tesla,
3) RTX 40/50 being paired with Tesla/Fermi/Kepler class cards.

This is exactly correct, and these are the "branch gates" for mixed architectures. Release 570 drivers still support Maxwell-Blackwell but, the next branch might not.

Your sig is making me anxious.

Qiqi is love. Qiqi is life.

 

[LabRat 891]

As for the last question, most games disable all CUDA-related functionality if the primary adapter's vendor is not Nvidia. Very long time ago, when @Regeneration still ran the NGOHQ forum, they had come up with a tool that would enable dedicated PhysX with a Radeon primary GPU, but it required a very specific set of drivers. I recall using it with my HD 5970 back in the day. Not sure if anything like it still exists but they might know.

I am yet again reminded of, how very much 'legal protection' high-tech companies like nVidia get, strictly from the Technical Complexity of a given issue.
Not unlike like nForce 2/3 actively gimping ATI cards, that's flat-out anti-competitive. -and, good luck ever proving that the choice was anything other than 'technical and reliability concerns'

 

[Dr. Dro]

I am yet again reminded of, how very much 'legal protection' high-tech companies like nVidia get, strictly from the Technical Complexity of a given issue.
Not unlike like nForce 2/3 actively gimping ATI cards, that's flat-out anti-competitive. -and, good luck ever proving that the choice was anything other than 'technical and reliability concerns'

Mmm, I haven't been around that long , so I never really had the chance to play around with the socket 462 or 754 stuff, but from what I could read is that Nvidia chose not to update the nForce 2/3 boards to support Vista correctly and that generated a lot of compatibility problems.

As I understand it, though, the tool I mentioned earlier simply patched the PhysX driver so that it wouldn't do a vendor check on the primary adapter so using a dedicated card for PhysX was possible. It's been too long, I don't remember any of it really.

 

[b1k3rdude]

AFAIK 2 different versions of the same vendor's driver cannot coexist on the same PC, I've seen it done with AMD GPUs before, but it involved driver mods and stuff - unsure if would work at all with NV

I believe it can, as I am fairly sure that I have seen the MS driver and the nV driver coexisting on the same machine. I will verify this by installing an old GT710 into my desktop machine and see what device manager says.

That's a good one! Even a K2200 would be about the same 50% faster than a 1030 and those are about the same price.

GeForce GT 1030 vs Quadro K2200 vs Quadro K1200 vs Quadro P620 [videocardbenchmark.net] by PassMark Software

The perf seem to vary from marginal to noticable -

- https://gpu.userbenchmark.com/Compare/Nvidia-Quadro-K2200-vs-Nvidia-GT-1030/2839vsm283726
- https://www.videocardbenchmark.net/...s-Quadro-K2200-vs-Quadro-K1200-vs-Quadro-P620

And if its an issue for some, the K2200 uses over double the power of the 1030, that said I am finding K2200's on ebay for dirt cheap....

 

[chrcoluk]

I feel driver support is important in case of new major feature or for future OS.
But I dont usually keep my drivers up to date as is often regressions, I only yesterday downgraded back to 551,86, because everything after 55x on drivers has problems.

If using GT 1030, I dont think there is a need for the latest driver. I cant think of any improvement that would benefit the card, especially if its only used for physx purposes. the problem will be future generations of cards where they have a min supported version, and that version no longer supports 10 series.

Seen the spreadsheet, my 1030 GT currently idle, might stick in the PC now.

 

[Dr. Dro]

I believe it can, as I am fairly sure that I have seen the MS driver and the nV driver coexisting on the same machine. I will verify this by installing an old GT710 into my desktop machine and see what device manager says.


The perf seem to vary from marginal to noticable -

- https://gpu.userbenchmark.com/Compare/Nvidia-Quadro-K2200-vs-Nvidia-GT-1030/2839vsm283726
- https://www.videocardbenchmark.net/...s-Quadro-K2200-vs-Quadro-K1200-vs-Quadro-P620

And if its an issue for some, the K2200 uses over double the power of the 1030, that said I am finding K2200's on ebay for dirt cheap....

Let me phrase this better, two drivers from the same vendor. Having two vendors' drivers present at once is not uncommon, most laptops and desktops with integrated graphics will do this, and laptops will even often use their dGPU headless (as is the case with mine). For example, my laptop has both a Radeon and a GeForce, and requires both AMD's and NV's driver to be installed. The Radeon iGPU is connected to the panel, and the RTX dGPU operates headless, with everything passing through the AMD GPU for display, unless the laptop is outputting to the HDMI port.

But for example, having the 566.36 and the 572.12 drivers both installed and each instanced at one different card, I'm not sure that configuration works. I've seen modified AMD drivers coexisting with official AMD drivers, for example, but that still isn't a feasible or supported configuration as far as I am aware.

 

[b1k3rdude]

But for example, having the 566.36 and the 572.12 drivers both installed and each instanced at one different card, I'm not sure that configuration works. I've seen modified AMD drivers coexisting with official AMD drivers, for example, but that still isn't a feasible or supported configuration as far as I am aware.

So it seems I was mistaken, I thought for sure this type of config would work. So as you rightly pointed out the driver have to support both cards, so nVreedia and Micro$haft could not place nice in the futere and remove all support for anything older than 1-2 generations. I just tried a GT710 and my RTX4080 and no matter what I tried, I could not have 2 different drivers. Would a quadro K2200 work as it has 570.xx support (but via the studio driver.) Looking on the driver page the Studio/GRD drivers don't officially have any crossover...

It does beg the questions of how the hell did this person managed to get a 4090 and a 750Ti working in the same system -

https://www.reddit.com/r/nvidia/comments/1j4lxgh

 

[Dr. Dro]

So it seems I was mistaken, I thought for sure this type of config would work. So as you rightly pointed out the driver have to support both cards, so nVreedia and Micro$haft could not place nice in the futere and remove all support for anything older than 1-2 generations. I just tried a GT710 and my RTX4080 and no matter what I tried, I could not have 2 different drivers. Would a quadro K2200 work as it has 570.xx support (but via the studio driver.) Looking on the driver page the Studio/GRD drivers don't officially have any crossover...

It does beg the questions of how the hell did this person managed to get a 4090 and a 750Ti working in the same system -

https://www.reddit.com/r/nvidia/comments/1j4lxgh

Quadro K is Kepler just like the GT 710, you will run into the same issue, if you had a Quadro M (Maxwell) it would work as that is identical to the 900 series architecturally. The 750 Ti exclusively is also Maxwell, so it is still supported. The other 700 series cards are all Kepler

 

[lexluthermiester]

And if its an issue for some, the K2200 uses over double the power of the 1030, that said I am finding K2200's on ebay for dirt cheap....

The power usage is not a something I care about. However, for those who are concerned, the K1200, P620 and P1000 are all single slot, sub 60W cards and will do the job just as well or better.

GeForce GT 1030 vs Quadro K2200 vs Quadro K1200 vs Quadro P620 vs Quadro P1000 [videocardbenchmark.net] by PassMark Software

 

[tugrul_LordOfDrinks]

I kept my 4070 as an extra compute power and a physx device and bought a 5070 since the money did not reach 5070 ti. I currently get nearly 15k points in kombustor physx benchmark. No overclock. 144 simulations per second, pcie 4.0 x4 slot fir 4070. 500-600 frames per second on 5070.

 

[ShrimpBrime]

It's not a GPU. But it does pretty good for physx as it was designed. But was never released.

NVidia PPU2 physx adapter.

 

[agent_x007]

Would a quadro K2200 work as it has 570.xx support (but via the studio driver.) Looking on the driver page the Studio/GRD drivers don't officially have any crossover...

It does beg the questions of how the hell did this person managed to get a 4090 and a 750Ti working in the same system

GTX 750 Ti (GM107) and 4090 can use the same (newest) driver, since they are Maxwell 1.0 and Ada.
Also, both Quadros (K1200 and K2200) use GM107 as base (so again, it's Maxwell 1.0).

 

[b1k3rdude]

GTX 750 Ti (GM107) and 4090 can use the same (newest) driver, since they are Maxwell 1.0 and Ada.
Also, both Quadros (K1200 and K2200) use GM107 as base (so again, it's Maxwell 1.0).

Huh, well I'll be a monkeys uncle. I didnt think driver support went that far back, noice!

 

[LabRat 891]

It's not a GPU. But it does pretty good for physx as it was designed. But was never released.

NVidia PPU2 physx adapter.

Is good 'ole G92 still supported at all for CUDA/PhysX? The cooler on that PPU2 brought me back.

 

[ShrimpBrime]

Is good 'ole G92 still supported at all for CUDA/PhysX? The cooler on that PPU2 brought me back.

If it has Cuda, it does physx. 8800 would work for physx, but probably not as well as a GT 710 or something small and low power I'd imagine.

 

[demirael]

And if its an issue for some, the K2200 uses over double the power of the 1030, that said I am finding K2200's on ebay for dirt cheap....

I wonder where that info is from, I'm using a K2200 (that I had lying around) as a dedicated PhysX card along my AMD GPU, and both nvidia-smi and GPU-Z report the board power to be 39W. (Idles at 1W, which is nice.)

 

[lexluthermiester]

I wonder where that info is from, I'm using a K2200 (that I had lying around) as a dedicated PhysX card along my AMD GPU, and both nvidia-smi and GPU-Z report the board power to be 39W. (Idles at 1W, which is nice.)

At full load, the K2200 can use up to 90W, which is more than slot power specs and why most models have a PCIe power connector, including mine. It has to be pushed hard to do that though. PhysX shouldn't be enough.