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

[DemonicRyzen666]

GitHub - NVIDIA-Omniverse/PhysX: NVIDIA PhysX SDK

NVIDIA PhysX SDK. Contribute to NVIDIA-Omniverse/PhysX development by creating an account on GitHub.

github.com

It was open-sourced and licensed under BSD 3 model around 3 years ago.

GPU accelerated physX isn't open source. You still need a Cuda supported graphics cards for GPU accelerated physX (which Nvidia won't license out the Cuda parts or a translation layer for Cuda), only CPU is open source. GPU accelerated physX also doesn't play nice/work with DX12 because of mGPU & some other low lever A.P.I changes from DX11 apparently.

 

[Dr. Dro]

?

Their chipsets ran hot and couldn't get FSB as high as comparable P35/P45/X38/X48 boards. Their only selling point was SLI capability.

I had a socket AM3 nForce, ASUS M4N75TD I believe. Worked relatively well for a low cost motherboard, ran a X6 1090T on it for a while. Unfortunately, I've no fond memories of that processor. Couldn't be slower if it tried vs. its contemporary Core i7 parts.

GPU accelerated physX isn't open source. You still need a Cuda supported graphics cards for GPU accelerated physX (which Nvidia won't license out the Cuda parts or a translation layer for Cuda), only CPU is open source. GPU accelerated physX also doesn't play nice/work with DX12 because of mGPU & some other low lever A.P.I changes from DX11 apparently.

I mean yes, CUDA isn't open source, but PhysX is. The problem is that since CUDA isn't, then PhysX remains restricted to NV unless it's completely refactored in order to run in any other compute API, like OpenCL or something.

 

[Athlonite]

I mean yes, CUDA isn't open source, but PhysX is. The problem is that since CUDA isn't, then PhysX remains restricted to NV unless it's completely refactored in order to run in any other compute API, like OpenCL or something.

Yeah using OpenCL for PhysX would work great if somebody could get it go on any GPU regardless of manufacturer as OpenCL is already doing alot of the stuff PhysX was designed to do anyways ie: Fluid dynamics and Particle Physics

 

[b1k3rdude]

Guys, Good news, GTX1030 is ENOUGH

Thanks for finding this!

 

[RootinTootinPootin]

will this work with thunderbolt gpu docks?

 

[b1k3rdude]

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

eh, the current driver still support the 900 series, so I would think the 1000 series has a while yet. But if and when they do then we have low profile 2060/3060/4060 to update too as those card get older and drop in price. Right now I can grab a 1030 for £28.

• https://whywelikethis.com/best-rtx-2060-low-profiles/#gsc.tab=0
• https://whywelikethis.com/best-rtx-3060-low-profiles/#gsc.tab=0

No, run dual video cards for gaming just for the PhysX thing. To be fair, I'm inclined to try it out. Sorry, I should have been more clear.

Same, specially when I can grab a 1030 for £28.

 

[lexluthermiester]

Same, specially when I can grab a 1030 for £28.

A 1030 seems a little lacking to me. A 1050 seems better idea.

 

[Selaya]

nonsense. a gt1030 is more than aplenty as a physx accelerator card.

 

[DemonicRyzen666]

nonsense. a gt1030 is more than aplenty as a physx accelerator card.

Fluidmark testing shows that it's barely faster than CPU, when the cpu isn't bogged down by other things during rendering. Increasing the emitters to 14 & particles to 256,000 on Fluidmark shows this issue.
The normal presets for 720p or 1080p for the benchmark cpu is fine, since it's not that many emitters or particles.

 

[Selaya]

yeah but the amount of ppl buying physx accelerator cards to run ... unrealistic benchmarks is probably exactly zero

 

[Dr. Dro]

eh, the current driver still support the 900 series, so I would think the 1000 series has a while yet. But if and when they do then we have low profile 2060/3060/4060 to update too as those card get older and drop in price. Right now I can grab a 1030 for £28.

• https://whywelikethis.com/best-rtx-2060-low-profiles/#gsc.tab=0
• https://whywelikethis.com/best-rtx-3060-low-profiles/#gsc.tab=0

Same, specially when I can grab a 1030 for £28.

I'm afraid not. The CUDA runtime support schedule is pretty reliable at predicting at how long driver support is going to hold up. These cards have 2-3 driver branches left at this point, and that's probably optimistic.

CUDA Toolkit 12.9 Update 1 - Release Notes — Release Notes 12.9 documentation

The Release Notes for the CUDA Toolkit.

docs.nvidia.com

• Architecture support for Maxwell, Pascal, and Volta is considered feature-complete and will be frozen in an upcoming release.

Once these are dropped, you won't be able to update your graphics driver for the 50 series card beyond what is commonly supported by the 10 series card as well

 

[Shou Miko]

It really depends on how much you need and want to spend because I can get A400 4GB for £170 brand new.

Is overkill, yes and a I don't want to talk about the T400 because it's only £160 so makes no sense going with a Turning based version when the Ampere ain't much more.

As nice as the GT 1030 is now that Nvidia is talking about dropping driver support for it it's not much good, leaving only T and A400 cards as a single slot option left.

 

[Viruzz]

It really depends on how much you need and want to spend because I can get A400 4GB for £170 brand new.

Is overkill, yes and a I don't want to talk about the T400 because it's only £160 so makes no sense going with a Turning based version when the Ampere ain't much more.

As nice as the GT 1030 is now that Nvidia is talking about dropping driver support for it it's not much good, leaving only T and A400 cards as a single slot option left.

Here is a better, more future-proof single slot option:

Yeston RTX 3050 6GB GDDR6​

 

[Shou Miko]

Here is a better, more future-proof single slot option:

Yeston RTX 3050 6GB GDDR6​

The RTX T and A400 are single slots and half height cards if you didn't knew.

The RTX T400 is Ampere and will properly be cheaper for a lot of people in europe than import Yeston's RTX 3050 sadly.

 

[Shihab]

Yeah using OpenCL for PhysX would work great if somebody could get it go on any GPU regardless of manufacturer as OpenCL is already doing alot of the stuff PhysX was designed to do anyways ie: Fluid dynamics and Particle Physics

OpenCL is a compute API. It isn't "designed to do" what PhysX does, per se. Not more than Intel SIMD intrinsics were. Sure, they make for a much faster implementation, but someone needs to put [a lot] of effort doing so. Especially if you're talking translation layers.

Besides, isn't OpenCL old tech at this point?
For more modern APIs, the 32bit requirement also seems to be an issue. See this discussion on the ZLUDA (project translating CUDA calls to vendor-agnostic APIs) related to PhysX.

CUDA implementation for PhysX · vosen ZLUDA · Discussion #31

This might seem not fitting, however I think here is the best place since ZLUDA is the only project I've seen so far that comes close to how I think it should work. More specifically - re-implement...

github.com

 

[b1k3rdude]

Once these are dropped, you won't be able to update your graphics driver for the 50 series card beyond what is commonly supported by the 10 series card as well

hmm, if I understand having support for the second card in the main driver is a desirable, but not a requirement for physx to work. As the once the card is dropped from the main driver windows will install the latest supported driver for that card and then presumably you can select that card for physx in the NvCpl...

 

[Dr. Dro]

hmm, if I understand having support for the second card in the main driver is a desirable, but not a requirement for physx to work. As the once the card is dropped from the main driver windows will install the latest supported driver for that card and then presumably you can select that card for physx in the NvCpl...

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

 

[InVasMani]

There are exceptionally few integrated GPUs that can run it, all very old. nForce has been gone for a very long time.

A dedicated card is the only option in most scenarios.

Yeah forgot all about the GPU side of it being proprietary for a moment was pretty tired. I mean technically if it weren't that it would probably make a nice option over CPU PhysX, but oh well. Best option is just using whatever option of two Nvidia GPU's pair best for PhysX or rendering. As to which GPU to prioritize to one or the other is likely not as straight forward as one might presume since the complexity of PhysX and rendering tasks really can vary heavily as well hardware capabilities for either task with GPU's. It's really a bit of a test situation and having two GPU's of similar performance characteristics, but well balanced is probably ideal compared to two that are more extreme opposites in design scope.

Actually I believe what I meant was use a APU or iGPU provided it's reasonably strong enough to do primary rendering, but using a Nvidia GPU for the PhysX could lead to some interesting results.

 

[lexluthermiester]

nonsense. a gt1030 is more than aplenty as a physx accelerator card.

Regardless, it still seems wise to go for a somewhat better card to allow for more headroom.

Here is a better, more future-proof single slot option:

Yeston RTX 3050 6GB GDDR6​

Like this one!

 

[Ninja Weedle]

Regardless, it still seems wise to go for a somewhat better card to allow for more headroom.


Like this one!

I also should note that GT 1030s are weirdly expensive for what they are- I see 1050s that go for lower than 1030s at times on ebay, which is why I recommend the Quadro P620- they're readily availiable for 30$ or less, have more features than the 1030 and are faster (like a theoretical GT 1040), are low profile so they wont block your fans, and they offer 4 video outs in case you need them. If you're gonna run a second card, might as well run something with more features for less money.

A GTX 1650 is the next cheapest option that wont be affected by the driver cutoff. a Yeston 3050 is a nice card but 200$ for physx is a bit ridiculous.

 

[lexluthermiester]

which is why I recommend the Quadro P620- they're readily availiable for 30$ or less

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

 

[agent_x007]

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

There are exceptionally few integrated GPUs that can run it, all very old. nForce has been gone for a very long time.

A dedicated card is the only option in most scenarios.

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).

 

[lexluthermiester]

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

Yes you can. You have to install The Blackwell drivers first and then install the Kepler drivers directly from the Device Manager without using the NVidia installer.

 

[agent_x007]

Yes you can. You have to install The Blackwell drivers first and then install the Kepler drivers directly from the Control Panel without using the NVidia installer.

Which still replaces Blackwell's driver files during installation (because both occupy the same folders and files are named the same in Windows driver library), and will cause a initialization failure of said card after restart.
Limitations :
1) Kepler driver does not know what Blackwell is and vice-versa,
2) You can only use one NV driver version concurrently/at the same time (it will be used on all NV cards present in PC).

 

[lexluthermiester]

Which still replaces Blackwell's driver files during installation (because both occupy the same folders and files are named the same in Windows driver library), and will cause a initialization failure of said card after restart.
Limitations :
1) Kepler driver does not know what Blackwell is and vice-versa,
2) You can only use one NV driver version concurrently/at the same time (it will be used on all NV cards present in PC).

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.