DosFreak wrote on 2025-02-19, 23:28:

What is the minimum GPU needed for PhysX and what GPU is overkill just for Physx
For the minimum GPU are drivers still supported and if so when is EOL?

Funny you should ask this, I was just researching a similar question about a month ago. I wanted to find a low-power secondary GPU for PhysX since my primary is a Radeon. I managed to get my hands on a Quadro K620 for the low price of $0. The card is EOL, but has Windows 11 compliant WDDM 2.0 drivers. Anything older/slower probably won't work. Luckily, even this card managed to double my fps in Fluidmark (~20 fps to ~40 iirc, not great, but much better than CPU only).

Use Fluidmark for benchmarking Physx: https://www.ozone3d.net/benchmarks/physx-fluidmark/

I was curious how bad the impact of CPU physx is going to be on a recent CPU and it's pretty bad, here's Arkham in comparison:
https://www.youtube.com/watch?v=mJGf0-tGaf4
And some games don't even allow full CPU physx apparently.

As far as I've understood, it's just been deprecated at a driver level.

Since this is the forum for making old stuff work on newer hardware/software, isn't there a way to force 32bit CUDA on newer 5000 series cards?
Maybe we can also finally get some kind of wrapper or translation layer that makes that thing work on non nVidia GPUs.

I never owned an high end nVidia GPU, and I currently run an RX5700 in my main rig, thus I never experienced Hardware Accelerated PhysX at more than a couple of frames per second.
Thankly it was never mandatory in games that supported it, so you could always revert to a lighter (and less impressive) software solution.

Digital Foundry covered this in a recent video.

They did a brief analysis using Arkham Asylum and it doesn't look good at all. That said, they did showcase a workaround where a compatible secondary card (e.g. an RTX 3060) can be used purely for Physx processing, while the main card renders the rest. But still, that's not an excuse for Nvidia dropping this for no good reason.

I'm a little curious why they didn't make a Cuda implementation (ie software) while they dropped the hardware.

I guess it wasn't used enough to justify

Putting a second card for PhysX is mostly impossible nowadays. Modern GPUs are already taking 3-4 slots and a lot of people shifted to Micro-ATX/ITX form factors. Also you can bet that current DLSS/DLAA will be just as useless somewhere after 10 years down the line.

myne wrote on 2025-03-01, 18:06:

I'm a little curious why they didn't make a Cuda implementation (ie software) while they dropped the hardware.

I guess it wasn't used enough to justify

PhysX had a software version that ran along side the hardware accelerated. The issue is that high settings are too much for even the best CPUs. And some games only allow their extreme settings is hardware acceleration is enabled.

This is disappointing, but really no different from the killing off of hardware accelerated sound with the XP to Vista switch. That ended A3D, and while OpenAL extended it a bit, it was also the end of the road for EAX.

I don't see why a wrapper to use OpenCL instead couldn't be created, it's just until now there wasn't much point. If you wanted all the fancy effects you bought an Nvidia card. So there just wasn't a need.

Yeah, I know about the CPU code. Which is why I'm surprised they didn't convert it to Cuda code. It's not likely to cost any noticeable performance

myne wrote on 2025-03-01, 19:08:

Yeah, I know about the CPU code. Which is why I'm surprised they didn't convert it to Cuda code. It's not likely to cost any noticeable performance

PhysX's hardware accelerated mode has been using 32-bit CUDA from all the way back in 2008 when Nvidia acquired Ageia. Only Nvidia cards can run CUDA, so why would anyone convert anything?

Nvidia also had the SDK source available since 2015, with a switch to the 3-clause BSD license in 2018. A wrapper to OpenCL should be possible, and that would allow current, and future cards that also aren't Nvidia to be able to run. But as mentioned, there just wasn't a need for someone to develop anything before now.

https://github.com/NVIDIA-Omniverse/PhysX

First Ageia and later Nvidia deliberately made CPU implementation slow as hell, running on 1 thread and pure x87.

Gamers Nexus released a video comparing the GTX 580 against the RTX 5080.

Yeah, the 580 wins when PhysX is involved.

More reason for Windows 7 rig heh.

The Serpent Rider wrote on Yesterday, 16:08:

More reason for Windows 7 rig heh.

I hadn't plans to build it for a few years... But now I'm gonna do it by the summer. The XP build is basically finished (the new GPU is on its way), and I can put some focus on the Win7 box.

There's an LGA 1150 system (with H81 chipset) in the closet in my office, it was of my boss. He alredy said that I can have it because, otherwise, it would end up in a landfill the second he wants to free up space in that closet.

And mind you, I never cared about PhysX for one main reason: I never owned a fast enough nVidia card to use it. A GT650m was never fast enough to run PhysX in games without them running at 10fps, and after that I only had Radeon GPUs that, just the same, would run those games at 10fps with PhysX enabled.
So I'm not doing it because I can't miss PhysX in those games. It's just that I have an hunch that those cards capable of running those PhysX games are gonna ramp up in prices in the near future, so I'm gonna play in advance.

I actually even have a 970 closeby, my younger brother has a PNY one, with a blower cooler. But he won't sell it even to me. It was his first dedicated GPU, and he ran it for almost 8 years, so he's sentimentally attached to it.
Meaning I'm alredy looking at GTX 10 and 16 series cards. I'll probably stay away from RTX cards for this build.

Minimum viable card for PhysX on RTX 50 is Maxwell 2.0/Pascal (for now), because there can't be two NV installed at the same time.
Windows 7 (64-bit), also only supports up to RTX 30 max, so - no problems there 😉

on the one hand, anyone who uses a card lower than 4070ti for physx is a loser
on the other hand, gtx240/260 is more than enough
gtx275 physx edition

agent_x007 wrote on Yesterday, 17:03:

Minimum viable card for PhysX on RTX 50 is Maxwell 2.0/Pascal (for now), because there can't be two NV installed at the same time.

Yep
Card mast be minimal from support driver list

The GT 710 that I sold probably has a purpose now

BEEN_Nath_58 wrote on Yesterday, 19:16:

The GT 710 that I sold probably has a purpose now

The point is that different driver versions have the same file names.
Windows will not be able to install two drivers of different versions, they will conflict.

Therefore, for owners of 5xxx series video cards, only the 572 driver is suitable.
Its minimum supported family is 7xx

Supported Products
GeForce 700 Series
GeForce GTX 750 Ti, GeForce GTX 750, GeForce GTX 745

Pretty sure windows deals with different drivers for different cards fine if you manually set it