• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

RTX 50 ends 32 bit phyx support

ChadD

[H]F Junkie
Joined
Feb 8, 2016
Messages
8,407
https://www.tomshardware.com/pc-com...ries-gpus-nvidia-ends-32-bit-cuda-app-support

I know this is older tech now. Guess it shouldn't be too shocking support has ended. It does make me wonder what exactly changed hardware wise that made supporting it on blackwell an issue.
Still some pretty popular games around that still use it, even if they are pushing a decade old now.
Toms lists these ones anyway; Batman Arkham trilogy, Borderlands: The Pre-Sequel, Borderlands 2, Metro: Last Light, Metro: Exodus, Metro 2033, Mirror's Edge, The Witcher 3, and some older Assassin's Creed titles.
 
I wonder if doing it on the CPU on modern systems helps make up for the fact that the GPU Now doesn't accelerate it anymore
Apparently still a bit too slow on some 32 bits only title.

Looking at it, a bit seems like it was a long time coming:
Note that since CUDA v6.5, only 64-bit code is supported (x64 architecture).

https://developer.download.nvidia.com/compute/cuda/6_5/rel/docs/CUDA_Toolkit_Release_Notes.pdf
(Linux) Support for 32-bit Applications on x86-based Linux DistributionsSeveral portions of the CUDA Toolkit are no longer available for developing 32-bitapplications on x86-based Linux distributions:

2017:

Note that as of v10.0 CUDA SDK no longer supports compilation of 32-bit applications.

6.5 that was around 2014, 970 gtx, they took a decade to slowly completely kill it "gently", less than 4GB only vram gpu stopped being a thing quite a while ago.

Wonder how much hardware is saved or if it is driver-compiler, etc... side mostly, maybe it would have needed some hardware something to still support 32 bits, like it happened on many cpu for many older bits affair that run in virtual machine.

Batman Arkham trilogy

Some of the later one could be 64 bits application, at least the 8GB of ram recommended on steam and 64 bits operating system required in the specs list make it sound like it is.
 
I wonder if doing it on the CPU on modern systems helps make up for the fact that the GPU Now doesn't accelerate it anymore
Apparently from the articles I read on it, no, the performance is still pretty bad. Though the last time I did any real reading on CPU accelerated PhysX was almost 15 years ago and at the time it was written using single threaded x87 code, but it's highly unlikely the CPU code has improved.
 
Some of the later one could be 64 bits application, at least the 8GB of ram recommended on steam and 64 bits operating system required in the specs list make it sound like it is.
The game itself is 64bit. Same with games like borderlands 2. but the Physx implementation is still 32bit.
I don't know if this is a big deal to most people, just an interesting new quirk for the 5000 cards. Retro gamers that might have wanted to play some of the old Physx titles with that setting maxed out won't be doing that.

I am sure no one is buying a new high end GPU to play 10 year old titles only... it is interesting though. What makes me curious is clearly its a hardware level issue. I mean support has not ended for 4000 or older cards. If you have the latest drivers you can use GPU physx on your 4090, but not on a 5090. Clearly whatever changed is internal, not just a removal of 32bit software libraries.
 
Apparently from the articles I read on it, no, the performance is still pretty bad. Though the last time I did any real reading on CPU accelerated PhysX was almost 15 years ago and at the time it was written using single threaded x87 code, but it's highly unlikely the CPU code has improved.
Yes even the big monster 9950x type chips are going to perform a lot worse then even a 5060 with Physix enabled would. I mean its a Nvidia feature, I'm not sure how popular its been anymore. Seems like game devs have basically abandoned it. Interesting quirk of the new hardware.
 
The game itself is 64bit. Same with games like borderlands 2. but the Physx implementation is still 32bit.
I was looking by mistake at Batman Knight (2016)
The 3 2013 or before show up there:
https://www.resetera.com/threads/rt...mpacted-mirrors-edge-borderlands-etc.1111698/

They seem all 32 bits, that older trilogy.

Boderland 2 was 32 bits I think, at least at launch, still run on a 32bits os according to the steam specs:

View: https://store.steampowered.com/app/49520/Borderlands_2/
 
Last edited:
i was going to post this. Not surprising. Mainly was thinking of BL2 and Mirrors Edge as the ones I remember using it for specifically.
 
https://www.tomshardware.com/pc-com...ries-gpus-nvidia-ends-32-bit-cuda-app-support

I know this is older tech now. Guess it shouldn't be too shocking support has ended. It does make me wonder what exactly changed hardware wise that made supporting it on blackwell an issue.
Still some pretty popular games around that still use it, even if they are pushing a decade old now.
Toms lists these ones anyway; Batman Arkham trilogy, Borderlands: The Pre-Sequel, Borderlands 2, Metro: Last Light, Metro: Exodus, Metro 2033, Mirror's Edge, The Witcher 3, and some older Assassin's Creed titles.
Many of the games on that list are pushing 20 years old.
I like to think that it is still 2015 and that I'm a decade younger, too. :oldman:
 
Last edited:
Disappointing. I overall like the additional debris and particle affects in the Ghost Recon Advanced Warfigher games. Just replayed the 1st game a year or two ago, was wanting to do another run with GRAW 2. I felt like it was overdone in Batman Arkham Asylum but still preferred it on over it being off. Kind of disappointing if these have to be disabled.

Any particular reason Nvidia decided to drop support?
 
Disappointing. I overall like the additional debris and particle affects in the Ghost Recon Advanced Warfigher games. Just replayed the 1st game a year or two ago, was wanting to do another run with GRAW 2. I felt like it was overdone in Batman Arkham Asylum but still preferred it on over it being off. Kind of disappointing if these have to be disabled.

Any particular reason Nvidia decided to drop support?
They are cleaning out legacy to lighten it up, the CUDA library is getting big and there is lots of logic being dedicated to it in the silicon and it’s getting expensive to maintain. I suspect that they will start taking out hardware support for legacy DX and OpenCL versions soon, probably go the Intel route and use translation libraries to maintain compatibility at the expense of some performance. Their AI should be good enough to help smooth out that mapping for them at the very least.

Getting rid of 32-bit CUDA and the old graphics API’s could probably shave a good 100 mm2 off the chips.
 
They are cleaning out legacy to lighten it up, the CUDA library is getting big and there is lots of logic being dedicated to it in the silicon and it’s getting expensive to maintain. I suspect that they will start taking out hardware support for legacy DX and OpenCL versions soon, probably go the Intel route and use translation libraries to maintain compatibility at the expense of some performance. Their AI should be good enough to help smooth out that mapping for them at the very least.

Getting rid of 32-bit CUDA and the old graphics API’s could probably shave a good 100 mm2 off the chips.
This is logical and all...
Its still pretty funny in its own way if they do lean on AI. Rendering old games wrong, and trying to make it work via ai. I mean they do it for new games why not start doing the same with the old stuff. ;)
You make a good point though. Intel out of the gate choose to not add hardware bits for older games. Obviously a bigger deal 2-3 years ago then now... and it will matter even less in a year or two. I know this is just a decade old NV exclusive thing. The next step will be dropping even more hardware. I imagine AMD when they move do UDMA will do the same thing. As long as they get enough of a bump in performance translation stuff should be pretty unnoticeable. Intels implementation wouldn't have gotten much flak if it wasn't crashy at launch. As long as the latest games benchmark well... who really cares if old games run at 200fps instead of 250.
 
Getting rid of 32-bit CUDA and the old graphics API’s could probably shave a good 100 mm2 off the chips.
It's for software maintenance and software support not the chips. The instruction set is teeny tiny. It's a subset. The bulk of the size is the cache, number of cores, packaging and memory interconnect.
 
Witcher 3 will be 10 years old in may. 2015 was 10 years ago. :)
However no doubt we are both old. lol
Crazy that it's one of the newest games in that list, and does not seem like it has been a decade.
 
  • Like
Reactions: ChadD
like this
Apparently from the articles I read on it, no, the performance is still pretty bad. Though the last time I did any real reading on CPU accelerated PhysX was almost 15 years ago and at the time it was written using single threaded x87 code, but it's highly unlikely the CPU code has improved.
I thought I recall something back then about how if they had just compiled it differently than x87 the code would of ran much faster on CPU. They suspected Nvidia intentionally keeping it as x87 so you had to use a GPU to get decent performance from it.
 
  • Like
Reactions: Rizen
like this
It is so funny as I own Metro: Exodus, Metro 2033, Just started testing the game on Intel
 
Last edited:
If there is a second PCIe or M.2 slot one could use a second 40 series GPU or lower and just dedicate it to PhysX.
What's the problem making a 32-bit PhysX driver to support these games? Think I remember saying something years ago how proprietary standards end up dying just like this and this sort of thing ends up happening. Someone is probably going to make software that allow PhysX to work on any GPU, but probably on Linux.
 
Someone is probably going to make software that allow PhysX to work on any GPU
Think about the kind of millions in effort in the last decade to make CUDA work on any GPU.... PhysX running on GPU a death tech would be the push ;) ?

What's the problem making a 32-bit PhysX driver to support these games?
A simple 960 has more performance than those old PhyX Ageia card I think, did they made windows 10 drivers for those old card ?
 
Think about the kind of millions in effort in the last decade to make CUDA work on any GPU.... PhysX running on GPU a death tech would be the push ;) ?


A simple 960 has more performance than those old PhyX Ageia card I think, did they made windows 10 drivers for those old card ?
Oddly enough that is what I have been considering if I upgrade to a 5000 series card. I have a GTX 960 2GB in my wife's computer but she doesn't game at all. However, I think for the first time since the nVidia Riva I may fo AMD and get a 9070xt to upgrade my RTX3070.
 
I wonder if doing it on the CPU on modern systems helps make up for the fact that the GPU Now doesn't accelerate it anymore

The only thing that's dead is old ass Physx code.

Probably every major modern engine at this point is already shipping GPU particle sims that will happily render millions like it was nothing.
 
Think about the kind of millions in effort in the last decade to make CUDA work on any GPU.... PhysX running on GPU a death tech would be the push ;) ?
They already did, and it's called Zluda. It's not that it doesn't work but that AMD is afraid of legal action by Nvidia. From what I remember it works faster than AMD's own HIP/ROCm.
A simple 960 has more performance than those old PhyX Ageia card I think, did they made windows 10 drivers for those old card ?
I mean a 32-bit PhysX driver for the RTX-50 series cards so it doesn't run off the CPU. Can't tell me that the RTX-50 series cards aren't capable of doing something that 5 year old GPU's could do? What I'm referring is using Zluda to add proper GPU PhysX support. It's been talked about for a while.
 
I'm at least glad that someone else thought back to Multi-GPU PhysX and confirmed that it's a valid workaround.

This now raises the question of how little of a secondary GPU you can get away with before it's a bottleneck with PhysX enabled; something like a Quadro P620 or GTX 1050 Ti could just run entirely off the PCIe slot, no additional power needed. Can't go too old, though, or else the PhysX secondary GPU won't even be supported by new drivers. (That rules out Kepler and older. Maxwell is probably pushing it, but safe for now.)

If anything, the bigger problem is just having PCIe slots left over on modern motherboards - if they have any extra slots, they usually get covered up by today's 3-slot, even 4-slot thick HSFs, and what's left is usually a PCIe x4 slot run through the chipset. (I've been griping about this on LGA1700 and AM5 boards for quite some time, and it's forced me to look very carefully at what to use.)
 
Last edited:
I mean a 32-bit PhysX driver for the RTX-50 series cards so it doesn't run off the CPU. Can't tell me that the RTX-50 series cards aren't capable of doing something that 5 year old GPU's could do? What I'm referring is using Zluda to add proper GPU PhysX support. It's been talked about for a while.
It very possible for cpu to not be able to run 8bits-16bits native, (even 32 bits do not run native on some cpu now). Could be purely simpler drivers, CUDA -compiler, but there is simpler ALU-MMU possible to made if you can expect and have a single instruction-memory size.

If Zluda to add pHysX support (or just good cuda support in general) was never made, now that it has been 10 years since the idea to run that kind of physX on the GPU instead of the CPU stopped to seem like a good idea, with multicore cpu asking for something to do and GPU going on at 99-100%, it will work ?

Maybe, but seem simple to just build 64 bits version of any game (if any was worth it), it should not be that difficult.

How were AMD users doing this all along?
Anything remotely popular could have been run with GPU PhysX off.
 
Last edited:
Back
Top