Support NeoGAF

[Draugoth]

Afew months ago we learned that the new cards in the RTX 50 range didn't have support for 32-bit PhysX technology, which means that older games have trouble running on the new GPUs at their highest quality.

The result was that we saw an RTX 5080 catching up to a GTX 980Ti in old games with PhysX enabled. Although it doesn't sound like something NVIDIA itself will fix, at least it has given users the opportunity to do it themselves.

NVIDIA recently released the PhysX source code to the community, which means that the technology is now open source and anyone can tinker and modify it.

Since the release of PhysX SDK 4.0 in December 2018, NVIDIA PhysX has been available as open source under the BSD-3 license, with one important exception: the GPU simulation kernel source code was not included.

Today that changes. We're happy to share that the latest PhysX SDK update now includes the entire GPU source code, fully licensed under BSD-3!

With over 500 CUDA kernels powering features such as body dynamics, fluid simulation and deformable objects, the PhysX GPU represents one of the most advanced real-time simulation use cases for CUDA and GPU programming. We expect this release to be a valuable resource for learning, experimentation and development across the community.

Source

 

[Closer]

Here's hoping for someone to port and test it on AMD GPUs just for fun.

 

[TVexperto]

So they gave up and want us to fix it for them?

 

[RoboFu]

So they gave up and want us to fix it for them?

Ah they probably shifting to an ai physics engine which should a lot faster.

 

[justiceiro]

So they gave up and want us to fix it for them?

Yes?

They don't include this on their GPUs anymore, why bother mantaining the code closed.

Hell, I wish all companies simply released the code of the features they gave upon insteas of hiding it.

 

[metareferential]

Hopefully the community will fix this. I want to replay Mirror's Edge and I can't fit a second gpu in my new case. xD

 

[Beechos]

Cpus nowadays can't just brute force the physics effects from like 10+ years ago?

 

[Reizo Ryuu]

Cpus nowadays can't just brute force the physics affects from like 10+ years ago?

no

 

[Allandor]

Cpus nowadays can't just brute force the physics affects from like 10+ years ago?

More or less, yes. At least the Xbox/ps4 could use multi core physx code. Nvidia deactivated that for the PC market. Physx games were always forced to just use one core max. With SSE x and AVX instructions there would be much more possible. Heck even just using all cores would be way better. But Nvidia always limited the good stuff to GPU compute even though a gtx1030 is more than enough to get the work and much more done. It's not as demanding as Nvidia always wanted everyone else to think. It was a software written limit.
Yes there are limits what should be done, but just because it shouldn't cost to much GPU time needed for other stuff.

But just as they always do, it was just a vendor lock.


But well it is good they finally released it. Well it is more or less dead at this point but still good.

 

[metareferential]

Some games really are unplayable in their original form without it (or need a second gpu).

Mafia 2, Mirror's Edge and others.

 

[SScorpio]

Some games really are unplayable in their original form without it (or need a second gpu).

Mafia 2, Mirror's Edge and others.

"Unplayable" - AMD users played them all just fine. Hopefully someone is able to get this working on all cards.

 

[Fafalada]

Ah they probably shifting to an ai physics engine which should a lot faster.

 

[Fafalada]

"Unplayable" - AMD users played them all just fine. Hopefully someone is able to get this working on all cards.

In fairness - this means someone could get this working on non NVidia cards for the first time in 2 decades.
The fact NVidia wants positive PR for this is - well really low - but from preservation perspective - this ultimately 'is' a win.

 

[Fafalada]

Cpus nowadays can't just brute force the physics affects from like 10+ years ago?

APEX particles were restricted to a single-threaded implementation on CPU specifically to make sure noone but NVidia GPU enabled system ever had access to it.
But with this source 'unlock' someone could just fix that restriction (it may not even require a GPU port to other vendors). Modern CPUs are more than powerful enough to run this.

 

[Soodanim]

Credit where it's due - this is a good thing.

 

[RoboFu]

Huh yeah it could be a lot faster using prediction methods.

 

[Buggy Loop]

Huh yeah it could be a lot faster using prediction methods.

You're right

AI predicted fluid simulations / cloth and so on are leagues faster than any brute force method.









Bizarre of Fafalada of all peoples to go aggro on this

 

[Fafalada]

Huh yeah it could be a lot faster using prediction methods.

I took issue with the term 'physics AI engine' - not the fact that there's tons of cosmetics enhancements that we could do a lot faster without running real simulation on them (ML proof of concept for this is a decade old, it doesn't even need 'tensor cores' to run well). The thing is though - that's what this entire Apex etc. story has always been (cosmetics) - NVidia never took the ball away for the actual simulation (with good reasons).

The other bit is that for once - it'd be nice to talk about AI for something other than cosmetics - but I digress.

Bizarre of Fafalada of all peoples to go aggro on this

The papers you posted are precisely what I'm referencing above. It's all nice stuff - but it's not 'physics engine' even in the loosest terms. It could give us more dynamic feeling worlds to leverage that though - yes.

 

[SScorpio]

In fairness - this means someone could get this working on non NVidia cards for the first time in 2 decades.
The fact NVidia wants positive PR for this is - well really low - but from preservation perspective - this ultimately 'is' a win.

I'm really hopeful this is made to work on all systems.

But I wouldn't call it NVIDIA fishing for positive PR unless you found a press release or something. The linked source is a discussion thread on the GitHub repo that's been around for years. It was mostly nerds finally convincing their bosses to let them open source it after all internal support was officially dropped. It no longer something that can make them money, better to get it out there than die as hardware ages out.

 

[Puscifer]

Ah they probably shifting to an ai physics engine which should a lot faster.

Pretty much, I feel like 99% of what we saw physx doing at one point is already done in engine now

 

[RoboFu]

I took issue with the term 'physics AI engine' - not the fact that there's tons of cosmetics enhancements that we could do a lot faster without running real simulation on them (ML proof of concept for this is a decade old, it doesn't even need 'tensor cores' to run well). The thing is though - that's what this entire Apex etc. story has always been (cosmetics) - NVidia never took the ball away for the actual simulation (with good reasons).

The other bit is that for once - it'd be nice to talk about AI for something other than cosmetics - but I digress.

That's cool thing about ai . We can easy give you an answer for 2x2 and 4x4. But ai can remember 23463 x 45664 just as easily. It can remember what a ball that weighs .12 lbs hitting a concrete material at velocity of 5 mph with a north east wind of 10 knots will result in. It doesn't have to do all of math every time.

 

[Gaiff]

Cpus nowadays can't just brute force the physics affects from like 10+ years ago?

Considering PhysX was deliberately programmed to murder them and be in the order of 100x less efficient when running on a CPU, no.

Ask again in 15 years.

APEX particles were restricted to a single-threaded implementation on CPU specifically to make sure noone but NVidia GPU enabled system ever had access to it.
But with this source 'unlock' someone could just fix that restriction (it may not even require a GPU port to other vendors). Modern CPUs are more than powerful enough to run this.

This. Fafalda explained it better than I.