AMD Publishes FEMFX Deformable Physics Library on GPUOpen

[Deleted member 171912]

Nice to see FEM is also used this way.

 

[evernessince]

No.Can't be more wrong.
take a game that uses some sort of cpu physics,bf5 as a good example,and see what happens to cpu loads during explosion.

what we have is gpu architectures that pack more and more compute power into smaller and smaller power envelopes.


Exactly,like recommending buying 8c/16t workstation cpus for gaming cause of physics.

you got $700 to spend ? get a $200 cpu and a $500 gpu instead of packing a $350 cpu in there.

I don't really see what you are fighting against here. Are you against doing physics on the CPU thus leaving extra GPU cores to actually rendering more frames? If CPU core counts keep doubling every other generation like AMD has been doing, there's not reason to not run physics on the CPU.

 

[Zubasa]

PhysX isn't crippled, and Havok is still widely used (especially in multiplatform titles). The only downside, is that MS and Havoc haven't done anything to improve it since 2011, and DX physics is taking too long to come to fruition.

Havok was first purchased by Intel and then by Microsoft.
When you get WinTel, you get ultimate stagnation.
They are basically Landlords intent on collecting rent instead of actual tech companies.

I don't really see what you are fighting against here. Are you against doing physics on the CPU thus leaving extra GPU cores to actually rendering more frames? If CPU core counts keep doubling every other generation like AMD has been doing, there's not reason to not run physics on the CPU.

Yeah I don't understand the obsession with running everything on the GPU when modern CPUs have most of its cores idling in games.
Sure the GPU can do Physics fasters than CPUs when running in a vaccum, but in games you are competing for resources in the GPU.

 

[Flanker]

hmm, let me see how much SIMD it uses. I could try implement this with DirectCompute as a hoppy project lol

 

[Fouquin]

still,physics should be done on gpu.

Even nVidia shows they don't agree with that statement. They ported PhysX to CPUs (PowerPC AltiVec and x87) almost immediately after acquiring the technology, and rewrote huge portions of the codebase in 2010-2011 for a more unified approach with SSE and multi-thread support. The idea that physics simulations are impractical/worthless on CPUs purely stems from the limitations of general purpose CPUs circa 2003-2005 that had limited register space, lacking FP instruction options/hardware, and minimal parallel thread execution capabilities. Modern CPU architectures have solved all of those issues.

 

[jahramika]

yes but you've got GPUs having absolutely ridiculous compute power too.why spend extra on an 8c/16t when a 6c/6t is plenty and your gpu packs so much power. how much does a 5700xt/2070 super pack ? 8-9 tflops ? probaly 10 overclocked. And both can do fp+int or fp16. Your rdna2 console gpu will probably be close to that too.

so can off load the work their gpu to perform faster!

 

[R-T-B]

Because you say so, Nvidia backer says gpu physx only please.

AMD backer would say it too if they had it. It's a good thing. We just need an open version, badly.

 

[HwGeek]

NIce, with all those extra cores we are starting to see some benefits like RTRTraced shadowes on CPU in WOT and in near future Physics calc on CPU too.

 

[Frick]

Havok was first purchased by Intel and then by Microsoft.
When you get WinTel, you get ultimate stagnation.
They are basically Landlords intent on collecting rent instead of actual tech companies.

Off topic, but definitely not true for Microsoft as a whole. They're very much still one of the giants in tech, despite missing out on mobile.

 

[uuuaaaaaa]

Awesome to see real time FEM applied to video games. This will be ground breaking.

 

[cucker tarlson]

Even nVidia shows they don't agree with that statement. They ported PhysX to CPUs (PowerPC AltiVec and x87) almost immediately after acquiring the technology, and rewrote huge portions of the codebase in 2010-2011 for a more unified approach with SSE and multi-thread support. The idea that physics simulations are impractical/worthless on CPUs purely stems from the limitations of general purpose CPUs circa 2003-2005 that had limited register space, lacking FP instruction options/hardware, and minimal parallel thread execution capabilities. Modern CPU architectures have solved all of those issues.

yes,but are video cards from then just like gpus now ?

nvidia ported physx to cpu a decade ago,therefore they don't agree now it should be done on gpu.really ? does time work that way ?

and really,nvidia isn't the oracle to say that physx is the only way for improving game physics,amd and intel should step in too cause with physx being the only option there isn't much progress.

 

[csendesmark]

Impressive!
I am already waiting to see this in games!

 

[cucker tarlson]

Impressive!
I am already waiting to see this in games!

have they even announced any games using this ?

 

[jabbadap]

yes,but are video cards from then just like gpus now ?

nvidia ported physx to cpu a decade ago,therefore they don't agree now it should be done on gpu.really ? does time work that way ?

and really,nvidia isn't the oracle to say that physx is the only way for improving game physics,amd and intel should step in too cause with physx being the only option there isn't much progress.

PhysX is open source now-a-days(BSD license). Most of GPU accelerated stuff from it is strip away long ago and are part of VisualFX.

 

[Octopuss]

This thread reminded me of something that looked interesting in past, but likely nothing serious came out of it: https://www.youtube.com/user/AndersOerum/videos

 

[Fouquin]

nvidia ported physx to cpu a decade ago,therefore they don't agree now it should be done on gpu.really ? does time work that way ?

nVidia ported PhysX to CPUs at a time when CPUs did not integrate all of the required instructions to properly accelerate physics simulation to the degree of GPUs, that alone should speak for itself. What you originally suggested is that physics simulation should only be done by GPUs. Clearly that's not what nVidia thinks as they've rebuilt PhysX multiple times for CPUs.

and really,nvidia isn't the oracle to say that physx is the only way for improving game physics,amd and intel should step in too cause with physx being the only option there isn't much progress.

This is where AVX comes into play. Nearly a decade of support with every x86 licensee distributing products with it included, and with ever increasing capabilities. Computational physics takes advantage of it already at the academic level, and efficient vectorization of FEM is fairly well understood by now. Note how most of these papers at least mention that global distribution is inherently possible with their respective solutions. Hardware agnostic, so long as AVX is supported it can be deployed.

Intel Developer Zone

Find software and development products, explore tools and technologies, connect with other developers and more. Sign up to manage your products.

software.intel.com

N-body simulation for self-gravitating collisional systems with a new SIMD instruction set extension to the x86 architecture, Advanced Vector eXtensions

We present a high-performance N-body code for self-gravitating collisional systems accelerated with the aid of a new SIMD instruction set extension of the x86 architecture: Advanced Vector eXtensions (AVX), an enhanced version of the Streaming SIMD Extensions (SSE). With one processor core of...

arxiv.org

https://indico.cern.ch/event/570876/contributions/2347270/attachments/1358499/2059757/7-VecPhys.pdf

Vectorization of a spectral finite-element numerical kernel