Project CARS devs address AMD performance issues, AMD drivers to blame entirely, PhysX runs on CPU only, no GPU involvement whatsoever.

[spartaman64]

Most of the scientific computing algorithms are written in CUDA. Now that Intel is gaining ground it is helping computer scientists rewrite the algorithms in OpenMP, OpenACC, and OpenCL, but the old algorithms can still be run in CUDA on Intel iGPU and Xeon Phi.

really i thought openCL was computationally more powerful than cuda

[Victorious Secret]

oh right forgot about those but those are for science and research which i dont think uses cuda 

 

You are so full of it. I suggest you really start researching before you utter another word. 

 

Our own lab runs a mix of 780s and Titans. Why? We do research. Why? CUDA FUCKING ACCELERATION. CUDA accelerated based simulations. 

 

LOL. Yea, I guess in the Top 500 supercomputers on Earth, the couple Hundred that are Nvidia powered can't possibly be using CUDA acceleration...oh wait. Titan. AKA the namesake of the Titan GPUs. Yea, pretty sure that is a CUDA accelerated supercomputer

[Shakaza]

You are so full of it. I suggest you really start researching before you utter another word. 

 

Our own lab runs a mix of 780s and Titans. Why? We do research. Why? CUDA FUCKING ACCELERATION. CUDA accelerated based simulations. 

 

LOL. Yea, I guess in the Top 500 supercomputers on Earth, the couple Hundred that are Nvidia powered can't possibly be using CUDA acceleration...oh wait. Titan. AKA the namesake of the Titan GPUs. Yea, pretty sure that is a CUDA accelerated supercomputer

Oooh, since we are "widening the discussion," can I ask you what your job is and what simulations you do? :3

[patrickjp93]

really i thought openCL was computationally more powerful than cuda

No no no. If it was AMD's FirePros would be outselling Teslas very easily due to their raw advantage in Teraflops. OpenCL 1.x and 2.0 do not allow you to use C++ standard library algorithms and instead require you to hand-code everything too. It's not very programmer friendly, and at least up to now Linpack and other major parallel benchmarks seem to favor CUDA even with AMD's extensive help in building the OpenCL versions.

[spartaman64]

No no no. If it was AMD's FirePros would be outselling Teslas very easily due to their raw advantage in Teraflops. OpenCL 1.x and 2.0 do not allow you to use C++ standard library algorithms and instead require you to hand-code everything too. It's not very programmer friendly, and at least up to now Linpack and other major parallel benchmarks seem to favor CUDA even with AMD's extensive help in building the OpenCL versions.

ok thanks but amd havent really targeted that market

[Victorious Secret]

ok thanks but amd havent really targeted that market

 

...Because they aren't remotely competitive compared to Knights Landing, Xeon Phi or Tesla. If they were, they wouldn't be in such dire straights as a business. 

 

No one wants to build a supercomputer (which is big money for suppliers) from AMD GPUs. And with how things are today and kit like Knights Landing, no one needs "cpus" in the traditional sense anymore either. AMD fell so far behind the 8 ball with their enterprise affairs. You think consumer sales are what keeps these companies going? No. Its them selling a couple supercomputers worth of GPUs that keeps them going. 

AMD doesn't have that on the level that Intel and Nvidia do. 

[spartaman64]

...Because they aren't remotely competitive compared to Knights Landing, Xeon Phi or Tesla. If they were, they wouldn't be in such dire straights as a business. 

 

No one wants to build a supercomputer (which is big money for suppliers) from AMD GPUs. And with how things are today and kit like Knights Landing, no one needs "cpus" in the traditional sense anymore either. AMD fell so far behind the 8 ball with their enterprise affairs. You think consumer sales are what keeps these companies going? No. Its them selling a couple supercomputers worth of GPUs that keeps them going. 

AMD doesn't have that on the level that Intel and Nvidia do. 

the titan supercomputer that you guys are using as an example has amd opteron cpus

[Victorious Secret]

the titan supercomputer that you guys are using as an example has amd opteron cpus

 

The CPUs are used to dedicated tasks to the Teslas, which pack 2880 CUDA cores at...~740mHz each. GPUs are far more powerful in that system for getting simulations done than the CPUs. Which is the point of the Titan. Its a GPU based supercomputer that ranks #1 in the world because of those Teslas, not because of the CPUs. 

 

We're getting to a point where having a dedicated CPU will be redundant in a supercomputer. 

[Notional]

You don't understand how PhysX works. At all. So why do you mention something irrelevant?

Wrong. PhysX. Holy fuck how many times do I have to say it?

 

You seriously need to calm down, and stop spamming the forum.

 

https://developer.nvidia.com/content/introducing-nvidia-gameworks

 

We get that the physics engine of PhysX, can run on a CPU. We get it. We also get that the advanced PhysX effects (under Physics in the link, runs on NVidia GPU's. Now answer me this: Why is it, that when NVidia users, set PhysX to run on the CPU, it tanks performance? Surely, if these advanced effects, only runs on NVidia GPU's, this should never happen?

 

Now read this post one more time:

once again parallel processors 

 

PhysX (ALL OF IT) uses floating point processing, NOT integer. You know what a dedicated floating point processor is often called? A GPU! Which is a parallel processor.

Still don't understand? Here is the original design of Aegis' dedicated PhysX card (before advanced PhysX was a thing): http://www.blachford.info/computer/articles/PhysX2.html

 

Notice how it is A FLOATING POINT PROCESSOR? Spartanman64 is 100% right, and YOU don't seem to grasp how PhysX actually works. So stop spamming the same retarded statement over and over. A CPU's floating point FPU's are generally weak, which is why AMD invented HSA, to offload FP processing to the integrated GPU. If this game spams the hell out of the CPU's FPU's, it will run like shit, even it is only physics PhysX, and not advanced.

[AlwaysFSX]

You seriously need to calm down, and stop spamming the forum.

 

https://developer.nvidia.com/content/introducing-nvidia-gameworks

 

We get that the physics engine of PhysX, can run on a CPU. We get it. We also get that the advanced PhysX effects (under Physics in the link, runs on NVidia GPU's. Now answer me this: Why is it, that when NVidia users, set PhysX to run on the CPU, it tanks performance? Surely, if these advanced effects, only runs on NVidia GPU's, this should never happen?

 

Now read this post one more time:

 

PhysX (ALL OF IT) uses floating point processing, NOT integer. You know what a dedicated floating point processor is often called? A GPU! Which is a parallel processor.

Still don't understand? Here is the original design of Aegis' dedicated PhysX card (before advanced PhysX was a thing): http://www.blachford.info/computer/articles/PhysX2.html

 

Notice how it is A FLOATING POINT PROCESSOR? Spartanman64 is 100% right, and YOU don't seem to grasp how PhysX actually works. So stop spamming the same retarded statement over and over. A CPU's floating point FPU's are generally weak, which is why AMD invented HSA, to offload FP processing to the integrated GPU. If this game spams the hell out of the CPU's FPU's, it will run like shit, even it is only physics PhysX, and not advanced.

You're missing the point of what the task is and how simple it is to compute that even a CPU can do it. If games use a PhysX  physics engine and don't run like shit, what do you do to explain that?

 

Again.

 

Wrong. PhysX.

[Notional]

You're missing the point of what the task is and how simple it is to compute that even a CPU can do it. If games use a PhysX  physics engine and don't run like shit, what do you do to explain that?

 

Again.

 

Wrong. PhysX.

The why does performance tank, when set to CPU processing on NVidia systems? Why are we seeing tanked performance on even 5960x processors? Do you believe, that it is impossible to make the physics engine bottle neck a CPU?

 

PhysX, both "types" are floating point, and runs on parallel processors, nothing to discuss. In this game, it seams that it has problems running on the CPU. And that is somehow AMD's fault?

Do we even know if this game forces advanced PhysX on the CPU or not yet?

[AlwaysFSX]

The why does performance tank, when set to CPU processing on NVidia systems? Why are we seeing tanked performance on even 5960x processors? Do you believe, that it is impossible to make the physics engine bottle neck a CPU?

 

PhysX, both "types" are floating point, and runs on parallel processors, nothing to discuss. In this game, it seams that it has problems running on the CPU. And that is somehow AMD's fault?

Do we even know if this game forces advanced PhysX on the CPU or not yet?

@Victorious Secret You feel like explaining the difference between the physics engine and PhysX particle engine again or should we leave this one to stew with not knowing games using the PhysX physics engine run perfectly fine?

[Victorious Secret]

@Victorious Secret You feel like explaining the difference between the physics engine and PhysX particle engine again or should we leave this one to stew with not knowing games using the PhysX physics engine run perfectly fine?

Not particularly. If he can't understand that PhysX acceleration and PhysX physics are two different implementations, what's the point?

It's funny that he tells you to stop spamming the forum. Pot. Kettle. Never mind.

Maybe Patrick can explain to him the fundamental differences. Or maybe he can tell us about his business major and how it makes his viewpoints more valid. Who knows, this thread has been a fantastic train wreck so far.

Hard to talk with facts when the other side doesn't care if it doesn't fit the narrative.

I feel like Jon Stewart showing up to a round table involving Hannity and O'Rielly and Olberman. Disastrous all around.

[AlwaysFSX]

Not particularly. If he can't understand that PhysX acceleration and PhysX physics are two different implementations, what's the point?

It's funny that he tells you to stop spamming the forum. Pot. Kettle. Never mind.

If it's Apex and you know it clap your hands...

[Notional]

@Victorious Secret You feel like explaining the difference between the physics engine and PhysX particle engine again or should we leave this one to stew with not knowing games using the PhysX physics engine run perfectly fine?

 

Not particularly. If he can't understand that PhysX acceleration and PhysX physics are two different implementations, what's the point?

It's funny that he tells you to stop spamming the forum. Pot. Kettle. Never mind.

 

Please stop the arrogant circle jerking. I know damn well the difference, and even linked to the definitions of both PhysX and Gameworks in general.

 

How about you just answer the question? Smoke is part of advanced PhysX, yet seems to be rendered on AMD systems. If so, some advanced PhysX effects might be force run on a CPU, which would explain the massive bottle necking a lot of people are experiencing (even NVidia users, who had set PhysX to run on the CPU).

[AlwaysFSX]

Please stop the arrogant circle jerking. I know damn well the difference, and even linked to the definitions of both PhysX and Gameworks in general.

 

How about you just answer the question? Smoke is part of advanced PhysX, yet seems to be rendered on AMD systems. If so, some advanced PhysX effects might be force run on a CPU, which would explain the massive bottle necking a lot of people are experiencing (even NVidia users, who had set PhysX to run on the CPU).

Apex PhysX system =/= PhysX physics engine.

 

Do you understand this much?

[Briggsy]

Please stop the arrogant circle jerking. I know damn well the difference, and even linked to the definitions of both PhysX and Gameworks in general.

 

How about you just answer the question? Smoke is part of advanced PhysX, yet seems to be rendered on AMD systems. If so, some advanced PhysX effects might be force run on a CPU, which would explain the massive bottle necking a lot of people are experiencing (even NVidia users, who had set PhysX to run on the CPU).

 

another glaring issue is that PhysX only runs on a single cpu thread. Nvidia have openly admitted that they haven't tried to make it multi-threaded, and they aren't going to try. Its very passive aggressive, but in no way can anyone pin any fault on Nvidia for doing this, even if it degrades performance for the competition, or even for themselves in some cases.

[Notional]

Apex PhysX system =/= PhysX physics engine.

 

Do you understand this much?

 

Generally that has been the case. Can you conclude, without a doubt, that this is also the case in this particular game?

 

another glaring issue is that PhysX only runs on a single cpu thread. Nvidia have openly admitted that they haven't tried to make it multi-threaded, and they aren't going to try. Its very passive aggressive, but in no way can anyone pin any fault on Nvidia for doing this, even if it degrades performance for the competition, or even for themselves in some cases.

 

Wait really? Do you have a source for this? Sounds absolutely insane!

[AlwaysFSX]

Generally that has been the case. Can you conclude, without a doubt, that this is also the case in this particular game?

Without a doubt. Because otherwise the game would be churning along at a wonderful 15 fps.

[Briggsy]

Wait really? Do you have a source for this? Sounds absolutely insane!

 

Its true for the most part. Nvidia physx 3.0 brought in support to use more than one cpu core, but games that use an older version of physX are stuck using only a single cpu thread, and it is those games Nvidia aren't concerned with, because its up to the developer to enable Physx to use more than a single thread. I'm not certain whether physX 3.0 and newer have multi-threading enabled by default, but it feels like it wasn't that long ago that PCPer were discussing this very issue on a podcast, and Nvidia weren't all that concerned about making it multi-threaded (obviously).

[AlwaysFSX]

Its true for the most part. Nvidia physx 3.0 brought in support to use more than one cpu core, but games that use an older version of physX are stuck using only a single cpu thread, and it is those games Nvidia aren't concerned with, because its up to the developer to enable Physx to use more than a single thread. I'm not certain whether physX 3.0 and newer have multi-threading enabled by default, but it feels like it wasn't that long ago that PCPer were discussing this very issue on a podcast, and Nvidia weren't all that concerned about making it multi-threaded (obviously).

3.0 came out in '11 iirc, and before then games really weren't made to run well on multiple cores either so it was a problem everyone had.

[zMeul]

nVidia rep, on reddit, said: https://www.reddit.com/r/pcgaming/comments/366iqs/nvidia_gameworks_project_cars_and_why_we_should/crc3ro1

The assumptions I'm seeing here are so inaccurate, I feel they merit a direct response from us.

I can definitively state that PhysX within Project Cars does not offload any computation to the GPU on any platform, including NVIDIA. I'm not sure how the OP came to the conclusion that it does, but this has never been claimed by the developer or us; nor is there any technical proof offered in this thread that shows this is the case.

I'm hearing a lot of calls for NVIDIA to free up our source for PhysX. It just so happens that we provide PhysX in source code form freely on GitHub (https://developer.nvidia.com/physx-source-github), so everyone is welcome to go inspect the code for themselves, and optimize or modify for their games any way they see fit.

Rev Lebaredian

Senior Director, GameWorks

NVIDIA

 

so, I wanted to test it for myself,

 

 

 

yes, I know .. the system is not top-notch  ; highest settings, 1080p, Vsync OFF, MSAA

• CPU: Intel Q9550 @2.8Ghz
• RAM: 8GB DDR2 800Mhz
• Video Card: Gigabyte GTX970 G1 @factory defaults
• drivers: GeForce 347.88 / PhysX 09.14.0702
• OS: Windows 7 x64

 

---

 

since quite a few people hinted that the PhysX toggle in the nVidia Control Panel doesn't work, I decided to run the test on Windows 10 x64 Preview

with the added bonus of the race being in light rain:

• drivers: GeForce 352.84 / PhysX 09.15.0324

 

---

edit: added avg/min/max FPS chart

 

edit 2: added W10 x64 Preview benchmark numbers

[mr moose]

The CPU side is open source, AMD can easily recreate the exact same function calls and get those to run on a GPU instead, making it very easy for the game developers.

 

PhysX runs on the CPU if there's no NVIDIA GPU available, AMD can create exactly the same API (names, functions etc) and make it run on the GPU.

 

AMD is having issues because they have done some wrong steps, not because nvidia is a huge company. AMD is still #1 in the console industry.

 

Nvidia's size does play a role in it,  with 75% of the pc market share that's who developers are going to favour when deciding who's middle ware to use. 

 

EDIT: or you could look at it from the other side, if you were undertaking a big project like this would you choose to use middleware from a company that can't afford to support it properly or at worst go bankrupt tomorrow? 

[AlwaysFSX]

nVidia rep, on reddit, said: https://www.reddit.com/r/pcgaming/comments/366iqs/nvidia_gameworks_project_cars_and_why_we_should/crc3ro1

 

so, I wanted to test it for myself,

 

yes, I know .. the system is not top-notch  ; highest settings, 1080p, Vsync OFF, MSAA

• CPU: Intel Q9550 @2.8Ghz
• RAM: 8GB DDR2 800Mhz
• Video Card: Gigabyte GTX970 G1 @factory defaults
• drivers: GeForce 347.88 / PhysX 09.14.0702
• OS: Windows 7 x64

I'm amused that your FPS is higher running PhysX on your CPU, and that me and Vic were correct in what we said. Thanks for the data.

[Dabombinable]

I'm amused that your FPS is higher running PhysX on your CPU, and that me and Vic were correct in what we said. Thanks for the data.

I'll run it with Nvidia's PLA benchmark, 1 test on my GTX 970, 1 test on my GTX 650 Ti, 1 Test on the CPU, and a test with the 650 Ti as the PhysX card. And I can say right now that although I've lost the original SS, running PhysX on the CPU makes the effects worse and reduces the FPS by a lot-and that was with my i5, not my Xeon X5450.