[jefierro]

Newcomer here.

I just upgraded to maxwell4, thanks Black Friday, knowing that it only uses CUDA for parallel GPU, we mostly use vray for production work, we have just upgraded hardware at the office, and I basically made the decision to incorporate amd's firepros into the new machines just to take advantage of vRayRT's OpenCL utilizing both CPU and GPU in conjunction for parallel processing for faster previews, final renders are always done by CPU, mostly huge memory hungry scenes.

As from last year that I win a promotional maxwell3 license I have been loving it increasingly, I'll hope to use amd's GPU processing in the future into maxwell, to keep pushing this workflow, if this starts a trend using CPU&GPU at the same process, will only work in a non proprietary implementation, I think, or if only nvidia star not caring in only selling you GPUs.

[luis.hijarrubia]

Well, Apple seems not to love OpenCL either. No mac supports OpenCL over 1.2, wich is a 6 years old standar. And let me assure we cannot implement a decent engine on OpenCL 1.2 (or at least one with good performance).

[dmeyer]

Just because majority believes so & have chosen CUDA over OCL, that doesn't makes it a standard (not yet)...

Actually, that is exactly what it means, for all practical purposes.

I've got nothing personal against OCL but explain to me why it would be a first priority for limited development resources when:

1. nVidia has around 71% of discrete GPU sales volume.. I would suspect it's even higher when looking at the workstation subsegment but I didn't feel like paying $2500 for the latest workstation market share report.

2. Apple, OpenCL's biggest proponent in the past, hasn't shipped a Mac that supports higher than OpenCL 1.2, which is 6 years old. Apple is also now encouraging developers to use their Metal API instead.

I understand the whole religious argument that "OPEN IS BETTER!" is appealing but personally I prefer to use the right tool for the job.

Perhaps some day NL will add OpenCL support. And then Metal support so OSX can use it.

[Mihai]

2. Apple, OpenCL's biggest proponent in the past, hasn't shipped a Mac that supports higher than OpenCL 1.2, which is 6 years old.

That same chart shows Apple also only supports OpenGL 4.1, which is also 6 years old

How does this support differ from Windows? Where can we find out which is the latest OpenGL version to have Windows support? Or does it work differently for Win? Does it automatically support the latest version by updating your GFX card OpenGL drivers?

[dmeyer]

How does this support differ from Windows? Where can we find out which is the latest OpenGL version to have Windows support? Or does it work differently for Win? Does it automatically support the latest version by updating your GFX card OpenGL drivers?

Yes the OpenGL support is part of the driver package for both Windows and OSX. However, Apple is the only supplier of drivers for AMD cards in OSX. Occasionally nVidia will do a 3rd party driver package for OSX. For instance, you can run up to a Maxwell-arch Titan in a 2012 Mac Pro tower on PCIe 2.0.

Pascal on OSX is not supported by either Apple natively nor nVidia.

[dk2079]

How does this support differ from Windows? Where can we find out which is the latest OpenGL version to have Windows support? Or does it work differently for Win? Does it automatically support the latest version by updating your GFX card OpenGL drivers?

usually the application you run is more likely the limiting factor
Rhino3d V5.0 for example uses openGL 2.0 and shader 1.2
the next release 6.0 will use openGl 3.3
maya viewport 2.0 uses openGL 3.2

for workstation application stability is primary focus and from what I read it seems pretty difficult to update a complex display pipeline to higher openGL versions and maintain compatibility/stability with all sorts of devices/gpus.

games have it easier in that respect, but these days most is D3D, I am wondering myself what software is making use of the latest openGL 4.x specs.

[Mihai]

Thanks for the clarifications

So I guess it's safe to say the latest OpenGL available today won't be used for at least another 8 years in most applications, including renderers....Bleeding edge my ass

[dk2079]

including renderers....

cuda and open>CL< are a different story though, the latest cuda release reflects the capabilities of the latest GPU chip generation. so for a renderer to be really fast and optimized they have to move to the latest cuda release rather sooner than later.

[Mihai]

But a renderer is more reliant on CUDA version vs openGL version? Anyway....I think you really need to be passionate and slightly masochistic to develop something for GPUs It seems a lot more tricky than for CPU in regards to everything, including debugging, testing 20 different configs and drivers...

[jefierro]

I'll share this benchmark we did to assets performance on a 2012 macpro (trashcan).

CPU = 12 core xeon E5.
GPU = amd firepro D700.

Complex multi-light scene using 3dsmax in win10.

CPU mode only 552 KPaths/s

OpenCL_CPU 1090.4 KPaths/s
OpenCL_1xGPU 721.6 KPaths/s
OpenCL_2xGPU 1461.9 KPaths/s
OpenCL_1xGPU+CPU 1744.8 KPaths/s
OpenCL_2xGPU+CPU 2482.8 KPaths/s

Is there a logical path to employ both cpu & gpu to process under maxwell? I see advantages in doing so.