nMP and CUDA, a discussion.

[DesterWallaboo]

I have never seen the mercury playback engine work in CS6 on a Mac running AMD cards. Please explain. It does not work and is the reason Nvidia is preferred at this time and before enterprise takes the hugely expensive CC experiment head on.

Adobe's choice of direction at the time CC came out. Though they've stated that they are planning on moving off of CUDA and going full OpenCL implementation.

----------

Technically Nvidia cards support OpenCL as well... but they do it poorly. Something along the lines of 1/10th the speed of AMD card OpenCL implementation.

 

[prfrma]

Maybe we should track the announcements of once CUDA powered, GPU accelerated programs?

My main ones are Octane, Iray and V Ray RT.

Also, I wonder when CPU bound rendering units like Renderman and Arnold make the switch to GPU rendering.... Probably when 12gb VRAM GPU's are standard... Which is a shame, as a 6 or 8 core dual socket machine would have been perfect for a standalone 3d workhorse.

 

[DesterWallaboo]

Maybe we should track the announcements of once CUDA powered, GPU accelerated programs?

My main ones are Octane, Iray and V Ray RT.

Also, I wonder when CPU bound rendering units like Renderman and Arnold make the switch to GPU rendering.... Probably when 12gb VRAM GPU's are standard... Which is a shame, as a 6 or 8 core dual socket machine would have been perfect for a standalone 3d workhorse.

VRay RT supports both CUDA and OpenCL rendering.

Octane currently renders on CUDA but has openly stated they plan full OpenCL implementation.

IRay.... from everything I read they have no plans for OpenCL. But of course that could change.

I've read nothing regarding any moves for Renderman or Arnold moving to GPU rendering. Who knows?

 

[riggles]

VRay RT supports both CUDA and OpenCL rendering.

Octane currently renders on CUDA but has openly stated they plan full OpenCL implementation.

IRay.... from everything I read they have no plans for OpenCL. But of course that could change.

I've read nothing regarding any moves for Renderman or Arnold moving to GPU rendering. Who knows?

Doesn't VrayRT only support OpenCL on NVIDIA cards at this point? I know Octane has said OpenCL is "planned", but they haven't given any timeline, or shown any working code to my knowledge, and have a strategic partnership with NVIDIA for their new cloud services. Not saying I don't want them to add OpenCL, just saying I wouldn't hold my breath.

Hopefully Pixar will adopt OpenCL (or GPU rendering in general) sooner than it took them to add physical lights and materials into Renderman. But I doubt we'll see it until it's been put thru it's paces in handling a new feature film.

 

[prfrma]

It's strange, even now there are new CUDA only GPU accelerated utilities and renderers out there... If anything, being a low-level, poorly documented open sourced foundation is a massive hinderance on, especially small developers with limited resources starting out, especially when a higher level, fully documented alternative exists.

Wax lyrical as much as you like about the openness of Opensource. The reality is that Nvidia has created a pretty good ecosystem for CUDA that's enabled lots of startups interested in GPU acceleration for all sorts of applications develop market ready, and reliable products. AMD, Intel and Apple have a ton of work to do to get the developing (sdks, documentation, drivers) environment to the same standard for OpenCL or else the vision of what the nMP is supost to offer will only be limited to a few programs maintained by a few big developers with the resources to waste on an ever increasing minority (pro-users on OSX).

 

[flat five]

(pro-users on OSX).

i wouldn't really worry about that too much.. osx rendering was pretty bad 5 years ago and it's only gotten better since.. and continues to.

these new breed developers know if you're not making cross platform applications then you're blowing it..

and yes, mac users are in the minority but not with the same overall mac vs pc stats.. it's more likely for a mac owner to want to use those types of apps so their percentage of osx buyers is a lot higher than the 10-to-1(?) ratio of the general population.
------------------------

regarding the openCL thing.. they all want to use it.. even Thea's developers who just released presto which is gpu only & cuda only have said they felt openCL wasn't ready for prime time when they started the project which is why they went with cuda.. followed by him saying he wrote with openCL in mind so when it matures further, they can switch to it as it's their preferred path..

i don't know.. apple is seemingly saying "hey, let's do this thing" to the entire industry.. i think they're a lot more influential than nvidia.. it's not like cuda is even that established (as in- most people- even around here- don't know what it is).. pretty sure in 5 years time, nobody is even going to remember the great cuda vs opencl battle of 2014.

----------

here's the exact quote regarding presto:

In the beginning there was only one path for us, the standard path, that is OpenCL. But we have found out (the hard way) that OpenCL is not really "tuned" for complex staff. There is a varying performance and quite a lot of fighting with getting the code to actually run. Nevertheless, this is the path we would like to follow and this is why, the language used during programming has been made with OpenCL in mind and to exactly help the transition to the standard when it is more mature. We will be visiting this topic frequently to see how we can make Presto running on ATI/Intel graphic cards.

http://www.thearender.com/forum/viewtopic.php?f=17&t=12500

 

[vladi]

This is really not about OpenCL vs CUDA and there is no such thing as one is better than the other. That "battle" only exists to those who don't use none of those technologies, much like Flash vs HTML5 "battle".

If you have to choose one or the other then you choose it by your workflow, end of story.

If you are working in SolidWorks then OpenCL is your weapon of choice all the way and CUDA has nothing on it as almost any AMD card will do the better job.

But if you take Catia, app from same developer as SW, then CUDA is the tool by huge margin where even Quadro 2000 pisses on W9000.

Adobe is similar story, most of their GPU accelerated apps run wild on OpenCL but when it comes to Ray-tracing in AE its only CUDA.

See how GPU requirements vary even when apps are made by same software developer?

You have to know your tool in other to choose what's right for you.

there is no one PC will do it all. While nMP will give you an edge in Maya due to great OpenCL support, it will really crumble down on you in compositing stage cause Nuke is all CUDA infested.

 

[Anim]

Autodesk Maya and Mudbox both support OpenCL in 2013 releases
http://www.youtube.com/watch?v=TzGvkvDrLqc

OpenCL being Open will get more and more industry support especially as both next gen consoles use AMD hardware (no CUDA).

I think that industry has to embrace OpenCL over CUDA now as more and more people get on board with it. This isn't a bad thing for Nvidia either as the cards can support both technologies.

I am personally hoping that ChaosGroup and their Version 3 release of Vray-RT will fully support OpenCL on FirePro cards.

 

[motegi]

especially as both next gen consoles use AMD hardware (no CUDA).

This is a very good point. The next 8~ years of next-gen gaming has been set. Gives OpenCL a good chance of maturing.

 

[prfrma]

The "because it's open" argument is the same nonsense fandroids use to describe how android development will eventually take over iOS.

Right now, Nvidia cards were a no brainer, can run CUDA, and OpenCL, better workstation drivers and reliability history. There's far more to consider than *Openness* and a lot of this OpenCL optimism is predicated on the nMP being a run away success.

 

[Anim]

The "because it's open" argument is the same nonsense fandroids use to describe how android development will eventually take over iOS.

Right now, Nvidia cards were a no brainer, can run CUDA, and OpenCL, better workstation drivers and reliability history. There's far more to consider than *Openness* and a lot of this OpenCL optimism is predicated on the nMP being a run away success.

I agree. CUDA is well ahead at the moment. But. lets look at who is betting/using OpenCL technology so far; all major operating systems, all next gen consoles systems and software giants like Autodesk and Adobe.

Nvidia is just one company, they can not ignore OpenCL and so will embrace it as much as they have with CUDA only on an open platform which brings it's own issues. But If all the big players out there are adopting it then Nvidia has to as well.

For a more in depth analysis into both technologies, have a read of this:
http://streamcomputing.eu/blog/2011-06-22/opencl-vs-cuda-misconceptions/

Anim

 

[Why??????]

The "because it's open" argument is the same nonsense fandroids use to describe how android development will eventually take over iOS.

Right now, Nvidia cards were a no brainer, can run CUDA, and OpenCL, better workstation drivers and reliability history. There's far more to consider than *Openness* and a lot of this OpenCL optimism is predicated on the nMP being a run away success.

Although Nvidia cards can run OpenCL, their implementations are suboptimal. I mean, even the Iris pro beats a 750M in OpenCl performance, which says a lot.

Why would Nvidia threaten its own creation with a good OpenCL implementation?

 

[riggles]

Wax lyrical as much as you like about the openness of Opensource. The reality is that Nvidia has created a pretty good ecosystem for CUDA that's enabled lots of startups interested in GPU acceleration for all sorts of applications develop market ready, and reliable products. AMD, Intel and Apple have a ton of work to do to get the developing (sdks, documentation, drivers) environment to the same standard for OpenCL...

I really do think this has been the heart of the issue so far. From a business standpoint, when developers choose which technology to support (or build their product on), if their fixed amount of resources can get them a more feature-rich app quicker and more reliably than another technology, why wouldn't you go with it? OpenCL 2.0 seems to be heading in the right direction, but I think AMD has work to do before more developers take another hard look at it.

 

[prfrma]

Although Nvidia cards can run OpenCL, their implementations are suboptimal. I mean, even the Iris pro beats a 750M in OpenCl performance, which says a lot.

Why would Nvidia threaten its own creation with a good OpenCL implementation?

Because at the end of the day, they sell hardware and if OpenCL performance becomes a bigger deal, they'll release better drivers/hardware to adjust.

 

[flat five]

i guess it's worth pointing out that it's not an either/or situation.. applications can support cuda and openCL for gpu acceleration.. such is the case with indigo.

 

[Dr. Stealth]

My main renderer is Bunkspeed. It's CPU and/or GPU (CUDA) based. When in hybrid mode it's smoking fast and produces beautiful renders.

Bunkspeed was recently purchased by RTT and now Dassault Systèmes (Catia & SolidWorks) has announced they will purchase RTT. Interesting times....

 

[chevalier433]

apple has fairly clearly said to developers--
"if you want to write for our platform, use openCL instead of cuda"

do you think there are any developers out there who are wishing to continue selling a mac program that are trying to make any cuda based enhancements for the platform?

for all intents and purposes, cuda is dead on mac.. as in-- if you want cuda based gpgpu, leave mac now..

Thats not right what about their Macbook pro line and iMac they use nVidia cards there a lot of creative business pro's that use them.

 

[flat five]

Thats not right what about their Macbook pro line and iMac they use nVidia cards there a lot of creative business pro's that use them.

you have to try to look at it from a developer's point of view.. ultimately, they're the only ones who matter in any of this.

so say you're a developer using gpu processing/acceleration and you were at wwdc2013.. what would you do? how would you feel about coding cuda on mac?

 

[riggles]

so say you're a developer using gpu processing/acceleration and you were at wwdc2013.. what would you do? how would you feel about coding cuda on mac?

That would really depend on whether I'm a Mac-only developer, no?. If I am, I'd be taking a hard look at OpenCL. If I supported multiple platforms and am already invested in CUDA, I'd wait to see how the dust settles a bit before making any big roadmap decisions. NVIDIA holds a very good position in the PC market, as well as powering all the current MacBook Pros and iMacs, which are quite popular among creative pros.

 

[flat five]

That would really depend on whether I'm a Mac-only developer, no?. If I am, I'd be taking a hard look at OpenCL. If I supported multiple platforms and am already invested in CUDA, I'd wait to see how the dust settles a bit before making any big roadmap decisions. NVIDIA holds a very good position in the PC market, as well as powering all the current MacBook Pros and iMacs, which are quite popular among creative pros.

yeah.. i don't know. i'm not a developer.
i quoted two devs in this thread and could do the same with three others. they're all saying similar things.

go try to convince your developers of this stuff instead of me.. they're the only ones who matter.. i'm not going to argue about it anymore.

 

[riggles]

go try to convince your developers of this stuff instead of me..

I am. They're telling me the same thing: "We're not supporting OpenCL at this time, but perhaps in the future."

I didn't mean to make you upset. You said that CUDA was dead on the Mac. I simply think that it's too soon to come to that conclusion.

 

[flat five]

You said that CUDA was dead on the Mac. I simply think that it's too soon to come to that conclusion.

i personally think cuda is going to go away in general.. if openCL matures and becomes easier to implement and the developer won't have to pay for it? and your potential client pool becomes much larger since a lot more hardware can run the program?
--------

of course the statement "cuda is dead on mac" is not conclusive.. but people have buying decisions to make.. but hey, if you want to believe cuda is alive and well on mac then be my guest.. keep buying macs for your cuda apps. i'd strongly recommend against it but it's not my money.

 

[jasonvp]

I didn't mean to make you upset. You said that CUDA was dead on the Mac. I simply think that it's too soon to come to that conclusion.

It isn't, though. CUDA is dead on the Mac Pros, as of now. If you want your application to sing on a Mac, then you want it coded to work on a Mac Pro. the iMacs and Macbook Pros are very good computing platforms, but they're not Mac Pro level. And don't be too surprised to see Apple push similar direction down to those platforms.

 

[riggles]

..but people have buying decisions to make.. but hey, if you want to believe cuda is alive and well on mac then be my guest.. keep buying macs for your cuda apps. i'd strongly recommend against it but it's not my money.

I don't want to believe CUDA is still alive, it just really is still alive. I don't mind switching to OpenCL, whatever works better for me. But in 3D rendering, OpenCL just hasn't been an option, Mac or PC. So for someone who needs a computer now, like I did two months ago, it's predominantly NVIDIA if you want GPU rendering and that ruled out the nMP right there. I could bet on OpenCL, but how long would I have to wait before my investment in dual AMD graphics starts repaying the money I spent on it? None of my app's developers have even said they're actively working on supporting OpenCL (except Photoshop, where I don't really need it). I feel more comfortable having CUDA support for the next 2-3 years, and then reapproaching it. Again, I'm speaking about the 3D world.

 

[deconstruct60]

I don't want to believe CUDA is still alive, it just really is still alive. I don't mind switching to OpenCL, whatever works better for me. But in 3D rendering, OpenCL just hasn't been an option, Mac or PC. So for someone who needs a computer now, like I did two months ago, it's predominantly NVIDIA if you want GPU rendering and that ruled out the nMP right there.

What is the conflict with "dead on Mac" when you are ruling out Macs because you need CUDA? You own choice is indicative that it has future growth problem in at least part of the Mac space.

Frankly, if Nvidia doesn't get cracking and upgrade to OpenCL 1.2 ( follow relatively quickly with OpenCL 2.0) they will likely die off in the Apple GPU bake-offs unless there is something else spectacularly aligned with what Apple wants on that design iteration as one of the criteria. Intel is delivering OpenCL 1.2 . AMD is delivering OpenCL 1.2. Nvidia isn't. When there is a bake-off they are very easily going to loose on that check list item. [ It wasn't all that long ago that AMD had position in all iMacs and MBPs. It is quite creditable they could win the next round for those dGPUs if they get their power/price/performance mix right. And Intel is quite visibly squeezing dGPUs out of more Mac models. ]

None of my app's developers have even said they're actively working on supporting OpenCL (except Photoshop, where I don't really need it). I feel more comfortable having CUDA support for the next 2-3 years, and then reapproaching it. Again, I'm speaking about the 3D world.

Any app developer betting the farm on CUDA on a Mac over next 2 years is seriously gambling. Perhaps they only have a relatively small Mac application business or have a customer base they feel will cling to 2010-2012 models for several years. Vendors retreat/enter the Mac ecosystem over time. Those still wondering about OpenCL support are probably also wondering about staying in an OS X port of their app.

CUDA isn't going to die overnight or even in a couple of years. But there are a couple of factors. One is that in terms of units Intel and AMD are shipping more GPUs. CUDA's long term trend is to a Mac like sized niche of GPGPU.