nMP and CUDA, a discussion.

[prfrma]

So, it's launched, it's handling FCP pretty well an appears to have a few decent benchmarks. It's not the reincarnation of the dual socket monster we all wanted but it would have been an easier pill to swallow if the workflow disruption caused by no CUDA compatible GPU options wasn't the case.

Who will jump first? Will Adobe and others who currently offer CUDA accelerated rendering jump first and switch to OpenCL? Or are Apple working on providing Nvidia options with the next generation of Quadro cards?

I guess using thunderbolt with a GPU enclosure is a possibility, but what a faff...

Which OpenCL render engines use OpenCL with multiple GPU support?

If I wanted to use Arnold, why not just get a dual socket PC in stead?

Maddening.

 

[Pressure]

I suppose you haven't been following Adobe that closely, let alone their blog.

They are currently working on replacing CUDA entirely with OpenCL in the Adobe CC suite.

Currently, as far as I know, only Compressor 4 is fully compatible with OpenCL.

Update: Just to clarify, both Photoshop and Premiere Pro already supports OpenCL.

 

[ZnU]

DaVinci Resolve 10 also enhances OpenCL support (which started showing up in 9, if I recall correctly) to the point where it's on equal footing with CUDA.

 

[flat five]

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..

 

[jasonvp]

Update: Just to clarify, both Photoshop and Premiere Pro already supports OpenCL.

To clarify: Premiere 100% supports OpenCL for everything and anything it can do with CUDA. And the CC version of Premiere adds support for multiple GPUs, something not seen before with it. CC also adds support for OpenCL on Windows; prior to that (CS6), OpenCL support was only available in OS X.

 

[PuppiesCakePizz]

apple has fairly clearly said to developers--
for all intents and purposes, cuda is dead on mac.. as in-- if you want cuda based gpgpu, leave mac now..

Worth pointing out iMac has GTX 775...not to mention bootcamp.

Speaking from personal experience, lots of people need Cuda and use macs just fine. OpenCL is forward thinking, but cuda is right now.

That said, nMP is pretty handicapped with lack of nvidia. Interesting also seeing the imac with a 775M go toe to toe / outperform the dual firepro (7970?) nMP in a variety of tests. The excuse people are giving? Lack of driver support! Oh the irony..

http://www.nvidia.com/object/mac-driver-archive.html

^ dead..?

 

[derbothaus]

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.

 

[iSee]

OpenCL vs. CUDA

Short-term (say, a 2-3,4 year timeframe) there will be some pain and tradeoffs.
But long term for GPGPU, I think OpenCL will flourish and CUDA will die.

OpenCL works for CPU cores, non-nVidia GPUs, and is a better position to support potential future, difficult-to-predict paths to more computational power... (let's see, I'll make up a term here: XPUs)

I just can't see why any rational software developer would invest in CUDA over OpenCL unless nVidia was sending over wheelbarrows of cash to tip the scales.

The only way I see CUDA surviving is by essentially becoming OpenCL (e.g., nVidia opens licensing, starts supporting CPU, etc.) But my guess is they'll simply press their short-term advantage while it lasts... Get while the Gettn's good.

However, if I'm investing in workstations and related information systems today I think you'd be a sucker to bet on future software releases: If workable OpenCL solutions exist in your domain now (not necessarily comparable to the CUDA options but close enough for your uses), you can invest in OpenCL with confidence. If not, I'd have to go for CUDA with the understanding that I may need to retool yet again. If I'm a Mac shop that would prefer not to change but OpenCL isn't currently an option for the tools I use, I'd do whatever I could to delay, delay, delay in the hopes of getting more clarity.

In general, the new Mac Pro looks like Apple's bet on a future direction for workstations. Apple has a pretty good track record in predicting (if not setting) the future direction of various tech segments. But as the potential customer of the new Mac Pro, you're probably more interested in what you need to get done today or in the next few months. Today, buy tools for today. Buy tools for tomorrow... tomorrow.

 

[jasonvp]

I have never seen the mercury playback engine work in CS6 on a Mac running AMD cards.

It works just fine on OS X. There's an OpenCL supported text file in the same place the CUDA supported text file exists. Delete said file, or add the exact GPU you're using into it (I always found deleting the file to be easier) and voila: Acceleration.

 

[derbothaus]

It works just fine on OS X. There's an OpenCL supported text file in the same place the CUDA supported text file exists. Delete said file, or add the exact GPU you're using into it (I always found deleting the file to be easier) and voila: Acceleration.

Well... crap. 10.7 or later may be why I am thinking this. Thanks for the info.
What about AE?

 

[flat five]

Worth pointing out iMac has GTX 775...not to mention bootcamp.

equally worth pointing out that openCL runs on nvidia gpus

OpenCL is forward thinking, but cuda is right now.

cuda was was up until ~6months ago.. from that point on, it's highly unlikely developers keep coding cuda on os x other than finishing up something they were already far along in -or- just porting over improvements from windows versions..

That said, nMP is pretty handicapped with lack of nvidia.

no it's not.. just take a wild stab at how many mac pro users are going to be affected by no nvidia.. surely it's a fraction of one percent, right?

Interesting also seeing the imac with a 775M go toe to toe / outperform the dual firepro (7970?) nMP in a variety of tests. The excuse people are giving? Lack of driver support! Oh the irony..

the imac is going to be outperforming the nmp in the cpu realm as well in certain tests.. understand the tests- understand your requirements.. and pick the right computer.

http://www.nvidia.com/object/mac-driver-archive.html

^ dead..?

10 updates in 2012.. 5 in 2013

but besides, nvidia can't really be used as the example because it's their property..
quote me the developer who has said "yes, we are going to continue along our osxcuda roadmap" after wwdc2013..

 

[Cubemmal]

Apple will never support a proprietary standard like CUDA when they have OpenCL.

 

[PuppiesCakePizz]

cuda was was up until ~6months ago.. from that point on, it's highly unlikely developers keep coding cuda on os x other than finishing up something they were already far along in -or- just porting over improvements from windows versions..

Uh why?? iMacs and Macbooks have had ongoing nvidia cards. Why did the support for cuda suddenly end 6 months ago?

no it's not.. just take a wild stab at how many mac pro users are going to be affected by no nvidia.. surely it's a fraction of one percent, right?

Right, interesting way to look at it. So a 'pro' computer which does not really currently support a heck of a lot of pro software is not handicapped by this fact because supposedly a fraction of a percent of the users are apparently using pro software. So is the new 'mac pro' in actuality a hobbyist machine to your mind?

the imac is going to be outperforming the nmp in the cpu realm as well in certain tests.. understand the tests- understand your requirements.. and pick the right computer.

I understand them.

I should clarify one point, I'm not saying this looking one year down the road, I'm saying right now. Obviously Nvidia also supports OpenCL so that's not a concern. The point is right now and probably for the next year having a nMP without nvidia is very limiting.

 

[flat five]

Uh why?? iMacs and Macbooks have had ongoing nvidia cards. Why did the support for cuda suddenly end 6 months ago?

is that a serious question? do you really not understand what i've said up there? i mean, it's totally fine for you to disagree with me but you have to at least understand what i'm saying first.

So is the new 'mac pro' in actuality a hobbyist machine to your mind?

huh? is that really what you think i'm saying?
or is this just one of those sweet interweb arguments where it doesn't matter what anybody says? i mean, i'll join you for the fun of it but it's going to be more along the lines of me taking stabs at you..

I understand them.

cool.. then why all the surprise/irony by seeing the imac going toe-to-toe etcetec?

 

[jasonvp]

Obviously Nvidia also supports OpenCL so that's not a concern.

In this instance, it should be a concern. Either through Apple or nVidia's doing, the OpenCL performance on the vast majority of (even expensive) nVidia rigs is... lacking. Could this change with optimized drivers? Perhaps. But who would write them? Apple or nVidia? The latter certainly has no real reason to do so until they come to the conclusion that they're wasting their time with CUDA.

So that leaves Apple...

 

[prfrma]

Meanwhile AMD's diver support for OpenCL is...

 

[PuppiesCakePizz]

In this instance, it should be a concern. Either through Apple or nVidia's doing, the OpenCL performance on the vast majority of (even expensive) nVidia rigs is... lacking.

Sure, but that's not a huge issue ...yet. At the point at which it becomes one nvidia will have likely stepped up OpenCL performance. And if not easy enough to swap to amds latest.

At which point those 7970s in the nMP will be begging for an upgrade anyway.

 

[goMac]

In this instance, it should be a concern. Either through Apple or nVidia's doing, the OpenCL performance on the vast majority of (even expensive) nVidia rigs is... lacking. Could this change with optimized drivers? Perhaps. But who would write them? Apple or nVidia? The latter certainly has no real reason to do so until they come to the conclusion that they're wasting their time with CUDA.

So that leaves Apple...

It's Nvidia. I believe the CUDA performance isn't so great on the current cards either. The thinking is that Nvidia threw GPGPU performance under the bus for the GeForce line to boost their OpenGL performance.

Uh why?? iMacs and Macbooks have had ongoing nvidia cards. Why did the support for cuda suddenly end 6 months ago?

Apple never supported CUDA.

They included Nvidia cards for OpenCL and OpenGL, but they never officially supported CUDA on those cards.

Nvidia, entirely independently of Apple, made a CUDA driver available, but it was never Apple endorsed.

 

[Rich.Cohen]

prior to that (CS6), OpenCL support was only available in OS X.

Does CS6 support OpenCL?

 

[Quu]

CUDA will die off eventually and OpenCL will become the mainstay. That is my opinion as someone that has bought 10 NVIDIA GPU's and only 1 ATi/AMD GPU in the past decade.

It's just really obvious I think because OpenCL runs everywhere and CUDA only runs on NVIDIA cards. Any software development tool advantage that CUDA presents will be wiped out by market penetration of OpenCL compatible hardware.

I mean literally you can be on 16.1% or so of GPU's with CUDA (roughly NVIDIA's share including Intels IGP's) or you can be on 100% with OpenCL. When presented with that kind of penetration it's just too logical not to use OpenCL over CUDA.

Btw if anyone is wondering here is the marketshare breakdown:

NVIDIA: 16.1%
AMD: 21.9%
Intel: 62.0%

That is including add-in cards and integrated chips such as the ones in Intels Haswell CPU's and AMD's APU's but it is for Desktop and Notebook chips only, not smart phones or ARM based tablets.

I give NVIDIA 3 years at most with CUDA before they throw in the towel and go full speed in to OpenCL with 3rd party development tools and much better performance than what they offer now on the OpenCL side.

 

[jasonvp]

Does CS6 support OpenCL?

Premiere Pro CS6 on OS X does indeed support OpenCL. It does not under Windows.

 

[thekev]

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, switch to Linux and grow a beard.

Fixed it for you.

Well... crap. 10.7 or later may be why I am thinking this. Thanks for the info.
What about AE?

I really suspect it will be a long time before they deprecate CUDA support in AE. In some cases developers have added OpenCL support for Macs only or for Macs before Windows and Linux, but it often seems to come with a reduced feature set. That was the case with Davinci Resolve, not that I get to use the full version.

 

[riggles]

Personally, I think the adoption of OpenCL hinges on two things: software developers and AMD drivers. These are bigger factors than better performance vs CUDA. Video production is just one segment of the market that is starting to adopt OpenCL (mostly thanks to Adobe and Apple). But there are other markets, 3D for instance.

If you look at the GPU rendering landscape in the world of 3D, CUDA has a significant advantage/lead. Octane, iRay, Arion, Redshift, Bunkspeed … these are all CUDA only. These developers have to have a real reason to either stop developing on CUDA, or split their efforts to support and advance both. I think they can be motivated, but why didn't they start with OpenCL earlier. Well, some have, and they've said two things held them back: OpenCL functionality and AMDs OpenCL drivers. In fact, Vray RT does actually support OpenCL unofficially, but even then only on NVIDIA's OpenCL drivers. To me that says something.

AMD needs to up their driver game for it too succeed.

 

[prfrma]

personally, i think the adoption of opencl hinges on two things: Software developers and amd drivers. These are bigger factors than better performance vs cuda. Video production is just one segment of the market that is starting to adopt opencl (mostly thanks to adobe and apple). But there are other markets, 3d for instance.

If you look at the gpu rendering landscape in the world of 3d, cuda has a significant advantage/lead. Octane, iray, arion, redshift, bunkspeed … these are all cuda only. These developers have to have a real reason to either stop developing on cuda, or split their efforts to support and advance both. I think they can be motivated, but why didn't they start with opencl earlier. Well, some have, and they've said two things held them back: Opencl functionality and amds opencl drivers. In fact, vray rt does actually support opencl unofficially, but even then only on nvidia's opencl drivers. To me that says something.

Amd needs to up their driver game for it too succeed.

qft

Further more, know one really knows if AMD has any hands on control with these property amd GPU's.

 

[flat five]

an interesting little chain of events via Martin Wengenmayer-- developer of cheetah 3d
http://www.cheetah3d.com

neat seeing a dev go from 'wtf apple?!' to "sweet!"



june 22-- nmp is announced

Hi,
to be honest I'm a little bit speachless concerning Apples announcement. Until the announcement of the new MacPro I would have taken every bet that OpenCL is an end of live technology on the Mac since Apples interest in supporting OpenCL (fixing bugs, answering emails concerning OpenCL, etc.) was close to zero.

In its current state OpenCL is not very reliable for complex computations. For example OpenCL theoretically runs on Intel CPUs, AMD GPUs and NVIDIA GPUs. But in practice I found cases where I got three different results depending on which CPU/GPU I ran the OpenCL kernel.

Apple also has the tradition not to build in the fastest GPUs in its hardware. The result is that probably in 95% of the installed Macs the CPU is much more powerful for OpenCL than the GPU. Even on the latest Retina MacBook Pro the CPU is dancing circles around the GPU when it comes to OpenCL performance. So why limiting the Ram to 1 GByte of GPU VRAM when the CPU has access to 16 GByte or more? Why using OpenCL if it only pays out for maybe 5% of the current users?

For simple task like calculating image filters the picture might be different. For complex calculations like raytracing Apple still has to do quite some homework on the software (OpenCL framework). Lets hope Apple finally wakes up and gets some serious interest in improving OpenCL. Then the developers will finally use it.

CUDA can be considered as pretty dead on the Mac. With no Nvidia option in the MacPro I would be surprised if too many developers would support it on the Mac anymore.

So the current situation on the Mac is that CUDA makes no sense anymore and OpenCL is pretty immature.

For the near future the CPU is therefore the best bet for raytracing on the Mac. And I have quite some improvements to it in the queue. In parallel I will work on my experimental OpenCL renderer so that I have something to release once the time has come.

Bye
Martin

P.S. On the PC the situation is completely different of course.

december 19-- nmp available

Sweet, I already ordered my GPU monster. Looks like this years christmas is in January.

Bye
Martin



-----
Yes, I couldn't resist. The D700 upgrade was priced very fairly so I took the big GPU. I can't await to start playing with OpenCL on it.

december 20

Cheetah3Ds renderer and some other performance critical algorithms are multi-threaded and therefore profit from additional cores. Especially rendering profits almost linearly from more CPU cores.

I'm quite sure that the interest in OpenCL will increase once the new Mac Pros ship. And therefore also more applications will use OpenCL. So having good OpenCL performance it definitely important for the future.

Photoshop CS6 already supports OpenCL. I think it is already turned on by default in the performance preferences.

Quote:
Originally Posted by Shift Studio
Wow, so a 12-core machine could render ~ 2x faster than 6-core all else being equal.

No, not really since the 12 core CPU runs at just 2.7 GHz while the 6 core runs at 3.5 GHz.

The number of cores multiplied by CPU clock rate gives a better idea of the real performance.

6 * 3.5 GHz = 21,0 GHz
12 * 2.7 GHz = 32,4 GHz

And as you can see it's not even close to twice as fast.

Originally Posted by Shift Studio
How about d500s vs d700s? whats the approximate performance gain there?

At the moment Cheetah3D doesn't profit too much from the GPU but I want to change that. Since I don't even have my MacPro yet it's difficult to say when that will be the case. v7 would also have a rich feature without a OpenCL renderer.

I personally picked the d700 not just because it's faster but also because it has twice the memory. And you can never have enough video RAM once Cheetah3D offers a GPU renderer.

So the faster GPU is a good future investment at the moment.

Originally Posted by Shift Studio
So if Cheetah use was my main objective, the upgrades would be well worth it. Right?

I picked the 6 core since the price for the 12 core is indeed pretty steep. Especially for an approx. 50% speed up.

Originally Posted by ZooHead
I'll probably have to get a quad, and the d700 cost 1k more in that configuration.
Is it worth it, or should I sell my 1970 custom Les Paul Custom and get the six core d700.
Anyone interested?

Hi,
sounds like you love your guitar. So maybe its better to wait until Cheetah3D really can utilize the new GPUs.

Bye
Martin

http://www.cheetah3d.com/forum/showthread.php?t=9670






----------

Fixed it for you.

haha.. i missed this when you first posted it.