The claim initially comes from the Octane forum and expanded upon here.
Supposedly, AMD has mentioned plans to drop OpenCL altogether amid the lack of interest by both Apple and Nvidia to support it properly. If this is true, then this would sadly be yet another case where open standards lose to their proprietary equivalent (meaning a better chance of the owner using that to try to achieve a monopoly via the concept of vendor-lock). Perhaps the writing was already on the wall when AMD announced their project to give their cards the ability to run CUDA code.
Assuming the rumor is true, this would not affect Cycles so much as nothing would be lost and the core team has not spent near as much time on it as they did on CUDA (not to mention that the split-kernal stuff will still likely benefit Nvidia cards), but it’s likely that it would end up hosing the Luxrender developers. Also with the case of shifting GPU API’s, I’m left to wonder if Cycles could’ve been that super-fast render engine on all hardware combinations if the time spent tweaking GPU code was instead spent on optimization work for CPU-rendering (amid claims of how engines like Corona manages to be super fast on the CPU).
I feel like AMD doesn’t have the right people working for opencl. Maybe it’s their hiring process that have failed them. It has been proven time to time that the sole issue with amd are their drivers. It’s like they make fast cars but can only go to the 3rd gear.
That’s why I’m asking whether it’s credible or not (there’s a number of Octane users here for one thing who probably saw that post).
Also, why I was asking is because its credibility could affect the development direction for Cycles (does it mean we are better off removing all OpenCL kernels and using the full extent of CUDA’s functionality)?
I don’t even know if AMD has much say in this matter (Nvidia crippled OpenCL on purpose and I think it’s actually Apple who writes the drivers for their hardware on Mac). Nvidia has achieved a virtual monopoly in the GPGPU space and has much of the power as to its direction and how open it is (Blender being kicked out of that area due to CUDA becoming incompatible with FOSS licenses, they could do it).
So they would bee working on a very fast pathtracer, a machine learning program framework, etc. knowing they will drop their only way to get into GPGPU? Same as second thread, don’t feed the rumor…
No rumors, but speculations. Bad and some even illegal.
Ah, Ace you show symptoms of an addiction, you know that? News or gossip… it’s sickness.
Although.
Probably just another misinterpretation.
Scientist tends to confuse the consumer/user by knowing too much and explaining too little.
Probably something along the lines of “Boltzmann Initiative”. If so, OCL doesn’t die, but evolves on & soon will be possible to run CUDA apps on the AMD cards.
“Whaaat? I don’t get it.” “Aay, just burn’em at the stake!”
You guys do realize that rendering isn’t the only use for GPGPU, right, and there are plenty of applications that rely on OpenCL exclusively. I can sort of see AMD not supporting OpenCL on consumer drivers, but I have a hard time seeing them leaving it entirely given that it’s used extensively in visual effects and scientific settings.
It isn’t like Nvidia doesn’t support OpenCL, and ever since Maxwell OpenCL performance has been good. Abandoning OpenCL outright by AMD would pretty much hand the high end market over to Nvidia. Being that they’ve been putting a lot of effort into their professional GPUs lately I really have a hard time seeing this.
If that’s true, I’ll just say “good riddance”.If OpenCL was well-conceived to work on GPUs, it would’ve been implemented properly by now. It’s better to not support it at all than to support it poorly.
OpenCL isn’t the only “non-proprietary” way to do GPGPU. You can use Compute Shaders or at least Vertex/Fragment Shaders on every GPU nowadays. OpenCL was a way to (re-)use “regular” C code (with many restrictions) on the GPU, but that’s probably not such a good idea in the first place.
There’s also many technologies that can turn (a subset of) some language into instructions for a particular GPU architecture, into IR (like SPIR-V or PTX), or into shader code. Some of that stuff is open-source, some by AMD itself. If they really ditch OpenCL, they’ll provide for a migration path like they are doing with CUDA (HIP).
Are there really that many applications using OpenCL to a vital degree? That’s not my impression. The programs used in scientific settings aren’t necessarily that complex and may well be easily portable.
Abandoning OpenCL outright by AMD would pretty much hand the high end market over to Nvidia. Being that they’ve been putting a lot of effort into their professional GPUs lately I really have a hard time seeing this.
NVIDIA pretty much owns that market anyway. I don’t see that much effort going into their professional GPUs either, besides marketing. Those are basically the same chips, running more-or-less the same software, sold at a premium (with some extra bells and whistles).
I stand corrected. I thought that Adobe and Autodesk both used OpenCL more than it appears they do. In fact, Adobe doesn’t seem to use it much at all. This pretty much leaves Houdini. Not sure what Bullet 3 is supposed to be using? I’d assume OpenCL.
HAhaha, Really? OpenCL is the only compute platform AMD have, Scratch that, That ALL vendors other than Nvidia have.
And Baron going on about compute shaders is even more laughable, The only way i could see this happen is if Otoy give their so called miracle cross compiler open source, but last i heard they were complaining about trying to support drivers which is why it hasn’t been released yet in any product form.
Funny that, AMD get killed about drivers even though they work on par with Nvidia drivers for at least 12 months now but yet were supposed to beleive AMD would just Dump OpenCL. Phhhfffffffff. what a load of shite
AMD have ground breaking new hardware coming out and about to be released but you think they would let them have no dedicated compute platform?? One of the biggest new markets will be in GPGPU compute with cars etc, like Nvidia have done but im sure faster and cheaper.
OpenCL 2.1-2.2 is as capable if not more capable as can run in unison with new CPU’s but cuda can never take such advantages, has the ability to actually kill Cuda in performance and compatibility at that point. Just cant see this as true, UNLESS OTOYS MIRACLE CROSS COMPILER is open sourced (and i still half believe Otoy are full of shit on this).
Non sense. If you beleive CHsociety forums for things like this your mental, As one dick head said lately, Fake news.
This is just not true. Look at the Boltzmann/HIP thing, it doesn’t target OpenCL, it targets AMDs own compiler (and CUDA). Boltzmann will let people run a more CUDA-like subset of C++ on the GPU. HIP will migrate CUDA code to this language. OpenCL isn’t in the picture at all.
And Baron going on about compute shaders is even more laughable, The only way i could see this happen is if Otoy give their so called miracle cross compiler open source, but last i heard they were complaining about trying to support drivers which is why it hasn’t been released yet in any product form.
I don’t see what’s laughable about Compute Shaders, they’re widely applied in games nowadays. Nobody uses OpenCL there. Before OpenCL/CUDA, GPGPU was done with just vertex/fragment shaders. It works, it just doesn’t let you pretend you’re writing a C program, like OpenCL.
Funny that, AMD get killed about drivers even though they work on par with Nvidia drivers for at least 12 months now but yet were supposed to beleive AMD would just Dump OpenCL. Phhhfffffffff. what a load of shite
Let’s put this story into context: There is a claim by an Octane developer, that AMD told “them” they were about to drop OpenCL, in favor of Boltzmann. It’s plausible. Octane doesn’t have to target OpenCL, they can target HIP.
HIP/Boltzmann right now does not look attractive, so why would anyone target it when there’s OpenCL? If you remove OpenCL, all you have is either CUDA or HIP/Boltzmann. HIP/Boltzmann lets you target CUDA too.
AMD have ground breaking new hardware coming out and about to be released but you think they would let them have no dedicated compute platform?? One of the biggest new markets will be in GPGPU compute with cars etc, like Nvidia have done but im sure faster and cheaper.
AMD has a terrible reputation with OpenCL. Even if they’re supposedly better now, that reputation isn’t going away. Might as well “start fresh” with something else.
I certainly wouldn’t want to ride a self-driving car running OpenCL software on an AMD GPU…
OpenCL 2.1-2.2 is as capable if not more capable as can run in unison with new CPU’s but cuda can never take such advantages, has the ability to actually kill Cuda in performance and compatibility at that point. Just cant see this as true, UNLESS OTOYS MIRACLE CROSS COMPILER is open sourced (and i still half believe Otoy are full of shit on this).
I never saw this CPU/GPU union as a big argument. If you want to take advantage of a GPU, program it like a GPU. The same goes for CPUs. If you run the same code on both, you’ll lose performance on either. If you’re writing code twice, you might as well use two different languages.
Non sense. If you beleive CHsociety forums for things like this your mental, As one dick head said lately, Fake news.
The actual quote is from an Octane developer, on the Octane forums.
What makes you think I’m angry about this? Is it my avatar?
I’m not invested into GPGPU, so I don’t really care what exactly happens to it.
All this is pure speculation for now. Got it?
An Octane employee is on the record saying that AMD employees told them to drop OpenCL and use the Boltzmann stuff instead. That part isn’t speculation, it’s not even a rumor, it’s a fact. I don’t see why that employee would fabricate this, so I’m willing to speculate that this really happened.
Of course, there could be a misunderstanding here as to what exactly “dropping OpenCL support” means here. It might as well mean “we’re not fixing our driver bugs for Octane” and “we really want to focus on Boltzmann”. In that sense, AMD users wouldn’t lose anything they already have (i.e. broken OpenCL support).
Lastly, Cycles might target Boltzmann just like it targets CUDA, OpenCL and C. It’s probably not that much work. Don’t worry about OpenCL going the wayof theDodo.
Well this page is dedicated to GDC 2017 and Pro render and Radeon Rays in a Gaming Rendering Workflow (Presented by AMD)
As we all know Radeon pro render uses OpenCL, If AMD are doing a tech talk on real world integration and utilization it goes against everything that Otoy employee has said. Never believe a word Otoy say.
More likely Otoy are a bit worried about how fast Pro render is.
Not necessarily. It uses OpenCL 1.2, which can be somewhat easily ported over to something like CUDA or Boltzmann, unless you use stuff like runtime compilation (which afaict they don’t). It’s rather that going from CUDA/Boltzmann to OpenCL isn’t straightforward (or else Boltzmann could just target OpenCL).
ProRender may also run just fine on current drivers, unlike whatever Otoy has (not) running, which also may require OpenCL 2.x.
Either way, it’s not that unusual for one division of a company to work on something that is in conflict with another division, or to have their project scrapped because some higher-ups decide they need the people elsewhere (or not at all).
Never believe a word Otoy say.
This is not a “company statement”, it’s just one employee saying something on a forum.
More likely Otoy are a bit worried about how fast Pro render is.
I don’t see how they should be worried, ProRender has a pretty basic path-tracer feature set, so the only thing it really has got going for itself would be speed and cost. I don’t think users care about speed that much, if it’s still in the same ballpark with other path tracers (and why wouldn’t it be?).
Anyhow, the take-away is to just not get too invested into one particular API. OpenCL would be the third graphics/compute API that AMD has dropped in the past ten years.
…but googling Boltzmann and opencl and there are many results.
There are many OpenCL programs using Lattice Boltzmann methods.
In the context of this thread, I’m talking about the AMD Boltzmann initiative. It’s a set of technologies that allow you to program GPUs, both from NVIDIA (running on top of CUDA) and from AMD, using a C++ subset. It also includes a tool to migrate CUDA code. It isn’t compatible with OpenCL, but there might be a migration path just like with CUDA.
And I don’t see the problem with dropping OpenCL if they have a solution that will make it work as smooth as CUDA. Then the hard work has to be on AMD side to make it work instead of developers of the software.
