Low GPU utilization in 16 Beta

Just watched a video on nVidia's plans for utilizing their Maximus GPU power in compositing:

- and have a question:

- why in Fusion SA 16 as well as in Fusion page of Resolve 16, rendering a comp uses much more CPU than it does GPU?

Is it a matter of the Betas 16 not being optimized yet? In Resolve's (also 16 Beta) color page, both my Titans Xp tend to be loaded up to 100% (with my 2990WX CPU only taxed at some 50-75%, very rarely 100% when an OFX is used in a node), which used to make me think my GPUs are actually creating bottleneck in the system as it's configured at the moment, and plan replacing the Titan Xps with 2x RTX 2080ti in order to remove this bottleneck and make the CPU/GPU more "balanced". But if Fusion (SA or Resolve-integrated) is going to only use a single GPU at 20% max (while the TR CPU can be taxed even at 100%), such investment wouldn't really make sense at all*... SO BMD - please tell me what to expect in the more matured Betas, or final Release 16 of both Resolve's Fusion Page and Stand Alone Fusion - will both (or more generally - ALL) GPUs be used, and at higher loads too?

*...Or would quite an opposite conclusion make more sense: if many Fusion 16 Beta operations are heavily dependent on the Turing architecture, it would sorta explain why my Pascal GPUs aren't used efficiently? And my idea of replacing my 2x Titan Xps with 2x RTX 2080ti (or even Titan RTX, as the 11 GB of VRAM on the 2080ti cards might prove not enough, with many comps already touching the 12 GB limit of my Titan Xps)? With the caveat being that even if the Pascal architecture of my current GPUs is already sub-optimal, it doesn't explain only using one of them by the 16 Beta (unless I have mis-configured something in my Fusion Settings; is so - what could it be)?

BMD - please give us some hints

Piotr

In the past (and I think right now) not all nodes in fusion using the GPU most will compute still on the CPU.
Depending on your actual composition you will see a lot of GPU used or not

In addition to what Michael said, Resolve was written to entirely take advantage of a multi-gpu system. You actually can’t run Resolve without a card.

Fusion in the other hand, as a compositor, needs to be much more flexible that so GPU support is optional on a node-by-node and system-by-system basis. Not all modes have support yet, and comps still need to be renderable on CPU only workstations or farms.

Cheers,
Mike

Do not buy a new GPU for Fusion at this time.

What you need for Fusion to run well is fast storage, lots of memory, and fast CPU(s). A GTX 1070 is enough for most cases right now, and anything beyond that is diminishing returns, especially if you don't have the other components.

BMD has said that multiple GPUs will be supported at some point in the future, but by that time there may be new GPU's or other considerations. Right now many operations aren't supported on a single GPU, and we don't know what's the higher priority for BMD, making >99% of the tools run on a single GPU or making 50% of the tools run on multiple GPUs.

Smart planning for the future MIGHT involve making sure you can add more GPU's to your systems, assuming you aren't paying for rack space. Generally speaking getting systems with enough RAM pretty much assures that you can get additional GPUs, provided you have the chassis for it, and the chassis is not a significant part of the system cost.

Some people are still seeing better performance with OpenCL than CUDA. Things are in a weird place, at the moment.

Some people are still seeing better performance with OpenCL than CUDA. Things are in a weird place, at the moment.

Video posted above made all the confusion about Fusion implementing CUDA down the line which clearly was and is not the case to this day.

Some people are still seeing better performance with OpenCL than CUDA. Things are in a weird place, at the moment.

Video posted above made all the confusion about Fusion implementing CUDA down the line which clearly was and is not the case to this day.

I am not sure how to interpret this post. What do you mean?

Sent from my SM-N960U using Tapatalk

Video posted above made all the confusion about Fusion implementing CUDA down the line which clearly was and is not the case to this day.

That video is from Siggraph 2012 or 2013 (not sure which) and hardly has any bearing on the future of Fusion since Eyeon died on the inside and BMD acquired Fusion. Fu16 adds support for CUDA as a backend, alongside OpenCL which was introduced in Fu9. According to BMD, every node in Fu9 added support for OpenCL, whereas CUDA I believe is still working on finishing the port. If BLD could clarify that, it would be appreciated.

Stability of the GPU engine is still up for debate, and needs thorough testing from the community. However, I believe OpenCL will perform better as it’s had more time to mature and the CUDA port is mostly (seemingly) a tacked in byproduct of the Resolve integration.

Cheers,
Mike

One thing sometimes easily forgotten is that nothing happens automatically. For any processing to take action on GPU, these operations must be ported to suitable GPU compute platform, either general compute framework OpenCL and CUDA or OpenGL shader based like GLSL. Simply sticking a GPU into computer does exactly nothing unless a programmer sitting somewhere has written a gpu processing kernel for each and every operation one wants to do on gpu. Another way would be creation of generic auto-vectorization system that produces byte code for target platforms, be it SIMD CPU or GPU etc. In Nuke it is called Blink framework and I believe there is something in these lines in Fusion also, but it also means that one must write every op in that intermediate code so that code translator can do its work.

Another piece is that if GPU and CPU operations are mixed (decoding inputs, writing outputs, some ops are cpu based, others gpu based, some are single-threaded due to their nature), a considerable amount of time is spent sitting idle waiting for data from another compute device. Up to a point where gains from using mixed devices are diminishing, unless scheduling and threading logic is seriously amped up.

According to BMD, every node in Fu9 added support for OpenCL, whereas CUDA I believe is still working on finishing the port. If BLD could clarify that, it would be appreciated.

I'm not sure where they said that. I'd be surprised if even half the tools supported OpenCL at this point.

I'm not sure where they said that. I'd be surprised if even half the tools supported OpenCL at this point.

I'm pulling that statement from Bryan's blog post from Aug 2017 after Fu9 launched"

"Blackmagic claims that every tool has been rewritten to offer GPU acceleration with OpenCL. ... Although I can say for certain that almost every tool does, indeed, now run in OpenCL mode"

http://www.bryanray.name/wordpress/fusi ... eleration/

Their site material as well also makes it seam like 3D nodes are now GPU accelerated, but I'm not seeing anything of the normal GPU enable/disable sort on any of the nodes. Unless they mean they use the GPU for OpenGL, which is an entirely different meaning of the word accelerated in this context.

Cheers,
Mike

Well, since I've been brought into this…

That statement was based on conversations at SIGGRAPH. The claim was that "everything" was GPU accelerated. Investigation showed that most of the 2d image processing stuff did, indeed, have OpenCL modes. 3D and particles did not. Given how many tools fall into those two categories, Chad may be right that around half support OpenCL. I don't have any inclination to count up the total number of tools with a Use OpenCL switch on them.

Please do note the theme of that article, though: Although much of Fusion was capable of accessing the GPU at that time, very few tools saw any kind of improvement from doing so. CPU mode was faster on the majority of workstations. I suppose slowing things down is, technically, still acceleration…