GPU Memory, OPENCL/CUDA in Resolve 1x

Hello,

I am having major issues editing UHD video from my Sony PXW-X70 since the upgrade to Resolve 14 Studio due to the dreaded "GPU Memory Full" error message. I have 3 GB of RAM on my Radeon R9 280X which is lower than the recommended specs.
Using GPU memory monitoring tools, I noticed that even though the card has 3 GB, the program stops working and displays the error message when GPU memory usage reaches 2-2.5 GB. This makes it very difficult to know exactly how much memory is really needed.

I do not need a superfast GPU for complex color grading, I guess just more GPU RAM in order to be able to simply use the software for editing. When I built my new PC, CPU and RAM were the most important (I was planning on using Final Cut Pro), now it seems to have shifted to GPU with Resolve 14. I do not do huge color grading operations but appreciate the power of Resolve when I need it. I am a low end amateur, I do not use Resolve as a Pro. I output to compressed h.264. Because I invested in the Studio version license, I am pretty much stuck with Resolve now and need to find a fix to the GPU Memory Full problem.

Looking at all the information about CUDA vs. OpenCL, it seems like BMD has clearly chosen NVidia and has put more optimizations in place to take advantage of their GPUs vs. AMD. I am open to switch from AMD to NVidia in order to take advantage of all this, but the problem is that an 8GB graphics card equates to a very high end and expensive GPU. Radeon cards with 8 GB are more reasonable, but still. Whereas on a motherboard, one can choose the CPU and memory independently, on a graphics board there is a direct link between the amount of RAM and the GPU class you have to buy. As a consequence an 8 GB graphics card costs a minimum of $300 for a Radeon RX 580 and $410 for an NVidia GTX1070. I wish they made a GTX1050 or RX470 with 8 GB for the DaVinci Resolve market but that's not the case (yet?), or made cards that have RAM slots to allow for upgrades. Until then I am forced to buy more GPU power than I need, just to be able to fit the UHD video I process into the graphics card's memory.

I would prefer a GTX to be able to take advantage of the h.264 and h.265 encoding/decoding acceleration today, but given my budget that forces me to buy a GTX 1050 with only 4 GB or a 1060 with 6 GB. How can I know if 4 or 6 GB of RAM will be sufficient to process UHD video? Has anyone been able to process UHD with 4 GB?
Secondary question, is it possible for BMD to start implementing more OPENCL optimizations in the future that will put Radeon cards on equal levels with GTX cards in terms of h.264/265 acceleration? Is CUDA so unique? Will BMD start supporting h.264 acceleration from Intel chips (quicksync)? If QuickSync or Radeon h.264/265 acceleration are on the roadmap for a future release of Resolve, I could save money and buy a Radeon card with 8 GB of Ram, knowing that I will not miss on the CUDA specific acceleration for too long.

I am sure that I am not the only one facing a potential investment in hardware to support Resolve 14 and up, it would be nice to have an idea of what we should invest in today that Blackmagic Design plans to support in the next versions of their software.

I had similar problems with my Radeon HD7970 3GB. Got GPU Memory Full error, very frequently. Then thumbnails looked like random noise, and usually Resolve stopped displaying the correct video frames when playing back.

I had given up on trying to make this gpu work for my audio workstation and bought the cheapest GTX 1050 Ti 4GB I could find (For about 180 euros), and that solved the problem. I'm not getting these GPU Memory Full errors anymore, and can even edit 4.6k raw from the Ursa Mini Pro. Playback was not realtime, but there were no GPU Memory Full errors. I could add up te 8 correction nodes, with one having a temporal noise reduction of 3 frames, then pressing playback choked Resolve, and it finally threw a GPU Memory Full Error.

Hi Nick,

Thanks for sharing your experience with a GTX 1050 4 GB. It looks like 4 GB is not enough RAM for 4k processing as you are still getting the Memory Full error message when playing back the timeline.
What happens if you do not attempt to playback, are you still able to use the Deliver tab and render the video without getting the Memory Full error?

I found the solution: downgrade from Resolve 14 to 12.5. Everything works, only very slightly slower. But so much more stable!
At first I tried to keep all of the 3 GB GPU memory for R14 processing and enable my iGPU for Resolve's GUI. No luck, Resolve would crash at startup before displaying the list of projects. Tried to uninstall and reinstall, hoping that the base settings would get reset and R14 would adapt its GUI to the iGPU. Still no go, so I installed 12.5.6 instead.

Version 12.5.6 Studio works very well, takes advantage of both iGPU and the AMD RX 280 with its 3GB of RAM.

So for me, until BlackMagic gets their act together and fixes the GPU memory allocation bugs according to the resources available in the system and not some arbitrary number, I will stick to version 12.5. It just works.

By completely rewriting the graphics processing exclusively for high end GPUs in v14, BlackMagic took a step back in their quest to open up their platform for editing enthusiasts and luring Premiere and FCP users. Version 14 is back to being 100% targeted at color grading professionals. It cannot do simple UHD editing without a super high end video card.

I refuse to buy a $400 GPU just because of the vague recommended 8 GB requirement of Resolve 14 without knowing the direction of the platform. Until we can clarify CUDA vs. OPENCL and have accurate GPU RAM requirements for NLE of UHD streams, I will stick to 12.5.

Personally i used 960 4gb and now 1060 6gb, both with success edits your 4k footage, 960 being clearly slower.
As long as you don't require realtime 60 frames playback, or you optimize all the video to about 1/4 of the size, you're good to go.
Take into account that no nvenc/hevc (h.264/h.265) NVIDIA acceleration seems to take place on current resolve 15.
This is more about CUDA vs OpenCL and i wouldn't say ATI loses before i can get my hands on some recent Vega chip or some other beast red card.
You can also see what BMD recommends on Configuration guide. The more VRAM - the better Remember that you're limited to 1 GPU only with a lite license, so it's better be the best one you can afford