Resolve needs an architecture change

Related threads have got me thinking of how Blackmagic might enhance functionality on platforms other than MacOS. Let me begin by outlining some facts.

1. Apple systems, with their integrated memory usage, can allocate as much RAM as needed to graphics functions. The result is that these computers can easily edit high resolution (above 4K) files without a problem. So long as you have 16 GB, you are good to go.

2. Windows users regularly get "Your GPU memory is full" errors if they do not have enough memory on their Nvidia [edit] graphics card. For example, 64 GB RAM might well go mostly unused, while 8GB on the GPU limits editing to 4K. Since you cannot easily add more GPU memory, this is a "hard limit" that forces users to upgrade to (very) expensive hardware.

3. The Studio version might improve performance in certain cases, but doesn't change the basic architectural problem.

4. I don't known anything about Resolve on Linux, but I imagine it's the same situation as Windows.

5. Other NLEs use the CPU for graphics functions and hence are not limited in the same way. This can make Resolve look bad by comparison, depending on the task. (A recent thread reports that even Shotcut, a free program, performs tasks Resolve cannot handle.)

6. Admittedly, performance on the GPU is faster than performance on the CPU, given today's hardware. But sometimes speed isn't everything. A slow render is better than a failed render.

Now my conclusion.

Blackmagic should innovate their code base, allowing graphics functions, especially those that are GPU accelerated by default, to use system RAM if there is not enough GPU RAM. Resolve would then lose its disadvantages vs. other software, and non-Mac systems would be on par with Apple computers regarding this functionality.

The cost barrier for high resolution editing would disappear. User frustration would decrease. All those tech support calls about "memory is full" errors would be a thing of the past.

This could be a huge marketing win.

2.

It's not completely true, my laptop show me 40GB GPU memory, 8GB integrated and 32GB shared.

Sent from my CPH2025 using Tapatalk

2. Windows users regularly get "Your GPU memory is full" errors if they do not have enough memory on the graphics card.

Hi.

It is only nVidea Graphics Cards that get "Your GPU memory is full" error in Windows. It never happen on AMD Graphics Cards.

Quote: 'While our 8K test project in CUDA mode with an RTX2080Ti repeatedly breaks off due to memory overflows despite 11 GB DDR6 RAM, strangely enough, this does not happen with the 6GB AMD card. This acts much more slowly, but does not stop processing because the memory is full.'

From this Google translated German link: https://www-slashcam-de.translate.goog/ ... r_hl=en-US

Regards Carsten.

I have observed that Quadro cards and drivers may make better use of shared GPU memory. I had T600 4GB card report GPU memory full only on the 6 TNR test of the UHD candle benchmark version of test. Speedwarp also passed on 4K footage as far as I can remember. Watched the shared memory graph and it filled up quite nicely. I was positively surprised.
On the other hand, never a memory full with AMD, even on 8K footage . It indeed seems to make good use of shared memory.

Sent from my Mi 9T using Tapatalk

It is only nVidea Graphics Cards that get "Your GPU memory is full" error in Windows. It never happen on AMD Graphics Cards.

Thanks for the clarification. I do wish I'd known this before I specified, priced, purchased, and built a new system.

That might explain why we had far less troubles in our student labs with Apple systems of the same VRAM size than with PCs. Apple has not been using Nvidia for quite some time, while our IT guys insisted on Nvidia for performance.

On Linux - it's a relatively easy install nowadays, next next done.

Generally performance is noticeably faster on Linux though render performance has got more competitive with Windows catching up on 1080p and being much faster with 4k footage (not sure if devs are working to improve a huge performance knock on Linux).

I don't use any GPU encodes as the quality is just too crappy. I have never experienced any "memory full" errors either, though my system is all AMD based. I generally output to sequenced images or cineform and wav then run through ffmpeg for my final output. This way I have the best quality with a very small file. GPU encodes just can't touch CPU encodes at this point quality for a given size wise.

Thanks for that interesting info. It's always nice to know what the Linux experience is.

I went with Nvidia because everything I read indicated that these were faster cards overall. The third gen RTX have advanced features like ray-tracing that are simply not present with AMD. I try not to be swayed by the game sector, but most of the GPU reviews are indeed targeted to those users.

Even release notes from Blackmagic pointed to optimisations that were specific to Nvidia. Shouldn't they be recommending AMD explicitly? Or at least documenting this issue?

I teach digital video and so am responsible for disseminating information to many people. I have just updated the document I distribute to students, to fully incorporate AMD GPU information.

In the meantime, the basic thrust of the OP is still correct: Blackmagic need to optimise their architecture to fluidly use memory and eliminate these memory errors.

I wonder how I can bring this to their attention?

Nvidia are technically faster, but do you really save time when your work is interrupted by "memory full" ? We found AMD generally more reliable, even if slower.
I second the recommendation to separate rendering and encoding (to GOP codecs), since it seems to provoke the problem with Nvidia more frequently and the quality is better when encoding externally anyway.

I understand the OP's request, but of all the things that BMD could be spending their resources improving, this would not be at the top of my list (even though I have also complained about the lack of scalability in the past). Resolve is a relatively inexpensive software (actually, free unless you pay for the $299 studio license) but that masks the fact that Resolve requires very expensive hardware to run as intended. This can catch many users off-guard because it seems at first glance like a very cheap way to get into advanced post-production.

I do agree that the issue with nVidia hardware should be addressed, if possible (assuming it isn't something fundamental to the nVidia architecture).

I do agree that the issue with nVidia hardware should be addressed...

Yeeeesss!!!