Blackmagic Forum

The ones they removed the graphics ports from.. so that miners would buy those & not the gaming cards.

Since the CUDA cores are still there.. seems like a good bet for DV Resolve too no? Assuming the price is interesting of course..?

Seems like it could work...

Can you provide a specific model number ?

Turing: 30HX, 40HX, 50HX

Ampere: 90HX

I notice the these have no display connectivity, as it is not designed to have monitors connected to it.

And they'll be all sourced by the miners too…

Peter Chamberlain wrote:I notice the these have no display connectivity, as it is not designed to have monitors connected to it.

Yes - but that's not a problem is it? Either if accessed remotely or as 2ndary cards the display ports are not necessary....

And if they are cheaper because gamers don't want them, that's still better for Resolve-ers, yes? Point is it can be an -option-, if the choice is there.

My question is would they still be useful for Resolve since the CUDAs are technically still intact..?

Nothing?

From reading this forum for several years it looks like BMD now recommend that the same card is used for display and compute, so yes, it could be a problem.

For V15 and earlier they used to recommend separating the cards for best performance, but no longer for V16 and up.

Mike Warren wrote:From reading this forum for several years it looks like BMD now recommend that the same card is used for display and compute, so yes, it could be a problem.

For V15 and earlier they used to recommend separating the cards for best performance, but no longer for V16 and up.

How does that equate with systems with multiple graphics cards though, since there it's obviously not possible to use --all- of the compute cards for display output.

And further, I like to use RDP, so in some cases I don't care about the display output. But I'd like the shaders or whatever Resolve likes best to still be available for use. So if these cards are just as good for Resolve as a display-toting card, but then cheaper because of no output.. then that's great for Resolve-ers too.

So there are at least two reasons to find this interesting:

- Cheaper cards for the 2ndary, 3tiary processor
- Cheaper graphics processor for headless/remotely-accessed systems.

Reason to be interested, no?

They're not particularly well suited on paper for Resolve. There's no point to use additional cheaper cards for graphics operations, as Resolve will be limited by the slowest card in the system. As a graphics processor on a headless workstation that won't do any color critical work, maybe, but that's an extreme edge case for this software.

Hmm ok.. Then maybe I misunderstood something.. when using multiple GPUs for Resolve you only ever get performance equal to the slowest card? Then what is the benefit then of using multiple GPUs?

No, performance does add (no way linear though, so 2 cards doesn't really mean 2x performance). VRAM doesn't, so if you have 1 slow card with little VRAM and 1 very fast with way more VRAM then you will be de-grading yourself in reality when used both.
If you want to use multiple cards then the best is to use the same model.

footofwrath wrote:what is the benefit then of using multiple GPUs?

This was posted by staff in another thread.


Resolve uses the GPU memory on every card it has in the system that is selected for Resolve use.
Generally it doesn't share memory to other cards.. lets not get into exceptions.
If you have a GPU in the system but its not selected in preferences, its just a waste of space.
If you have mis matched GPUs, the slowest GPU will define how fast the system can process.
If you have GPUs with difference memory capacity, the lowest capacity will define how big an image and the type of processing before its runs out of memory. The cards with extra memory waste that memory.
If you have a separate UI GPU, that's something we generally don't favor now, whereas we did 10 years ago.
If you have one more powerful and more memory GPU compared to a second, use the better one for the UI but not as UI only. It might make a little difference if the gpu performance is close.
One very powerful GPU with lots of RAM is better than two average GPUs, for many reasons....
Anyone who says multiple GPUs are a waste of time are picking a use case that doesn't need them, after all the large facilities that put 4 x 24GB GPUS in their system are not spending that money for fun.
Most users will be happy with one good GPU.
All image processing is in GPU.
Some new GPUs, with the Resolve Studio version. also support compression and decompression in a hardware stage in the GPU.
GPU driver selection makes a big difference in reliability.

That seems very very inefficient.. why not allow Resolve to use each card up to its actual capacity..? I get the VRAM deal for the UI, you can only use the bank in one card to actually hold each block, but ignoring half a card just because there's another card in the system that is weaker (regardless of the config for UI or otherwise) seems veerrrryyyyy un-ideal.

I have two PCs for Resolve, both with Ryzen 9s. Should I sell these off, buy Ryzen 3s instead and put the cash towards much better GPUs?

footofwrath wrote:That seems very very inefficient.. why not allow Resolve to use each card up to its actual capacity..? I get the VRAM deal for the UI, you can only use the bank in one card to actually hold each block, but ignoring half a card just because there's another card in the system that is weaker (regardless of the config for UI or otherwise) seems veerrrryyyyy un-ideal.

As far as I know this is how CUDA processing works over multiple cards so this is not up to BMD to fix.

footofwrath wrote:That seems very very inefficient.. why not allow Resolve to use each card up to its actual capacity..? I get the VRAM deal for the UI, you can only use the bank in one card to actually hold each block, but ignoring half a card just because there's another card in the system that is weaker (regardless of the config for UI or otherwise) seems veerrrryyyyy un-ideal.

I have two PCs for Resolve, both with Ryzen 9s. Should I sell these off, buy Ryzen 3s instead and put the cash towards much better GPUs?

It's not easy to do at all and it would probably require re-writing whole Resolve engine from scratch, which means were are almost back to Resolve 1 with all possible bugs etc. Huge task and decision not mentioning money needed for it.

Reading about the limitations of using multiple GPUs, I would just make sure the non-display GPU matched or exceeded the specs of the display GPU. As others have said, using 2 GPUs doesn’t fully double the performance, so it may be better to buy a single, higher spec GPU than 2 lower spec GPUs.

True, problem is when you need even more speed and there is no "faster" GPU out there anymore

OK, that makes sense, I am never going to need the power of an e.g. 3090 so I'll get the single fastest card that fits my purpose.

But should I then sell my 3900x's and put R3 instead, and use the surplus to pump the GPU? Or are the threads still useful in some tasks?

CPU is required for any encoding or decoding that isn't GPU accelerated, which is most of it.

Hi.

In Resolve the CPU is used to run the app, disk I/O, fusion, compression and decompression of codecs.
Resolve does all its image processing in the GPU on the graphics card. More CUDA/OpenCL Cores are better.

Regards Carsten.

So what's the general consensus then, max cash on GPU and mid-range CPU, or relative balance of CPU/GPU market point?
Again, I am running two Ryzen 9 machines. Would I be better off (i.e., CPU-sufficient in Resolve) with a Ryzen 3 w/ 3080, or the 9 with a 3070? I realise this is probably the most common question on this board ) Sorry for that

Don't forget disk I/O and RAM.

There's a lot of stuff going on and many opportunities for bottlenecks. Sometimes I wish Resolve would better report when you're hitting those. Like a floating profiler window or something.

footofwrath wrote:OK, that makes sense, I am never going to need the power of an e.g. 3090 so I'll get the single fastest card that fits my purpose.

But should I then sell my 3900x's and put R3 instead, and use the surplus to pump the GPU? Or are the threads still useful in some tasks?

If your timeline is relatively easy (no NR etc.) then during export CPU with many cores will be useful.
If your timeline is difficult your GPU will saturate way earlier than CPU (unless you are exporting h264 on CPU), so having many cores won't help for render times.

I would use your idea may be actually good one and overall give you better system. It all depends how your typical Resolve projects will look like. Many NR, blur, effects ? Or simple ones ?

Disk I/O problems are easier to handle.
RAM - well it's what it's I don't think you can gain/loose that much, by using faster/slower models (5%?). Not in Resolve itself.