What To Upgrade Next? CPU vs GPU

Just a datapoint... I've been using an RTX3060 for several months now (upgraded from a GTX1060/6) and I chose it as it had the best price/performance ratio (for use in Resolve) on the budget I could afford.

So far it has performed more than adequately on lightly graded 4K 50FPS and 60FPS timelines using H265 encoded source material with H265 delivery. The 12GB of VRAM was the key point for me and I see that even using my relatively light grading and a bit of temporal noise reduction Resolve regularly uses 9GB or more of VRAM -- which makes me very glad I didn't go for a card with only 8GB of VRAM even if (like the 3060ti) it might have had more CUDA cores.

I'm happy to sit where I am for a few years until the 5000 series of RTX cards is announced, at which time the 4000 series may represent another reasonable price/performance ratio.

You don't need the "best" to do the job, "adequate" is a perfectly good place to be.

Short version is that I'm kicking myself for upgrading my PC's 8 months ago... but hey, my 6700XT and 5900X combo is working quite nicely so I'm not complaining too much. Anyway, hope that's some useful food for thought.

Why kick yourself for having good performance for 8 months. I thought the same when I bought this 6800XT for 800 in the middle of the mining insanity...But the difference paid itself over and over and over...

You don't need the "best" to do the job, "adequate" is a perfectly good place to be.

Nothing wrong with the 3060...but I was referring to price/performance which is not good. Same with the 4080 12GB...The difference in price is small-ish (relative) vs the 4090 but the difference in cores is HUGE. 7,680 vs 16384...

Is the 4080 12GB adequate? Yes. Is it good value? Nope. Its actually a joke @ 900usd.

What you talking about? 1080Ti does not even have tensor cores that Resolve is increasingly using.

So what? Makes absolutely no difference. Every improvement I've seen (including the new RTX4000 series presentation where the 4090 was 70% better than a 3090ti in "Ai") comes from extra CUDA cores. 70% more CUDA cores --> 70% more performance.

As far as I know Resolve is FP32...and the tensor cores are ...NOT.

From 2018

What are Tensor Cores?
While already available on the more expensive Titan V GPU, the RTX line introduces tensor cores at a more reasonable price point. These tensor cores operate alongside the normal CUDA cores that traditionally do the heavy lifting, but are designed specifically for machine learning inference (running already created and trained machine learning models). Blackmagic has already announced that they will be using these cores, but exactly how they do so and where exactly it will improve performance is still to be seen.

What are RT Cores?
RT cores are brand new in this generation of graphics cards, and are specialized for a single type of operation: ray tracing. It is possible that Blackmagic may utilize these cores for ray tracing in the Fusion tab, but if or when they will take advantage of these RT cores is currently unknown.

From this forum before last Mac's

Rohit Gupta
Forum: DaVinci Resolve
Topic: Exports in Beta 7 : Mac vs PC - what's wrong? SpeedWarp
Replies: 11
Views: 1411
Re: Exports in Beta 7 : Mac vs PC - what's wrong? SpeedWarp
Thu Aug 01, 2019 7:14 am

DaVinci Neural Engine which SpeedWarp is based on, and plenty of other algorithms introduce in v15/v16 are most optimised for Metal and CUDA. On Windows, you should get significantly faster renders than your Macbook with a modern NVIDIA GPU, like the RTX or Volta series which have Tensor cores.

Rohit is Blackmagic developer.

So the 1080Ti probably cannot use Resolve Neural Engine while 3060 can.
Note that 4000 series appears to have got a big up in number of tensor cores.

Blackmagic has already announced that they will be using these cores, but exactly how they do so and where exactly it will improve performance is still to be seen.

...

"What are RT Cores?" Maybe when Resolve will add DXR raytracing...which will be maybe...never?

...

"DaVinci Neural Engine which SpeedWarp is based on, and plenty of other algorithms introduce in v15/v16 are most optimised for Metal and CUDA. On Windows, you should get significantly faster renders than your Macbook with a modern NVIDIA GPU, like the RTX or Volta series which have Tensor cores.

Rohit is Blackmagic developer.

Nothing he said means Resolve is using Tensor Cores... However, if Resolve would use Tensor Cores then it wouldn't run on AMD cards and it runs fine...Also if these tensor cores were used...you would see a better improvement than 70%...Where do these tensor cores matter?

Every single improvement Nvidia as made in Resolve has been a result of more CUDA cores. Exactly in line with their increase...Again...70% more CUDA cores for 70% more performance...

"Fourth-generation Tensor Cores with up to 1.32 Tensor petaflops — 5x over the previous generation using FP8 acceleration."

Where's the 5X? 500X? in Resolve...Come back to me when Resolve stops using FP32 and starts using FP8 so the tensor cores will matter...
If tensor cores are used anywhere in Resolve its where it doesn't really matter to make a difference.

So the 1080Ti probably cannot use Resolve Neural Engine while 3060 can.
Note that 4000 series appears to have got a big up in number of tensor cores.

Yes it can. Any GPU capable of FP32 is capable of FP16 and so on...And I doubt the "neural engine" is FP16...

Nothing he said means Resolve is using Tensor Cores...

DaVinci Neural Engine which SpeedWarp is based on, and plenty of other algorithms introduce in v15/v16 are most optimised for Metal and CUDA. On Windows, you should get significantly faster renders than your Macbook with a modern NVIDIA GPU, like the RTX or Volta series which have Tensor cores.

So why the Blackmagic developer made reference of tensor cores if they are irrelevant?

So why the Blackmagic developer made reference of tensor cores if they are irrelevant?

Because they're used in stuff that doesn't really matter in the big picture. Very limited items like maybe the Magic Mask or the face refinement stuff. Which can also be done on normal FP32 shaders...so in the end it doesn't really matter...

I would take 16K CUDA cores over fewer CUDA cores but full of whatever tensor core junk nvidia comes up with next...

Debayering, color corrections etc are all done using the normal shader cores.

So why the Blackmagic developer made reference of tensor cores if they are irrelevant?

Because they're used in stuff that doesn't really matter in the big picture. Very limited items like maybe the Magic Mask or the face refinement stuff. Which can also be done on normal FP32 shaders...so in the end it doesn't really matter...

I would take 16K CUDA cores over fewer CUDA cores but full of whatever tensor core junk nvidia comes up with next...

Debayering, color corrections etc are all done using the normal shader cores.

If he bothered to talk specifically about tensor cores in a context of faster significant renders then it matters.
Edit:
For DaVinci Resolve 16
https://www.nvidia.com/en-us/geforce/ne ... tive-apps/
Blackmagic Design Transforms Video Production With AI
Blackmagic Design announced the latest version of DaVinci Resolve with DaVinci Neural Engine. The application includes new AI-powered features to make video editing and color grading more powerful and easier.

The DaVinci Neural Engine uses several NVIDIA AI libraries and Tensor Cores found on the latest NVIDIA RTX GPUs to accelerate inferencing. New RTX-powered features include:

Speed Warp, which interpolates frames to change the playback speed of videos and convert clip frame rates
Super Scale to increase footage resolution by up to 4x
Auto color and color matching
Facial recognition for automatic tagging and tracking of characters for smart bins
Stylize for transferring looks from one clip to another
“DaVinci Resolve 16 Studio uses the latest multiple GPU innovations for AI and deep learning,” said Dan May, Blackmagic Design’s USA president. “With the new DaVinci Neural Engine using NVIDIA CUDA 10 and TensorFlow acceleration, our tests show a better than 2x performance increase over previous NVIDIA GPUs. These same GPUs are also used for decoding and debayering Blackmagic RAW images which makes them an attractive investment.”

These same GPUs are also used for decoding and debayering Blackmagic RAW images which makes them an attractive investment.”

This is what matters.

The rest don't matter. The rest work fine without tensor cores. He also didn't talk about faster renders...
He talked about those specific small sample of tools like Magic Mask and so on...

You keep thinking the Neural Engine works only on tensor cores for some weird reason...

This conversation has extended to reddit:

https://www.reddit.com/r/davinciresolve ... ng_on_the/

Puget replied

This conversation has extended to reddit:

https://www.reddit.com/r/davinciresolve ... ng_on_the/

Puget replied

So what's your conclusion from what he said there? As he seems to contradict himself in that post a couple of times. AMD is faster in 8K but slower in 4K? and so on... How does that make any sense?
Probably the same sense as a 3060 being faster than a 3090 in 4K and 8K in their benchmarks.

AMD is faster in 8K but slower in 4K? and so on... How does that make any sense?

My guess would be that NVidia is able to process Resolve images faster as long as the clip can be stored all in the GPU memory. 8K being bigger would likely need the image data moved from system RAM to GPU RAM. AMD's design may have an advantage there. Basically, the processing is the bottleneck on smaller images and system interconnect is the bottleneck on larger images.

AMD is faster in 8K but slower in 4K? and so on... How does that make any sense?

My guess would be that NVidia is able to process Resolve images faster as long as the clip can be stored all in the GPU memory. 8K being bigger would likely need the image data moved from system RAM to GPU RAM. AMD's design may have an advantage there. Basically, the processing is the bottleneck on smaller images and system interconnect is the bottleneck on larger images.

Maybe I'm not being clear enough. I think their tests are borderline weird.

The Puget reply was AMD is better than Nvidia in 8K, with a link to their 3090Ti review...Which has 24GB of RAM. Its not a matter of VRAM. If VRAM is out of the equation then it should be faster in 4K as well. Or at least closer, not losing in 4K by a large margin. Pixels are pixels in the end.

Also they are using Fusion on a GPU benchmark? A single core CPU dependent thing on a GPU review...Why?

They even said if you ignore Fusion, a 6900XT is 20% faster than a 3090Ti when using ProRes...yet I keep hearing how the AMD cards are garbage...And this is simply not true.

Lets also ignore AMD for a bit.
Here's another example:https://www.pugetsystems.com/pic_disp.php?id=66142
You have the 3060ti beating the crap out of a 3090 in 4K RED media and 8K RED media...

With all the above, I can't take any of these benchmarks seriously...

But that's just me and my 2 cents. People are free to follow whomever and whatever the data they want.

These same GPUs are also used for decoding and debayering Blackmagic RAW images which makes them an attractive investment.”

This is what matters.

The rest don't matter. The rest work fine without tensor cores. He also didn't talk about faster renders...
He talked about those specific small sample of tools like Magic Mask and so on...

You keep thinking the Neural Engine works only on tensor cores for some weird reason...

I would contend that many people don't use Blackmagic Raw. There are timelines at different resolution than the input videos, the viewer at different resolution that all counts for GPU calculations.

I would contend that many people don't use Blackmagic Raw. There are timelines at different resolution than the input videos, the viewer at different resolution that all counts for GPU calculations.

I wasn't referring to Blackmagic RAW in particular but RAW in general. And Resolve was always a RAW centric color grading software. Every input gets transformed into 32bit float (FP32) where all the calculations are done. And FP32 is the most important metric in Resolve...

Niche cases like Magic Mask where FP16 or FP8 (aka "tensor cores") could be used are that..niche. And even there the difference is not noticeable. 4000 series announcement has a 70% improvement in AI with a 70% increase in CUDA cores. Where's the tensor core improvement?

If FP32 is used in like 95% of what Resolve does, why would I not want more FP32 cores over some niche AI stuff? When even that niche Ai stuff can be done on regular FP32 albeit a tad slower since its higher precision.

AMD is faster in 8K but slower in 4K? and so on... How does that make any sense?

My guess would be that NVidia is able to process Resolve images faster as long as the clip can be stored all in the GPU memory. 8K being bigger would likely need the image data moved from system RAM to GPU RAM. AMD's design may have an advantage there. Basically, the processing is the bottleneck on smaller images and system interconnect is the bottleneck on larger images.

Maybe I'm not being clear enough. I think their tests are borderline weird.

The Puget reply was AMD is better than Nvidia in 8K, with a link to their 3090Ti review...Which has 24GB of RAM. Its not a matter of VRAM. If VRAM is out of the equation then it should be faster in 4K as well. Or at least closer, not losing in 4K by a large margin. Pixels are pixels in the end.

Also they are using Fusion on a GPU benchmark? A single core CPU dependent thing on a GPU review...Why?

They even said if you ignore Fusion, a 6900XT is 20% faster than a 3090Ti when using ProRes...yet I keep hearing how the AMD cards are garbage...And this is simply not true.

Lets also ignore AMD for a bit.
Here's another example:https://www.pugetsystems.com/pic_disp.php?id=66142
You have the 3060ti beating the crap out of a 3090 in 4K RED media and 8K RED media...

With all the above, I can't take any of these benchmarks seriously...

But that's just me and my 2 cents. People are free to follow whomever and whatever the data they want.

That benchmark is using DV 16.
A newer bench is here for DV 17.
https://www.pugetsystems.com/pic_disp.php?id=73914

I suspect an Nvidia driver and/or Resolve optimization issue(s).

I suspect an Nvidia driver and/or Resolve optimization issue(s).

So an 6800XT (a slower version of a 6900XT) went from 65fps to 40 for a 6900XT in Resolve 17? Is this still drivers? The better Resolve gets the worse H264/H265 asics get? The decoders are the same...

Same for 3090 and again the 3060Ti in 4K? And the difference in RAW between a 3070ti and a 3090 which is double in compute is 3fps?

8K RED - All GPUs there scored literally the same fps? (R16) - 69 and all dropped to 20ish in R17? Does this not seem odd to anyone? Mkay...

I think this benchmark is not good because it doesn't scale in any meaningful way. I mean if the difference between a 3060 and a 3090 is 5fps....then why even bother with the 3090? Everyone should dump their 3090s for 3060s or even better the 3050...

I suspect an Nvidia driver and/or Resolve optimization issue(s).

So an 6800XT (a slower version of a 6900XT) went from 65fps to 40 for a 6900XT in Resolve 17? Is this still drivers? The better Resolve gets the worse H264/H265 asics get? The decoders are the same...

Same for 3090 and again the 3060Ti in 4K? And the difference in RAW between a 3070ti and a 3090 which is double in compute is 3fps?

8K RED - All GPUs there scored literally the same fps? (R16) - 69 and all dropped to 20ish in R17? Does this not seem odd to anyone? Mkay...

You are comparing benchmarks of different generations. Both the software and hardware (in particular the CPU and likely the drivers) have changed. This can introduce different behavior (bottlenecks and optimization).

The point of the benchmark in this instance is self contained. Meaning, in the same environment how is GPU A compared to GPU B. Not GPU A from previous environment compared to GPU A or GPU B of this environment.

You are comparing benchmarks of different generations. Both the software and hardware (in particular the CPU and likely the drivers) have changed. This can introduce different behavior (bottlenecks and optimization).

The point of the benchmark in this instance is self contained. Meaning, in the same environment how is GPU A compared to GPU B. Not GPU A from previous environment compared to GPU A or GPU B of this environment.

Huh? What does this even mean?

You mean with newer CPUs that are better than the "old environment", the GPUs and especially the H264 / H265 decoders that have stayed the same get worse? As in much much worse?

Even better yet, the CPU and system is almost the same between the 2 benchmarks (TR 3970)...so...no that's not it.

We're talking about the same codecs here with a FPS counter...This shouldn't get worse with "newer generations"...The counter should go up, stay the same or in worse case scenario drop a few fps. Not 70 to 20...What kind of software optimization is that?

In the end I'm not even comparing benchmarks of "different generations". I'm pointing out discrepancies in the same self contained benchmark like you put it. 3060>3090. Same CPU, same driver, same software, same benchmark (R16 or R17)...

The rest is even worse...Resolve to Resolve versions things get worse? (of course it isn't true, things have stayed largely the same with the same codecs -> benchmark not good).

Again my opinion. I'm not here to change your mind...

You can ask Puget about what you perceive as discrepancy. I think your qualm is related to 4K/8K to h.264.
It seems likely a mistake on Pugets part regarding the test. The other test results seem as expected.

I don't like the fact that Puget has a single "score" for their bench. As Matt says in his reply on Reddit, different tasks can heavily favor different hardware, so you can't really say there is a "best" between flagships. It would be more useful to the community to break the down the data into a middle ground between one number and raw FPS data.

You can ask Puget about what you perceive as discrepancy. I think your qualm is related to 4K/8K to h.264.
It seems likely a mistake on Pugets part regarding the test. The other test results seem as expected.

There are a lot of mistakes on that chart. And my qualm isn't with some of the results either. My "qualm" is with : people are somewhat being deceived by that chart thinking a 3060 will be the same or close to a 3090 when dealing with 8K RED for example. And then coming back to this forum when their computer goes bananas with GPU memory full errors.

I would contend that many people don't use Blackmagic Raw. There are timelines at different resolution than the input videos, the viewer at different resolution that all counts for GPU calculations.

I wasn't referring to Blackmagic RAW in particular but RAW in general. And Resolve was always a RAW centric color grading software. Every input gets transformed into 32bit float (FP32) where all the calculations are done. And FP32 is the most important metric in Resolve...

Niche cases like Magic Mask where FP16 or FP8 (aka "tensor cores") could be used are that..niche. And even there the difference is not noticeable. 4000 series announcement has a 70% improvement in AI with a 70% increase in CUDA cores. Where's the tensor core improvement?

If FP32 is used in like 95% of what Resolve does, why would I not want more FP32 cores over some niche AI stuff? When even that niche Ai stuff can be done on regular FP32 albeit a tad slower since its higher precision.

FP32 is costly. Nvidia, ARM and Intel are now promoting FP8 for AI. Note the paper do not even give
much reference to FP32. Only talks about FP16.

https://wccftech.com/nvidia-intel-arm-b ... published/

https://www.intel.com/content/www/us/en ... model.html

Edit:

NVIDIA Tensor Cores offer a full range of precisions—TF32, bfloat16, FP16, FP8 and INT8—to provide unmatched versatility and performance.

.
A fast simple explanation of varius precisions and Tensor cores.

FP32 is costly. Nvidia, ARM and Intel are now promoting FP8 for AI.

And yet FP32 is what Resolve and GPU compute use...AI doesn't need high precision so yeah you could do it better on specialized tensor cores. Doesn't change the fact that more FP32 is better for you. And that you could do FP16 and FP8 on FP32shaders.

I'm not sure where you're going with this?

I don't like the fact that Puget has a single "score" for their bench. As Matt says in his reply on Reddit, different tasks can heavily favor different hardware, so you can't really say there is a "best" between flagships. It would be more useful to the community to break the down the data into a middle ground between one number and raw FPS data.

Bingo. Blanket statements like that are annoying as hell. Nvidia is better than AMD instead of Some Nvidia products are better than some AMD products.

Same with Intel, Apple and everyone else. Except Arri?

FP32 is costly. Nvidia, ARM and Intel are now promoting FP8 for AI.

And yet FP32 is what Resolve and GPU compute use...AI doesn't need high precision so yeah you could do it better on specialized tensor cores. Doesn't change the fact that more FP32 is better for you. And that you could do FP16 and FP8 on FP32shaders.

I'm not sure where you're going with this?

FP32 is available in tensor cores.

This is for Ampere the RTX3xxx 3rd generation tensor cores,

Third-Generation Tensor Cores
First introduced in the NVIDIA Volta™ architecture, NVIDIA Tensor Core technology has brought dramatic speedups to AI, bringing down training times from weeks to hours and providing massive acceleration to inference. The NVIDIA Ampere architecture builds upon these innovations by bringing new precisions—Tensor Float 32 (TF32) and floating point 64 (FP64)—to accelerate and simplify AI adoption and extend the power of Tensor Cores to HPC.

TF32 works just like FP32 while delivering speedups of up to 20X for AI without requiring any code change. Using NVIDIA Automatic Mixed Precision, researchers can gain an additional 2X performance with automatic mixed precision and FP16 by adding just a couple of lines of code. And with support for bfloat16, INT8, and INT4, Tensor Cores in NVIDIA Ampere architecture Tensor Core GPUs create an incredibly versatile accelerator for both AI training and inference. Bringing the power of Tensor Cores to HPC, A100 and A30 GPUs also enable matrix operations in full, IEEE-certified, FP64 precision.

FP32 is available in tensor cores.

Its not the same. Tensor Float 32 and FP32 from a CUDA core are not the same thing. A CUDA core does a lot of stuff and a Tensor Core does only mixed matrix math...Think of it as a dumber CUDA core...

Bingo. Blanket statements like that are annoying as hell. Nvidia is better than AMD instead of Some Nvidia products are better than some AMD products.

You've been complaining that the AI related stuff doesn't matter. A pretty big blanket statement.

The reason people look at those niche cases is because most of the other stuff runs just fine on older lower end cards. On a 1060 you start magic mask. Go to lunch. Come back. Maybe it's finished. The gain going to a 3xxx card is huge.

How exactly is 8K Red raw not a niche case? For every Red camera sold there are likely a million trying to deal with H.265.

You've been complaining that the AI related stuff doesn't matter. A pretty big blanket statement.

You're projecting here. I haven't complained about anything AI related. I've said the tensor cores are less important than the normal cores. And the AI stuff can be done on regular cores as well.
It runs absolutely fine on a 6800XT and I don't have to go to lunch.

The reason people look at those niche cases is because most of the other stuff runs just fine on older lower end cards. On a 1060 you start magic mask. Go to lunch. Come back. Maybe it's finished. The gain going to a 3xxx card is huge.

And this is exactly what I'm talking about. Even you seem to misunderstand me.
Most of the other stuff doesn't run ok on a lower end card like a 1060. Just look at the plethora of complaints in the GPU memory full thread. Be it 4K, 8K ...red raw, arri raw, braw, h264 it doesn't matter. You go to lunch with your 1060 because it has a meager 4TFLOPS (even less on the 3GB version) of compute power and its not enough.

The difference from going from a 1080Ti 11GB to a 3060 won't be huge since this was the discussion just because the 3060 has a very small amount of tensor cores.

When Resolve operates in float32 for almost everything then why is the concept of FP32 and memory size and bandwidth being the important metric when choosing a GPU seem so hard for some people?

How exactly is 8K Red raw not a niche case? For every Red camera sold there are likely a million trying to deal with H.265.

Having some H265 decoders and encoders won't make Resolve run faster. If that was the case this whole discussion would be moot. Everyone would run Resolve for every type of work on Intel iGPUs alone. You need compute...and lots of it.

In my experience I cant remember i seen the cpu go above 30%, while its eating up the gpu's 11g ram frequently. Especially in 18. This is purely based on playing raw and rendering 8k red r3d with gpu acceleration. So depending on if youre footage have gpu support or not it can differ.

Update:

If anyone was following my other thread, I was having issues getting both my RX 580s running at the same time due to power consumption. I upgraded my PSU and now they run fine.

I ran the PugetBench for DR on my machine and these were the results vs the more modern cards:

BenchMark Program - https://www.pugetsystems.com/labs/artic ... olve-1523/







------------------------------------------

The cool thing about this bench is that I know for sure that I'll see improvement by upgrading my card eventually.

With both cards running, I think I'll continue on with my system as is.

Thanks everyone for the conversation here.

Having some H265 decoders and encoders won't make Resolve run faster. If that was the case this whole discussion would be moot. Everyone would run Resolve for every type of work on Intel iGPUs alone. You need compute...and lots of it.

In my case I may get an ARC770 16G for its hardware decode and encode.

I think all these discussions are very dependent on the type of editing one does in Resolve. After being a Windows user for a very long time I bought a Studio Max mainly because at the time it was almost impossible to get a GPU upgrade for my Threadripper system with the 1080Ti. The reason was I had bought a new GH6 and intended to shoot 5.7K 60P which is h265. I shoot theatre multicam with GH5S, GH5 also AX100 so 3 or 4 camera shoots. My need was to run native files. Converting 8 hours of source not only takes time but a lot of storage. So hardware decoders in the GPU are top of the list. I use almost no effects just a little color adjustment in a few places ( all cameras shoot V-log other than AX100 of course ). Having a file from the GH6 on the timeline would grind the Threadripper to a halt. The 1080Ti could not deal with this file. Studio Max runs multicam just fine, just, a few hiccups every so often even for it with audio getting out of sync. As a comparison running just h264 files I see no performance difference between Studio Max and Threaripper 1080Ti marginal one playing Fusion credit roll about 8-12fps for the Studio Max and 5-10fps for the 1080Ti.

Of course it depends.

My point is. Decoders are nice to have (encoders who cares...). Tensor cores are nice to have.
But all of the operations inside Resolve need GPU compute and VRAM.

If all you do is decode h265, use one node (like in the benchmarks above) and then export that then that's all you need. But I don't think colorists only do that...hence why the other stuff matters more.

All in all OP's problem got solved and it was what I suspected all along. A power issue. And now that his non nvidia non tensor cards work (2x580) he'll keep on using them since they work fine...

And I might get an ARC 770 too if the raytracing performance is as good as the rumors.

If I may just interject a little advice regards timing of your purchase of a new GPU as quite a few things are about to take chunks out of the price of any existing cards and maybe make waiting a couple of months worth doing.

/snipped

Short version is that I'm kicking myself for upgrading my PC's 8 months ago... but hey, my 6700XT and 5900X combo is working quite nicely so I'm not complaining too much. Anyway, hope that's some useful food for thought.

How's that working for 4k video editing?

Asking because I'm looking to buy a new PC and I'm not sure about which graphics card to get (5900X vs 5700x is my CPU dilemma, but the price difference for the better one is not very big).

I read that Davinci works better with Nvidia (when similar raw-power cards are compared in terms of AMD VS Nvidia).

But I also read that some effects rely on VRAM heavily - and Nvidia gives ridiculously low amounts of VRAM.

For example, RTX 3080 has 10GB of VRAM, and it costs about $1000 at places where I can do any shopping from.

I'd like to be able to edit 4K 60 FPS videos (though most of my work is with 1080p).
And I'm not sure which graphics card to go for.

I can add more money if it really makes sense (bang for the buck). But I'd prefer to spend up to $700 if possible. For that kind of money, I can choose between:

RTX 3070 Ti
VS
RX 6800 XT

Which one is a better choice for Davinci video editing?
Are there any compatibility problems/hiccups with the RX 6800 XT AMD card model?

Can I get decent results with a cheaper card - and if so, which one would you recommend?

RTX 3070 Ti
VS
RX 6800 XT

What's your workflow? Is it RAW or H264/H265?

For RAW/Prores/H265: 6800XT
For H264 : 3070Ti

For H264/265 4:2:2 neither. Intel 12gen iGPU or Intel ARC GPU (if the drivers work properly) or a mac...

Same for the CPU, if its RAW go with the 12core. Else, the cheaper 8 core should suffice.

Also depending on where you live, the 6900/6950XT is around 700. And for the same price go for the better card.

Or if you can wait till November, new AMD cards should be revealed by then. Nvidia 4080 12 GB is a joke @ 900 and the 16GB is too expensive I think. Only "reasonably" priced 4000 series is the 4090.

I do 1080p YouTube videos 90% of the time.

Still learning the ropes. It's mostly compressed recordings (h264 for the 1080?).

However, I run into a 4K recording (GoPro usually - I suppose it compresses videos too).

And I fear 3070 might run into VRAM problems if I start playing with 4K more (as I expect to be the case).

Pricewise, RX 6800 (non XT version) is about 700 euros, 3070 Ti is abut 750 euros.

Edit:
Basically, my first question/dillema is: is RX 6800 stable with DaVinci Resolve?
Secondly, for editing GoPro 4K videos (would have to Google the format used, unless it lets you choose compression or raw), will one of the two cards be a lot better choice?

P.S.
6800 XT and 3080 are around 1000 euros. I'd rather not pay that much, unless the above-listed cards are both a very poor choice.

Basically, my first question/dillema is: is RX 6800 stable with DaVinci Resolve?
Secondly, for editing GoPro 4K videos (would have to Google the format used, unless it lets you choose compression or raw), will one of the two cards be a lot better choice?

It is stable. If you have a good PSU (not some non brand one) of at least 700W you're good.

If you're worried about vram issues, there are none on the AMD side. I haven't seen a GPU memory full with any material (from 4K to 8K RED and 12K BRAW).

The h264 encoder / decoder is better on the 3070ti. The one on the 6800XT is ok. I don't deal with a lot of H264 source material but it wasn't an issue when I did.

If you're in EU, the 6800XT has dropped in price to 679EUR. And the 6900XT is 769EUR.
https://www.mindfactory.de/product_info ... 20502.html

I think they deliver to any EU country.

If you do heavy grading then more VRAM and more FP32 is better. If you only edit and stuff, then even the low 3060 12GB would suffice. Or the 6700XT.

Basically, my first question/dillema is: is RX 6800 stable with DaVinci Resolve?
Secondly, for editing GoPro 4K videos (would have to Google the format used, unless it lets you choose compression or raw), will one of the two cards be a lot better choice?

It is stable. If you have a good PSU (not some non brand one) of at least 700W you're good.

If you're worried about vram issues, there are none on the AMD side. I haven't seen a GPU memory full with any material (from 4K to 8K RED and 12K BRAW).

The h264 encoder / decoder is better on the 3070ti. The one on the 6800XT is ok. I don't deal with a lot of H264 source material but it wasn't an issue when I did.

If you're in EU, the 6800XT has dropped in price to 679EUR. And the 6900XT is 769EUR.
https://www.mindfactory.de/product_info ... 20502.html

I think they deliver to any EU country.

If you do heavy grading then more VRAM and more FP32 is better. If you only edit and stuff, then even the low 3060 12GB would suffice. Or the 6700XT.

I'm in Serbia (Europe, but not EU).
And prices are outrageous (you can freely add 50+ percent on the regular US prices and you won't overshoot ).
Getting stuff from Hungary is an option, with its extra costs and caveats.
Ordering stuff online is quite tricky, especially if you expect some warranty (plus, ordering directly costs also about 40% on top of the retail price+shipping price).

It is what it is.

I've run so far on the computer from my signature, but it's frustratingly slow.
And the computer is almost 10 years old.

1080p 60 fps is impossible to edit and work with (sooo much stutter and out of sync and all kinds of hiccups). 4K - didn't even try it after that (for now, I bother a friend who does that regularly and has the tools).

As far as I can tell, any Nvidia that has at least 12 GB of VRAM (next tier is 24, no 16 GB option) costs well over a $1000. Based on my experience so far, getting a graphics card that is tight on VRAM makes it a very short-lived graphics card.

3070 Ti fits the budget nicely, but with 8 GB of RAM I expect it to start having capacity problems (despite using the super-fast DDR6x VRAM, it still lacks in capacity). 3080 is outrageously expesnive with only 10 GB, and even more so with (still quite conservative) 12 GB.

Hence, I'm leaning towards the 16 GB RX 6800. My biggest concern was the stability with DaVinci, as well as any harsh performance penalties.

So when you say that AMD it slower (for some tasks), do you mean it as unusably slow, or just takes longer but you can work normally?

For now, my editing boils down to cutting, merging, making two videos run "side-by-side" on the screen and similar basic stuff. Transition effects as well - for all I know it is considered to be very basic (could be wrong, still have a ton of stuff to learn).

P.S.
PSU I can find is Corsair RM750x (750 Watt) - the 850W version is not available.
My favourite, Seasonic, has problems with modern cards, especially Nvidia, so it's risky IMO.

P.P.S.
What's available for now:
SAPPHIRE 11305-01-20G Radeon RX 6800 16GB GDDR6 NITRO+ PCIE just under 700 euros
ASUS Radeon RX 6800 XT OC 16GB GDDR6 (TUF-RX6800XT-O16G-GAMING) just under 900 euros

So when you say that AMD it slower (for some tasks), do you mean it as unusably slow, or just takes longer but you can work normally?

It works normally here.

The render is slower than nvidia (its not slow just slower) with the AMD encoder option with H264. It has some weird limitation. On H265 it works faster for example...But I don't really use the encoder since I render everything to DNxHR / ProRes / Cineform and then make x264 (h264) via ffmpeg if necessary.

And the decoder (the part thats useful) works fine.

4K H264 60fps in a 4K timeline from that puget benchmark plays back in real time from a nvme drive (all i have to test right now). (3 tracks in split screen is also real time)

So when you say that AMD it slower (for some tasks), do you mean it as unusably slow, or just takes longer but you can work normally?

It works normally here.

The render is slower than nvidia (its not slow just slower) with the AMD encoder option with H264. It has some weird limitation. On H265 it works faster for example...But I don't really use the encoder since I render everything to DNxHR / ProRes / Cineform and then make x264 (h264) via ffmpeg if necessary.

And the decoder (the part thats useful) works fine.

4K H264 60fps in a 4K timeline from that puget benchmark plays back in real time from a nvme drive (all i have to test right now). (3 tracks in split screen is also real time)

Thank you very much for your time and sharing this information.

There's a ton of contradicting info on the Net, that I couldn't make heads or tails of.

First hand user experience you provided is what I find to be the most reliable info, and leads me to believe I won't go wrong with an RX 6800 card.

My goal is to work normally with 1080p and 4k 60 fps videos, and, hopefully, be able to see the finished (rendered) product of a 30-minute 1080p video within 10-ish minutes, instead of letting it work and coming back much later (or in the morning ).

A minor consideration would be if you do deal with h265 4:2:2 media (like from consumer grade video/hybrid cameras) then an intel processor 10th gen or up might be a worthy upgrade also since the integrated GPU does hardware decoding of this, which no AMD or NVIDIA GPU does, but you can still get the benefit of the hardware decoding by CPU with the hardware-accelerated everything else (studio version).

But that's a motherboard and whole cpu processor family change, so better bang for buck going with GPU upgrade. Agree with other comments that VRAM is where it's at.

A minor consideration would be if you do deal with h265 4:2:2 media (like from consumer grade video/hybrid cameras) then an intel processor 10th gen or up might be a worthy upgrade also since the integrated GPU does hardware decoding of this, which no AMD or NVIDIA GPU does, but you can still get the benefit of the hardware decoding by CPU with the hardware-accelerated everything else (studio version).

But that's a motherboard and whole cpu processor family change, so better bang for buck going with GPU upgrade. Agree with other comments that VRAM is where it's at.

Correction to this. Only 11th gen and forward for 4:2:2. iGPU only (there are KF variants without iGPU).
And 11th gen is not a good purchase.

12th gen (13th) or AM4/AM5 would be my choice with the AM4 5000 series being the best value.

A minor consideration would be if you do deal with h265 4:2:2 media (like from consumer grade video/hybrid cameras) then an intel processor 10th gen or up might be a worthy upgrade also since the integrated GPU does hardware decoding of this, which no AMD or NVIDIA GPU does, but you can still get the benefit of the hardware decoding by CPU with the hardware-accelerated everything else (studio version).

But that's a motherboard and whole cpu processor family change, so better bang for buck going with GPU upgrade. Agree with other comments that VRAM is where it's at.

Correction to this. Only 11th gen and forward for 4:2:2. iGPU only (there are KF variants without iGPU).
And 11th gen is not a good purchase.

12th gen (13th) or AM4/AM5 would be my choice with the AM4 5000 series being the best value.

I'm building everything ground up (even the case is not "reusable" because of the small size).
So everything is on the table.

Though it seems like Ryzen 9 5900X is a good choice overall (as far as I can figure it out).
Apart from video editing, other tasks are trivial, except maybe running a small/simple Linux instance using VirtualBox.

2x32GB RAM.
Decent B550 motherboard.
2 TB Samsung NVMe drive.
APC 1600 VA UPS.
And a 1 TB Samsung SSD drive and some 3 + 8 TB spare HDD storage I'm already using.

P.S.
Planning to skip the latest AMD and Intel CPUs (until all the bugs are sorted out), as well as the latest Nvidia graphics cards. Call it a gut feeling.

Planning to skip the latest AMD and Intel CPUs (until all the bugs are sorted out), as well as the latest Nvidia graphics cards. Call it a gut feeling.

5900X is a very capable CPU. Just make sure you get at least 3200CL16 DDR4. Ryzen is memory sensitive. The faster the memory the faster the CPU. 3200CL14/3600CL16 is the sweet spot. You'd lose performance with 2400 for example.

5900X is a very capable CPU. Just make sure you get at least 3200CL16 DDR4. Ryzen is memory sensitive. The faster the memory the faster the CPU. 3200CL14/3600CL16 is the sweet spot. You'd lose performance with 2400 for example.

I could source 3600 MHz / CL18.
CL16 is not available in 32 GB modules (only 16 GB), and using 4 sticks of RAM has some drawbacks.

Hope that the slightly higher latency will not make much of a difference.

As a follow-up here.... I purchased a second hand 3080TI for ~700

The performance difference between it and my 580s was insane... ; since my machine was crashing when trying to render with both 580s my work flow was to only render with one of the drives. (Issue outlined here - viewtopic.php?f=21&t=159841 )

So with one 4K supersized timeline I was rendering at 1 hour and 13 min on the RX 580....

The 3080TI rendered the exact same video in 10min. Truly awesome....

In addition the card didn't get early as hot and it was quite!