CPU Cores verses Clock Speed and Mobo's

I am upgrading my box (replacing Mobo, GPU and memory) and contemplating how Resolve uses system resources.

As I have not found many current specifications how Resolve 12.5 (PC) leverages CPU cores and Threads, in addition to system memory. I have heard that Resolve does not access more than 16GB of Ram however, is that accurate and if so is that specific to scrubbing multi-layered projects.

As my initial thought with respect to Mobo's would be to look at a Skylake, the X99 will accommodate six, eight and 10-core CPU's. While a 8-Core i7-6900k may not be worth an extra $400 plus verses the i7-6850 for an additional 2-cores and 4-threads, it is even worse when you go to the 10-core i7-6950X which lists for over $1,600

Going the other way, perhaps it might be worth losing the ability to overclock in favor of a Xeon e5-2640 v4, which while it sells for $900. One could add a second proc down the road for a build of 20-core/40-threads, this for about the same cost as the 6950X

So returning to the beginning of this thread, how many cores verses clock speed will Resolve leverage? Further, perhaps a good X99 with twin NVIDIA GTX-1080's would be a better spend as part of the overall system.

gbh

I have a 20 core Xeon in an X99 board, you can pick up v4, 12 or 14 core chips for under $500.

Yes, very reasonable and lot's of options - just not sure what Resolve uses as optimal - what they say is not really specific to these questions.

This is what is posted about GPUs and alludes to CPU's however, I'd like them to be more specific how and why CPU Cores do not enter into the mix...

Your GPU is everything! It is more important than your CPU or system RAM (both of which should not be skimped on either).

This should be a dedicated GPU just for image processing in addition to the graphics card running your desktop GUI (user interface). In the case that you are using a laptop or any system with a single, or integrated GPU you can still run Resolve, but performance will be compromised compared to a dual or multi-GPU system.

GPU RAM:

512MB – Forget about it.
1GB – You’ll be okay with basic HD ProRes work, checking RAW files but avoid noise reduction and optical flow.
1.5GB – Approaching the absolute minimum to use Resolve with some level of complexity in HD. Noise reduction and optical flow will still be problematic. I’ve made a 1.5GB Intel Iris Pro GPU work well on a iMac and managed to render a full 4K delivery (no noise reduction or optical flow speed changes):

[Peter Chamberlain]

Resolve does all its image processing in GPU, thus the reason we recommend the best GPU you can afford and certainly with high GPU RAM if you are working on anything higher than HD.

The CPUs are used for decoding and encoding compressed files and not image processing. In this case, and for the general operation of Resolve, we are multithreaded so having more CPU cores is generally preferable. I know of many customers using dual 18 core systems and Resolve uses all the real and virtual cores well.

If you only use Resolve on your system, no other apps, then 16GB RAM is generally sufficient, remember we are processing in the GPU, and so putting in 64GB or 128 is only going to help with other apps.

I find some of this reply curious but maybe I'm set up a little wrong. I've used resolve for years, built lots of computers.

I'm on win 10, 16gb ram I7 and just popped in a gtx1080 graphic card which supports h265 natively.
There were no "issues" with strange behavior with the card.

I just tested encoding h265 4k into h264 4k. I timed it with a gtx 770 card first. There was 0% speed increase with the 1080 card in both resolve and adobe media encoder.

I've changed preferences to use gpu, no diff. In watching the card it seems to be cruising, never using more than 50% of vram and never over 45% gpu usage. Both of these redlined with the 770 card.

Are you saying that for this kind of encoding it's all cpu? but for editing / coloring it's gpu?

I did notice that playbacks in coloring were smooth, about half play speed, but not so much hacking as before. Again the card was not stressed in the least. I have to wonder if resolve is using the gpu power available to it. It never seems to use more that 50% of gpu vram, no where near gpu 100% usage.

the 1080 might support h264 natively, but it is safe to assume that the software must be programed to call the 1080 instead of the cpu.

it is safe also to assume that resume still use the cpu code: in that way you don't get the speed benefit

Resolve does all its image processing in GPU, thus the reason we recommend the best GPU you can afford and certainly with high GPU RAM if you are working on anything higher than HD.

that's the bm response. So is transcoding not usng the gpu?

If resolve only uses the gpu for image processing - then stressful image processing tasks should use and stress the gpu, until the playback is full speed, not use half of it.

[Peter Chamberlain]

Transcoding is CPU bound, decoding then encoding, assuming your disk subsystem can provide and recieve frames faster than the CPU can process.

Peter,

To your post

Resolve does all its image processing in GPU, thus the reason we recommend the best GPU you can afford and certainly with high GPU RAM if you are working on anything higher than HD.

Understood, so a dual system with more GPU memory will provide faster cycle-time for processing however, is GPU clock speed part of the equation or just RAM?

The CPUs are used for decoding and encoding compressed files and not image processing. In this case, and for the general operation of Resolve, we are multithreaded so having more CPU cores is generally preferable. I know of many customers using dual 18 core systems and Resolve uses all the real and virtual cores well.

Decoding, at least from my perspective is more important as rendering can be done while the system "cooks" and I am doing other tasks, decoding/opening and dropping content into bins is a time factor however, how much of an improvement is seen if measurable?

If you only use Resolve on your system, no other apps, then 16GB RAM is generally sufficient, remember we are processing in the GPU, and so putting in 64GB or 128 is only going to help with other apps

Makes sense and memory is inexpensive relative to other costs

Thanks,
gbh

Coming late to this thread... find it interesting. I was considering building a new AMD Ryzen system with 64GB RAM and 1080 ti. I am rather curious why 16GB is a limit for editing? I would have thought the more RAM you have, the more Resolve can load in to RAM for fasting CPU use.. and to avoid any disk caching. I gotta believe larger HD and even small 4K projects must use well over 16GB of assets. Why if you have 32GB or 64GB of RAM would Resolve not make use of that to keep things off of the disk?

Now, I use all SSD based systems for editing these days.. e.g. smaller SSD boot disk, an assets SSD disk, and set up the scratch/editing on a separate SSD, with renders going to a plugged in HD or SSD depending on size of what I am rendering. But even so, I actually added RAM specifically because I assumed all this time that video editing would make full use of more RAM. Now that Resolve 12.5.5 is out, I am hoping Resolve 13 would look to support more RAM if it truly does not do so now.

Also, I want to throw out there.. to me with the low cost of SSDs and even multi TB HDs with hybrid SSDs, it seems like such a waste to try to edit with AVCHD/MP4 file formats. I use a fantastic little app called Convert (v4 now), that I first convert everything to DNxHD. It does not improve the quality, and the files are much larger, but editing is super smooth. I only work on small projects for my own use, never anything like a tv show or movie, so for me my 3 SSD system works well enough for one project at a time.

I am really looking forward to seeing what the Optane technology does for all this when it arrives later this year. Probably one reason I am leaning towards Intel chipset as I dont think it will work with AMD setups.

I am rather curious why 16GB is a limit for editing? I would have thought the more RAM you have, the more Resolve can load in to RAM for fasting CPU use.. and to avoid any disk caching. I gotta believe larger HD and even small 4K projects must use well over 16GB of assets. Why if you have 32GB or 64GB of RAM would Resolve not make use of that to keep things off of the disk?

I might not understand you correctly, but I get a feeling you think that a computer loads any files relevant to an application into RAM as long as there is space left? Like if I was editing a 4k movie, my media files would get loaded into RAM?

That's not how RAM works... It would be very impractical if you could saturate all of your RAM with a few media files, grinding your computer to a halt.

I am rather curious why 16GB is a limit for editing? I would have thought the more RAM you have, the more Resolve can load in to RAM for fasting CPU use.. and to avoid any disk caching. I gotta believe larger HD and even small 4K projects must use well over 16GB of assets. Why if you have 32GB or 64GB of RAM would Resolve not make use of that to keep things off of the disk?

I might not understand you correctly, but I get a feeling you think that a computer loads any files relevant to an application into RAM as long as there is space left? Like if I was editing a 4k movie, my media files would get loaded into RAM?

That's not how RAM works... It would be very impractical if you could saturate all of your RAM with a few media files, grinding your computer to a halt.

It was explained by Peter from BM earlier in the thread. Resolve is limited by the amount of RAM you have on your GPU card as all image processing is done on the GPU. So if you are lucky to have a GPU with 16GB of memory then that is your limitation, not the amount of RAM that your CPU uses.

Jules

So Ram is basically just for reading and buffering current playback footage with some look ahead I am guessing. Hence the importance of Fast drives to store your footage on.

If it is properly multithreaded you can assume threads * clock speed as a metric for the CPU.

So I guess the process for dtermining realtime playback is pretty much:

1. Have drives fast enough to read your footage in. (Test using BM Disk speed Util)
2. Have a CPU fast enough to decode the footage codec. (Test playback with no correction nodes, assume above test passed)
3. Have a GPU fast enough to process your corrections (nodes). (Test playback with nodes active, assume above tests passed)

And if rendering an extra step of:

4. CPU speed, output codec and drive write speed will determine the speed of encode.

Going the other way, perhaps it might be worth losing the ability to overclock in favor of a Xeon e5-2640 v4, which while it sells for $900. One could add a second proc down the road for a build of 20-core/40-threads, this for about the same cost as the 6950X

Overlooked 6950x will most likely outperform dual e5-2640 v4 system.
Simpel math:
-6950x- 10cores at eg 4.2GHz=42GHz of processing power
-dual e5-2640 v4=20x2.4GHz=48GHz of processing power, but such a high cores count means almost not a single operation in Resolve will utilise all cores at 100%, so this is not a representative number in real world

If you would used these CPUs with some 3D rendering app, then Xeons would probably win as these type of programmes can really use 100% of each core.

When you take for example ProRes decoding you can see that even this doesn't really scale linear way with number of cores (http://images.apple.com/final-cut-pro/d ... _Paper.pdf page 17). 12 cores give only real decoding speed of 7 cores (relatively to decoding speed of 1 core: 11.2/1.6=7). Encoding will be even worse and ProRes is quite simple codec and very well optimised by Apple. In case of h264 this is going to be way worse as codec is complex and it's hard to write decoder which can take 32 cores and use at 100%.

Dual e5-2640 v4 overall processing power in Resolve will be limited probably to something like 70% (this is probably way overestimated). So your 48GHz of power will be reduced to e.g. 33GHz. 10 cores of 6950x will be used in better % (because there are less cores), so it will rather outperform dual e5-2640 v4 in Resolve.

The main issue to solve is number of possible GPUs (PCI-E lanes) in case of single CPU like 6950x. You would have to check if your system will run out of GPU processing power or CPU. If e.g. 2x GPUs are running out of juice fairly quick (and you can't add 3rd one) then this is an issue and you need dual CPU system (but rather one with e.g. 16 cores at 3.5GHz than 20 cores at 2.4GHz).

You have wrong assumption- processing power scales with clock speed in a linear way (and this is the case for about every operation CPU can perform), cores don't scale linear for sure and for some operations 20 cores will give you only e.g. 3x boost compared to single core, not 20x!
You should always favour higher clock speed over number of cores, so 16 cores at 3.5GHz over 20 cores at 2.8GHz.

thoughts;

- i make do with dual 8 core xeon's, so 32 threads, currently i'm rendering source caches, alexa mini ProRez @ 4k -> DNxHR444_HDR in a UHD timeline, i'm seeing all 32 threads running about the same area, around 35-40%

- i have a SAS array, that's plenty fast enough not to have the processes disk bound

- one reel of film in this timeline, and my system is using 19 gig of ram, but i often see it much higher, and when using the media managemnt to delete unused caches on a feature i've seen it go as high as 70gig, my personal confort zone is 96 gig absolute minimum

edit to mention that the sys above was disk bound, incomeing from the nas, i see all 32 threads at 100% rendering when not disk bound

You have wrong assumption- processing power scales with clock speed in a linear way

It does, using 3D rendering as example (as you did) it pretty much scales linearly, twice the cores, twice the threads, half the time. And these algorithms usually run close to 100% usage.

With video the implementation of the codecs determines the efficiency which is more often the problem. But if other resources (GPU, Ram, Drives) can cope there is no need for the processing not to scale linearly, the threads can effectively (and often do) decode/encode there own frame or segment of video independantly. But you have to share the GPU, RAM and drives between each thread which is where the bottle neck occurs (and why Dermot is probably seeing 35-40% usage). So for example my machine has 36 threads which would mean each thread gets 1/36th of the GPU, Ram and Drive resources. This all becomes a balancing act as the resources aren't necessarily required at the same time.

That said my 36 threads is significantly faster that my previous high end i7 with resolve and I have no problem working with native 4K footage on feature length projects at full res with 4K monitoring.

But if I was building a bang for buck mid budget machine would look at the either of the 2 high end Ryzen chips, some largish SSD's and a 1080Ti (when the prices come down in a month or so).

You have wrong assumption- processing power scales with clock speed in a linear way

It does, using 3D rendering as example (as you did) it pretty much scales linearly, twice the cores, twice the threads, half the time. And these algorithms usually run close to 100% usage.

Processing power will scale about always linear with clock speed, but not with cores count and this is mainly related to fact how easily you can parallelise your computing tasks. 3D rendering is fairly easy to parallelise tasks, so it scales about linear even with number of cores.
As you said at some point you may hit other limitations like RAM speed, etc. Point is: more cores= more complications and using them efficiently is not so easy. Now we are not anymore at e.g. 4 core or even 8, but hitting big numbers like 44 cores. Not many programmes can use properly machines with these crazy core numbers, so choose wisely prioritising clock speed over cores number. Higher clock will always make you machine faster, more cores not necessarily.

It was explained by Peter from BM earlier in the thread.

Hi, Jules. I missed your previous reply.

Not that it greatly matters, but I still don't think we are on the same page. Yes, Resolve relies heavily on the GPU for its image processing. But to increase performance, you typically want to add more computing power (i.e. another GPU). Just adding more GPU RAM would not increase performance in itself (but as always: too little GPU RAM will hamper performance, but that's not the same thing). Once you have enough GPU RAM for your resolution or node types (temporal noise reduction?), adding more won't boost performance.

Anyway, what Peter said was Resolve by itself is happy with 16GB system RAM (not talking about GPU RAM here). I'd take his word for it.

Simpel math:...

Perhaps, but Resolve is not the only process running on the machine. My Linux system has literally hundreds of processes running. A fast single processor system would have to constantly flip between them. Each time a process is switched, the current context must saved, and this wastes clock cycles. If the number of processes is sufficiently large, the wasted clock cycles can promote thrashing.

So to the extent that multiple CPU's reduce context switching, the math is not quite linear. Resolve grabs two of my system's cores and holds onto them (running at 100%) for the DaVinciPanelDaemon alone. There are eight DPDecoder processes as well. All told, Resolve alone spawns more threads than my dual Xeon system has cores.

I imagine there is a performance crossover point at which the more linear clock speed wins out, but without actual testing, I'm not sure it is clear when that occurs.

Cheers.

These processes shouldn't worry you at all as these are very low level threading processes and there are 100s of them regardless if you have 1 or 20 CPU cores. These are very different threads.
If you think 20 cores at 2GHz are better in real world usage than eg. 12 at 3.5GHz than this is your choice

Back late to this..

So.. if you could answer this:

1) Timeline editing.. with nodes, FX, lots of clips on the timeline, etc..
a) GPU, CPU or both help when scrubbing through it back/forth... e.g. smoothly with no hiccups.
b) FX on timeline.. GPU bound (e.g. realtime playback and/or scrubbing.. use GPU) or CPU?
c) System RAM helps timelines or not? E.g. if I have 200+ 4K clips in my bins, and have them all out on the timeline, does more System RAM help.. or is 16GB still all you need, and GPU RAM is the winner here?
d) Multiple GPUs.. if I understand correctly, do nothing for timeline editing?

2) Rendering:
a) System RAM.. helps (if cpu bound) or 16GB is all that is needed?
b) If GPU based, GPU RAM helps/doesnt matter when rendering?
c) Multiple GPUs.. used or only one for rendering (if GPU bound).

I am trying to understand if towards the end of year it makes sense to build a loaded ThreadRipper (or possibly dual cpu EPYC) system with 64/128GB RAM and dual GPUs or not purely for use with Resolve. Would it be better to spend more on bigger NVMe drives (and optane if available by then) and less on system RAM? More on GPUs with more RAM.. e.g. one GPU with 16B RAM (possibly workstation class if that is better) or dual with 8GB RAM each?

Thank you.

I've found when building systems there is usually a sweet spot for just about every component. And there's also a reality check involved. For example, Resolve may only use 16 gig of RAM, but does that mean you actually only want let's say 24 gig of RAM in a brand new system? No, because you may be running other programs or Resolve may increase its RAM usage in the not so near future.

So asking "What should I prioritize when making a build?" is a good question, but pretty soon the answers become pretty routine, get the best value for the money you can afford for each component.

The only question that I think is still an open one is "Is it better to have multiple CPUs with matching video cards than having one faster CPU with more cores?" I imagine more CPUs is significantly better than one faster CPU with more cores.