Improving Resolve Performance on Windows

You'll need a very well balanced system.
Just adding GPUs doesn't cut it if the CPU or data transfer can't follow. I'm sure that one additional GPU is the sweet spot for my machine, I wouldn't expect three times the speed by adding just another eGPU. But if you assemble a box with lots of CPU cores, very fast memory and busses, the point of diminishing return might come when adding the forth.

Okay, the parts arrived yesterday and I have put the system together.

New parts:
Gigabyte X399 AORUS Gaming 7 motherboard
AMD 1950X CPU
Corsair 1000W PSU

Parts I already had:
2 x 16GB 2400 RAM in optimal slots (DDR4_2_2A & DDR4_2_2B)
2 x GTX1070 graphics cards in optimal slots (PCIEX16_1 & PCIEX16_2)
BMD Intensity Pro 4k (in PCIEX4)
Samsung EVO 960 250GB M.2 NVMe SSD in M2P_32G (C: Windows 10 Home)
Samsung MZVPV256 M.2 NVMe SSD in M2Q_32G (D: used for source and output for my test project)
3 x 3TB 7200RPM hard drives (not used in my test project)
2TB 7200 RPM hard drive (not used in my test project)

Test project: 1 clip 2160p25 H.264 70Mb/s 8 bit 4:2:0, 1 minute long with stereo audio.

Render output format: DNxHR HQX 10 bit 1920x1080 25p

Test render 1: Temporal and spacial NR plus a minor colour grade.
Render time: 2:21 (about 11 FPS)

Test render 2: No processing
Render time: 0:11 (136 FPS)

I guess this will do me, but honestly, I was expecting a bit better. On the render with a bit of processing nothing is even close to maxed out. Curiously, the simple render was very fast and did make good use of the hardware.

Maybe I still don't have something configured optimally.

For sure, there is a bottleneck somewhere.
It's not easy to find but with your material you have to get better

I did the same test :
Same TNR and NR setting as Dan and even CPU.
10bit 4: 2: 2 150Mbit UHD h.264 gh5 footage rendered to DNxHR HQX 10 bit UHD.


Your processor does not come into use.

Mine is + 98%, yours stagnates at 15%.

My render is made between 15 and 16 fps.
I tested with a clip on a HDD I specify.

Your Gpus are in the good slots? Bios setting ?, OS in max performance?

It is necessary to monitor CPU, GPU, RAM, HDD with stress tests.

No overheating? Not easy to diagnose.

CPU_GPU_WIN10.jpg

In your case it's source file type which makes your CPU so busy. 4:2:2 h264 will be CPU decoded.
What is your GPU?

In your case it's source file type which makes your CPU so busy. 4:2:2 h264 will be CPU decoded.
What is your GPU?

Andrew: At the bottom of the picture.

Had to scroll
So your result is just slightly better.
I assume this is about all what you can expect from 2x GTX1070.

Is there any difference between Windows 10 home & pro when it comes to multiple GPUs?

Is there any difference between Windows 10 home & pro when it comes to multiple GPUs?

Maybe?
3 little tricks:

On Windows 10 PRO (I do not know if the option exists on HOME? To check) there is an option to privilege GPUs:
Start => Parameters => Display => Graphic Settings =>
- Click on Browse and local RESOLVE.EXE
- Then Options
- select High Performance
- Record

and test.
For me 1 more fps. At the same time more use GPUs and less CPUs.

CPU_GPU_OS_PERFORMANCE.jpg

(sorry : is french)

GPU PREFERENCES.jpg

It is necessary to test if it is better or not according to its configuration.

Other parameter to test in Davinci Resolve => Preferences => Configuration:
(maybe reserved for the Studio version)
Enable the "User optimized image transfers for viewers" => test the result
Enable the "Decode H.264 / HEVC using hardware acceleration" => test the result
test the result by activating the 2 or one.

With these 3 tricks, I can read in COLOR tab in RT but not all the time ... sometimes it comes down to 22 or 23 fps ...

Results_tricks.jpg

Okay, the parts arrived yesterday and I have put the system together.

New parts:
Gigabyte X399 AORUS Gaming 7 motherboard
AMD 1950X CPU
Corsair 1000W PSU

Parts I already had:
2 x 16GB 2400 RAM in optimal slots (DDR4_2_2A & DDR4_2_2B)
2 x GTX1070 graphics cards in optimal slots (PCIEX16_1 & PCIEX16_2)
BMD Intensity Pro 4k (in PCIEX4)
Samsung EVO 960 250GB M.2 NVMe SSD in M2P_32G (C: Windows 10 Home)
Samsung MZVPV256 M.2 NVMe SSD in M2Q_32G (D: used for source and output for my test project)
3 x 3TB 7200RPM hard drives (not used in my test project)
2TB 7200 RPM hard drive (not used in my test project)

With only two RAM sticks, you're halving your memory bandwidth. TR can do Quad Channel.

Also Slot DDR4_2_2B doesn't exist, you meant DDR4_4_2B I assume? Try putting that in DDR4_6_2C. From the QVL Memory PDF-

"2 DIMM: Supports one pair of modules inserted into the same color slots as one pair of Dual-channel memory configuration. Install the modules into DDR4_2_2A, DDR4_6_2C for better compatibility."

Run CPU-Z and have a look at the Memory tab, see your RAM speed setting, and if it's running Dual Channel properly.

RAM speed also have a great effect on TR processing as the CPU interconnects run at the same speed. Well, half. Faster RAM, 3200-3400 odd helps.

Note: if your material is h264, you're cpu (for decoding) impact speed more than the gpu (for computing)

Your Gpus are in the good slots?

Yes, they are in the 2 16x slots

Bios setting ?

I don't see anything in there that stands out to me as needing to be changed

Note: if your material is h264, you're cpu (for decoding) impact speed more than the gpu (for computing)

I've proven the decoding, encoding and hard disk transfer rate are not the bottleneck since the exact same timeline with the effects switched off will render at 136 FPS.

I just did some more tests:

The same UHD25p 8 bit 70Mb/s 4:2:0 H.264 source file at 1 minute long.

SNR and TNR settings are the same as Dan Sherman's (see his image below)

Render times:
No effects: 0:11 (136fps)
Minor colour grade only: 0:11 (136fps)
SNR Only: 1:41 (15fps)
TNR Only: 2:07 (12fps)
All the above effects: 6:31 (4fps)

With only two RAM sticks, you're halving your memory bandwidth. TR can do Quad Channel.

If that's truly the problem and will make a big difference I'll add 2 more RAM modules. My gut feeling is it's not. I'll ask my supplier next week if I can buy 2 more modules and return them if they don't help.

Also Slot DDR4_2_2B doesn't exist, you meant DDR4_4_2B I assume?

Yes, that was a typo.

Try putting that in DDR4_6_2C. From the QVL Memory PDF-

I moved one module from DDR4_4_2B to DDR4_6_2C and the results were pretty much the same (very slightly worse in most cases)

Render times:
No effects: 0:19 (79fps)
Minor colour grade only: 0:18 (83fps)
SNR Only: 1:41 (14fps)
TNR Only: 2:07 (11fps)
All the above effects: 6:28 (4fps)

Render times:
No effects: 0:19 (79fps)
Minor colour grade only: 0:18 (83fps)
SNR Only: 1:41 (14fps)
TNR Only: 2:07 (11fps)
All the above effects: 6:28 (4fps)

Don´t worry Mike, with such material and such settings you will always get maxed load of GPU (s). On my setup with the same SNR + TNR on 2X1080TI I have 10-11fps full load on both cards and only 15% of CPU. Happy stories go from people with cache turned on or (and) timeline 2K

Render times:
No effects: 0:19 (79fps)
Minor colour grade only: 0:18 (83fps)
SNR Only: 1:41 (14fps)
TNR Only: 2:07 (11fps)
All the above effects: 6:28 (4fps)

Hello,

For me with this kind of material, he should get better.
Here with a clip GH5 10 bits UHD 422:
to deliver a clip of a minute + TNR + SNR in UHD it is necessary 01:35

There must be something that prevents you from going faster.

DELIVERY_TNR_SNR.jpg

I moved one module from DDR4_4_2B to DDR4_6_2C and the results were pretty much the same (very slightly worse in most cases)

Render times:
No effects: 0:19 (79fps)
Minor colour grade only: 0:18 (83fps)
SNR Only: 1:41 (14fps)
TNR Only: 2:07 (11fps)
All the above effects: 6:28 (4fps)

If you're keen, create a separate project with one video file and grade that you think is running badly and upload it. I'm happy to load it and see how it runs, others may, and at least give you a baseline to aim for.

Willing to try too.

I just created a shorter 20 second test ready to upload and I'm getting much better performance.

I've run out of time to go any further today so I'll have to get back to it in about 20 hours.

I'm really having trouble understanding what's going on at the moment.

In order to simplify things I created a fresh project with several timelines with different processing. I also wanted to use a shorter file to make it easier to upload/download.

Surprisingly, all the timelines in this new project perform significantly better.

Thinking this may be source file related, I imported the video file into the first project and applied the same processing. I'm getting similar poor results, so it's not the source file.

Maybe there is something wrong with the project settings, although I can't see anything different between the good and bad project.

So I've now simplified both projects to just 2 timelines; one with no processing, and the other with TNR and SNR applied.

The new source file is 20 seconds long and is DNxHR UHD HQX
See the attached screenshots:

Test 1 (problem project)
TNR+SNR applied. Render time: 2:13 (6.8 fps)
No Processing. Render time: 0:14 (35.7 fps)

Test 2 (good project)
TNR+SNR applied. Render time: 0:10 (50 fps)
No Processing. Render time: 0:04 (125 fps)

The 2 projects along with the test file are available here:
https://drive.google.com/open?id=1sr57J ... UivIsSOW0H

I'd love to know what the cause is because I'm seeing similar poor performance from a real project. This project is several hours long so the render times are quite painful.

[Rohit Gupta]

In the screenshots you are rendering one of the projects at HD and the other at UHD. Is that the difference?

Yep, but If there are no mistakes better numbers are for UHD export, which of course makes no sense

In the screenshots you are rendering one of the projects at HD and the other at UHD. Is that the difference?

Ah yes, I forgot to change that. I just tried a render at UHD and it still took 2:13.

I should mention that it's not just rendering. Playback performance is much better on the good project too.

I've just discovered the reason for the difference between the 2 projects.

I don't know how I didn't notice it, but on the good project the timeline resolution is 1080p, whereas it's 2160p in the problem project.

Okay, so that explains the difference, but doesn't explain why I'm getting such poor performance considering the hardware.

I just loaded the projects on my laptop running Resolve Studio 15b6 and experience similar performance difference between the 2 projects, as expected.

This computer has an i7 7700HQ CPU, a single GTX1070 GPU and 24GB or RAM.

Test 1 (2160p timeline resolution)
TNR+SNR applied. Render time: 2:39
No Processing. Render time: 0:26

Test 2 (1080p timeline resolution)
TNR+SNR applied. Render time: 0:56
No Processing. Render time: 0:24

The interesting result here is the one I've highlighted above. The difference between that and my Threadripper computer is not as great as the other results. So there is definitely a bottleneck somewhere.

Threadripper computer: No processing: 0:14 - Processing: 2:13 (9.5 x difference)
i7 Laptop computer: No processing" 0:26 - Processing: 2:39 (6.1 x difference)

Considering the extra hardware, I'd expect the Threadripper system to handle the extra processing much better than the i7 machine.

I'm wondering if having 2 x 16GB RAM sticks is the bottleneck.

RAM so so expensive at the moment that my local supplier won't keep anything more than the basic 8GB modules they use every day in stock. So this means I can't just try doubling the memory and return it if it makes no difference. I've spent so much upgrading this system for very little gain that I'm reluctant to throw another $500 at it on a whim.

My supplier will loan me 4 x 8GB modules, so I'll try that tonight, both with and without my existing 16GB modules installed.

That way I can test the difference between 64GB in dual channel mode and 32GB in quad channel mode.

[Rohit Gupta]

As per our internal benchmarks, UHD "Temporal NR - Better" runs at these speeds:

GTX 1080 - 11.7 fps
GTX 1080Ti - 17.43 fps
Titan V - 23.05 fps

So essentially, you need a faster GPU if you need faster Temporal NR performance at UHD.

As per our internal benchmarks, UHD "Temporal NR - Better" runs at these speeds:

GTX 1080 - 11.7 fps
GTX 1080Ti - 17.43 fps
Titan V - 23.05 fps

So essentially, you need a faster GPU if you need faster Temporal NR performance at UHD.

Okay, so you think what I'm getting is quite acceptable for my hardware?

The thing that got me is the very low GPU usage in Task Manager suggesting that the GPUs are not running as hard as they could.

Had I known that a single 1080 ti is better than two 1070s that's what I would have gone for in the first place and saved a few hundred dollars. I got lead astray by the support person in Australia who told me that Resolve loves multiple GPUs more than any other single factor. I see now that that statement could have been qualified.

But maybe for other Resolve tasks two 1070s will be better.

I'd really rather not waste any more time on this as long as I've got the best I can short of throwing out $1500 worth of brand new video cards.

At least the other processing seems to fly along on this machine.

Well, I took out the two 16GB sticks and fitted four 8GB ones.

No change to the render time.

I then added the two 16GB sticks back in to make the total 64GB.

Still no change to the render time.

So I put the memory back the way it was and will return the 8GB sticks tomorrow.

Looks to me like what I have is the best I can do without replacing the video cards.

Use Faster mode in NR settings.

I get 24 fps in an UHD timeline with temporal NR (1 frame radius, faster) and around 17 fps with "better".
That's on a 2017 iMac (non Pro) with one internal Radeon 580 and one in the eGPU. I don't think it's about RAM, even if I have 32 for Fusion, it's all about GPU.

Use Faster mode in NR settings.

It goes from 2:39 to 1:11 for a 20 second clip.

I get 24 fps in an UHD timeline with temporal NR (1 frame radius, faster) and around 17 fps with "better".
That's on a 2017 iMac (non Pro) with one internal Radeon 580 and one in the eGPU.

That's pretty good going.

I don't think it's about RAM, even if I have 32 for Fusion, it's all about GPU.

My experiment earlier seems to confirm that 32GB is enough RAM for this stuff.

Since nothing in those graphs is maxed out I don't know where the bottleneck would be.

Hi.

If you want to search for the bottleneck will I suggest you to make a RAM Disc and use it for testing?
Then will you be able to see if your disc's is causing the bottleneck.

Or if you tell us more about your system, can some member of this forum may be tell you where to look.
What motherboard is you using?
What two slot are you using for the GPU's
How many and what type of your hard drive and SSD?
etc.

Regards Carsten.

+1
I like the way you're thinking through this, Carsten.

ResolveStudio 14.3...014 | MiniMonitor | DTV 10.9.7 | Win8.1 x64 | ROG G751JL, 2.8GHz Intel i7HQ, 24GB DDR4, 1TB HD, 500GB EVO 850, 2GB GTX965M | Mackie MCU Pro | Softube Console 1 Mkii | Shuttle Pro 2

Test 1 (problem project)
TNR+SNR applied. Render time: 2:13 (6.8 fps)
No Processing. Render time: 0:14 (35.7 fps)

2:13 is not 6.8 fps, only 20s/133s*25fps = 3.8 fps

My result for your project:
TNR+SNR applied. Render time:1:05 (7.7 fps)
No Processing. Render time: 0:09 (55.6 fps)

benchmark.jpg (94.27 KiB) Viewed 12664 times

ASUS X299-A PRIME, i9-7900X
2XMSI GTX 1080 Ti SLI

Just a little note regarding your system info: SLI should be off for Resolve.

Just a little note regarding your system info: SLI should be off for Resolve.

I always thought that this is nonsense. But yes, I really got 2fps gain in that test turning off SLI. But 2/65= 3% is too small to be justifiable.

According to BM it's not only about speed, but stability too.

My result for your project:
TNR+SNR applied. Render time:1:05 (7.7 fps)
No Processing. Render time: 0:09 (55.6 fps)

ASUS X299-A PRIME, i9-7900X
2XMSI GTX 1080 Ti SLI

Thanks for that.

Just a little note regarding your system info: SLI should be off for Resolve.

SLI is off.

It wasn't for the first post in this thread because I didn't realise it was on by default.

Stupid questions: are the two cards in PCI slots that share the same lines? IN other words, in some motherboards you need a very specific order otherwise you will make speed conflicts...

(I remember reading other posts in merit.)

In the manual for your motherboard I see this sentence: "The PCIe 3.0/2.0 x16_2 slot is not recommended for VGA cards or PCIe cards running at x8 mode or above." To me, that does not bode well for your system, given that your second GPU is certainly running at x8 more or above. This admonition is present only for the PRIME X299-A, PRIME X299-DELUXE, and WS X299 PRO motherboards, and no others, which I can say now that I've looked in the manual of every single X299 motherboard they sell

Stupid questions: are the two cards in PCI slots that share the same lines?

I don't think so. These slots are specifically for video cards according to both the web site and the manual.

In the manual for your motherboard I see this sentence: "The PCIe 3.0/2.0 x16_2 slot is not recommended for VGA cards or PCIe cards running at x8 mode or above."

My motherboard is an X399. I can't find that line in the manual, but it does talk about fitting 2 16x cards in particular slots, which is what I did. See the attached screenshot from the Gigabyte web site.

Based on what I've found out here, what I'm getting is expected. For better NR performance I'm better off with a single 1080 ti.

So to that end, I've put the 2 1070s up for sale and I'll order a 1080 ti next week.

I've put the 2 1070s up for sale and I'll order a 1080 ti next week.

Hi.

I will recommend you to wait a few weeks to order a GTX 1080 Ti. There have been rumors on the Next Generation nVidea New Graphics cards for to long. But The world’s biggest gaming expo, Gamescom 2018, runs August 21-25 in Cologne, Germany:

https://www.nvidia.com/en-us/geforce/ne ... scom-2018/

Many is hoping/expecting to find out more about when the Next Generation nVidea Graphics cards will be available. They are expected to deliver more performance, at only a little higher price.

Regards Carsten.

In the manual for your motherboard I see this sentence: "The PCIe 3.0/2.0 x16_2 slot is not recommended for VGA cards or PCIe cards running at x8 mode or above." To me, that does not bode well for your system, given that your second GPU is certainly running at x8 more or above. This admonition is present only for the PRIME X299-A, PRIME X299-DELUXE, and WS X299 PRO motherboards, and no others, which I can say now that I've looked in the manual of every single X299 motherboard they sell

Mine is PRIME X299-A. I think this phrase means just not to put the single GPU into it. Because in very hard benchmark TimeSpy 4K the configuration 1 slot + 2 slot gives double increase in performance. Therefore both slots work at full speed.

@Mike,

You are only running 2 sticks of 2400 Mhz ram correct? Also what is your rams latency?

Threadripper supports quad channel, so only running dual channel is a limitation. Also the infinity fabric that allows the individual dies in your cpu to talk to each other runs at the same frequency as your ram. This is why people recommend the highest frequency ram that you can get to run stably.

@Mike,

You are only running 2 sticks of 2400 Mhz ram correct? Also what is your rams latency?

Threadripper supports quad channel, so only running dual channel is a limitation. Also the infinity fabric that allows the individual dies in your cpu to talk to each other runs at the same frequency as your ram. This is why people recommend the highest frequency ram that you can get to run stably.

+1

Ram Memory Performance is important

Here :
G.Skill 32GB DDR4-3200.
Internal memory: 32 GB
Memory layout (modules x dimensions): 2 x 16 GB
Internal memory type: DDR4
Frequency of memory: 3200 MHz
Memory Form Element: 288-pin DIMM
CAS Latency: 14

Memory Performance 14-14-14-34 (CL-RCD-RP-RAS) (the lower is the better)
Command Rate (CR) 2T

With setting XMP (if possible...) in Bios : good results.

@Mike,

You are only running 2 sticks of 2400 Mhz ram correct?

Yes, but as I mentioned a few posts ago, I tried installing 4 x 8GB modules and there was absolutely no change to render times.

Also what is your rams latency?

15-15-15-39. It's this:
https://www.gskill.com/en/product/f4-2400c15d-32gfxr

Threadripper supports quad channel, so only running dual channel is a limitation. Also the infinity fabric that allows the individual dies in your cpu to talk to each other runs at the same frequency as your ram. This is why people recommend the highest frequency ram that you can get to run stably.

I agree that this may come into play at some point, but I really think the actual problem I'm seeing is the GTX 1070 graphics cards. It looks like resolve's NR only uses one of them.

Okay, the 1080ti has just arrived and I will install it tonight.

Is there any advantage in keeping one of the 1070s is the machine?

Okay, the 1080ti has just arrived and I will install it tonight.

Is there any advantage in keeping one of the 1070s is the machine?

Not really, the GUI is not really heavy for the 1080TI

Okay, the 1080ti has just arrived and I will install it tonight.

Is there any advantage in keeping one of the 1070s is the machine?

Not really, the GUI is not really heavy for the 1080TI

And Resolve won't utilise the 1070 for some processing?

I seem to remember reading that different GPUs can't be used wiht Resolve, but I don't know it that was referring to totally different classes of GPU or included slightly different ones such as the 1080ti and 1070.