2 GPU not identical / PCIe 16x AND/OR 8x & GUI for compute ?

2 GPU not identical & PCIe 8x / 16x OR 16x PCIe / 16x & GUI for compute: good or bad?

I did a test from the "candle test" http://liftgammagain.com/forum/index.ph ... mark.3718/
and here are my results.

The PC (short) : ASUS 99 A-II & 6850k & 32 G RAM & W10 anniversary & OS on SSD + DEKLINK 4K EXTREM 6G.
All PCIe slots are GEN 3.
2 GPUs : GTX980Ti (6Gb VRAM) & GTX 1070 (8Gb VRAM)


1) GTX980Ti = PCIE X8 (GUI) / GTX1070 = PCIE X16 / "USE display GPU for compute" = ON / DECKLINK 4K = ON (in preference)
TNR
1= 24 (fps)
2= 24
4= 17/18
6= 11/12
BLUR
9= 24
18= 24
30= 22
66= 12

2) GTX980Ti = PCIE X8 (GUI) / GTX1070 = PCIE X16 / "USE display GPU for compute" = OFF / DECKLINK 4K = ON
TNR
1= 24
2= 24
4= 13
6= 8/9
BLUR
9= 24
18= 22
30= 13
66= 8/9

3) GTX980Ti = PCIE X16 (GUI) / GTX1070 = PCIE X16 / "USE display GPU for compute" = ON / DECKLINK 4K = removed
TNR
1= 24
2= 24
4= 17
6= 11/12
BLUR
9= 24
18= 24
30= 24
66= 12

4) GTX980Ti = PCIE X16 (GUI) / GTX1070 = PCIE X16 / "USE display GPU for compute" = OFF / DECKLINK 4K = removed
TNR
1= 24
2= 24
4= 13
6= 8/9
BLUR
9= 24
18= 22
30= 13
66= 6

5) GTX980Ti = PCIE X16 (GUI) / GTX1070 = removed / "USE display GPU for compute" = N/A / DECKLINK 4K = removed
TNR
1= 24
2= 24
4= 13
6= 10
BLUR
9= 24
18= 20
30= 12
66= 6

6) GTX1070 = PCIE X16 (GUI) / GTX980Ti = removed / "USE display GPU for compute" = N/A / DECKLINK 4K = removed
TNR
1= 24
2= 24
4= 12
6= 9
BLUR
9= 24
18= 20
30= 12
66= 6


7) GTX1070 = PCIE X16 (GUI) / GTX980Ti = PCIE X16 / "USE display GPU for compute" = ON / DECKLINK 4K = removed
TNR
1= 24
2= 24
4= 17
6= 11
BLUR
9= 24
18= 24
30= 24
66= 12

8) GTX1070 = PCIE X16 (GUI) / GTX980Ti = PCIE X16 / "USE display GPU for compute" = OFF / DECKLINK 4K = removed
TNR
1= 24
2= 24
4= 14
6= 10
BLUR
9= 24
18= 22
30= 23
66= 6

9) GTX1070 = PCIE X8 (GUI) / GTX980Ti = PCIE X16 / "USE display GPU for compute" = ON / DECKLINK 4K = ON
TNR
1= 24
2= 24
4= 17
6= 12
BLUR
9= 24
18= 24
30= 24
66= 12
(I kept this configuration)

10) GTX1070 = PCIE X8 (GUI) / GTX980Ti = PCIE X16 / "USE display GPU for compute" = OFF / DECKLINK 4K = ON
TNR
1= 24
2= 24
4= 15
6= 10
BLUR
9= 24
18= 24
30= 13/14
66= 6/7

0- testing HD and is not recent. Is it suitable for Davinci Resolve 12.5.3?
1- I held that: DECKINK ON or OFF or removed: it does not influence the results
2- GTX980Ti did not support 4K (memory), with the GTX1070 = yes, this is a workaround,
3- 2 cards for those who "almost" the same power, it's worth it to put ON the " USE display GPU for compute "
4- these 2 GPUs: the investment is made but for a new configuration, between 1 and 2 GPU, I think it is better only one very large GPU.
5- PCIE 16x/16X and 8X/16X : found no big difference

Now these two cards are not strictly identical: does it matter?

Going over the supplied information that you yourself(!) provided I have to draw the following conclusion:

'Use display GPU for compute" is BAD. It's very bad. Should be avoided.





…unless… you want more frames per seconds. Then it's good. Very, very good.

It's a hard choice to make for sure. See if you can get someone more qualified to pitch in.

It all depends on how powerful the GPU is, and how much VRAM it has - never seen any ill effects of using UI GPU for computing (but of course can imagine them).

On another note: I downloaded the Candle Test project, but only have 66 Blur nodes in it - how about TNR ones? What am I missing?

Piotr

What am I missing?

In the Color panel, press CTRL+N (or CMD+N on a Mac) to cycle through the Versions.

Thanks, Andree

BTW, on the system as per my sig, I'm getting 12 fps with 6 TNR nodes, and 8.5 fps with 66 Blur nodes (full 24 fps for 30 blur nodes). Of course, my GTX 1080 as the only GPU at the moment is also used for UI. So slightly better results than in Jean-Claude's first scenario (even though he uses 2 GPUs)!

Piotr

PS. I'm gonna add my GTX 750ti (the only single-width card that fits into my overcrowded system), set it for UI, and report back how my 1080 fares without the UI burden....

PPS. Oops - I only have 2 GB on the 750ti, so cannot use it for compute like J-C did...Anyway, unloading the 1080 from UI work is also worth testing...

Going over the supplied information that you yourself(!) provided I have to draw the following conclusion:

'Use display GPU for compute" is BAD. It's very bad. Should be avoided.





…unless… you want more frames per seconds. Then it's good. Very, very good.

It's a hard choice to make for sure. See if you can get someone more qualified to pitch in.

I think it is better to test with its own configuration. I may have misspoke (i'm french) but I think that I must be active "GPU compute" in my case to get the best results.

I think it is better to test with its own configuration. I may have misspoke (i'm french) but I think that I must be active "GPU compute" in my case to get the best results.

Yes, it was an attempt at irony. It's difficult in online forums.

The original poster had already made many tests with the following results:

USE display GPU: always faster
Do NOT USE display GPU: always slower

Then the question was: which is better??? =)

He had already provided the results! For the record I found his post very valuable. Tests and hard facts is all that counts. So many times theoretical calculations don't pan out when you go 'all in' and by the most expensive components.

I think it is better to test with its own configuration. I may have misspoke (i'm french) but I think that I must be active "GPU compute" in my case to get the best results.

Yes, it was an attempt at irony. It's difficult in online forums.

The original poster had already made many tests with the following results:

USE display GPU: always faster
Do NOT USE display GPU: always slower

Then the question was: which is better??? =)

He had already provided the results! For the record I found his post very valuable. Tests and hard facts is all that counts. So many times theoretical calculations don't pan out when you go 'all in' and by the most expensive components.

Andree (not easy),

The better is "for me" and after tests: USE display GPU for the GUI.

But, provided there is a balance between the power of GPU1 (use for compute) and GPU2 for calculations and if possible the CPU. What I see, if the 2 GPUs are not identical, the best is to put the GPU that has the most VRAM in GUI and the lowest in GPU2 (calculation).


GPU1 (for compute)
[0x00002774] | GPUManager | INFO | 2016-11-09 19: 35: 35,807 | Finished initializing CUDA board 0 <geForce GTX 1070> on platform NVIDIA CUDA GPU> clock rate <1784500> multiprocessor count <15> global mem size (MB) , SBM [0 | 0]

[0x00002774] | GPUManager | INFO | 2016-11-09 19: 35: 35,981 | Finished initializing CUDA board 1 <geForce GTX 980 Ti> on platform NVIDIA CUDA GPU> clock rate <1076000> multiprocessor count < 48, SBM [1 | 0]

So, (GPU1) is seen by Davinci Resolve having 6823 MB and must do the display + calculations
(It consumes on the 68236 MB) : it balances "at the best" the 5084 MB (remaining) of the 980 TI.

The curiosity is that when there is only one GPU, Davinci Resolve correctly displays the FULL VRAM of the GPU but when there are 2 GPUs it is always less:
8192 MB become 6823 and 6144 become 5084.

Andree (not easy)

Hi, Claude—let me start at the right end this time: I think your extensive testing with your cards is GREAT. It's the only way to really find out how things will actually play out.

The only thing I found mildly funny was, that you had already done the tests and in all the tests using the GUI card together with the compute card was faster. I don't see how there could be any questions about that at the end. And it sort of doesn't matter if it's "optimal" or not. With the cards you have, you should definitely use the GUI card too.

I didn't mean to add any confusion. I hope you have a system that works well.

Take care

EVERYTHING IS FINE. Thanks and CARPE DIEM

So, (GPU1) is seen by Davinci Resolve having 6823 MB and must do the display + calculations
(It consumes on the 68236 MB) : it balances "at the best" the 5084 MB (remaining) of the 980 TI.

The curiosity is that when there is only one GPU, Davinci Resolve correctly displays the FULL VRAM of the GPU but when there are 2 GPUs it is always less:
8192 MB become 6823 and 6144 become 5084.

Interesting, Jean-Claude; when I was experimenting with my 2GB GTX 750 Ti card as UI-only (in addition of course to my GTX 1080 as compute-only) - I still saw the entire 8192MB of the latter's VRAM as the max available. So would your observation only apply when both cards are set for compute, while the "first" or "system" one - also for GUI?

I since removed the GTX 750 ti as it was occupying the slot shared with M.2, so it could only be one or the other - I decided I need my M.2 cache drive more. But I'm still contemplating purchase of Titan X (Pascal) and using it in the exactly the same manner as you, i.e.

- Titan for UI and compute
- GTX 1080 compute

And I was hoping the entire 8GB would be available for Resolve computing, while the 12 - 8 = 4 GB of the Titan would be used for UI, displays attached and all other Windows stuff that gets loaded into the main "system" GPU's memory. But - from what you're saying - even if i install the 12GB Titan in the first x16 slot, and the 1080 in the second one, Resolve will not see entire 8GB as available for computing? Please elaborate! TIA

Piotr

I do not know how Davinci Resolve compute the available VRAM of installed GPUs but for my case there is only one case (in the logs of Davinci Resolve=> Finished initializing CUDA board....) where it is written that the GPU has all the VRAM: Is when there is only ONE GPU (it does not matter if the slot is in 8X or 16X).

As soon as there are 2 GPUs, there is less VRAM (displayed in the log). It remains true that the "USE display GPU for compute" is ON or OFF.

Only Blackmagic can tell us how the detected VRAMs are calculated. So between (may be) what is written and available and mostly used are (maybe) 2 different things.

TIPS : Neat VIDEO OFX displays all the VRAMs of the 2 GPUs and compute (in preference) its optimization from the 2 full VRAMs.

That's why I say: it is only by testing either that we know with certainty the VRAM when there are 2 GPU. I think that a TITAN-X 12 GB GUI + compute and a GTX1080 in compute will have about 1 GB less but as the 2 GPU will compute at the same time: so it will be a benefit.

Thanks, Andree

BTW, on the system as per my sig, I'm getting 12 fps with 6 TNR nodes, and 8.5 fps with 66 Blur nodes (full 24 fps for 30 blur nodes). Of course, my GTX 1080 as the only GPU at the moment is also used for UI. So slightly better results than in Jean-Claude's first scenario (even though he uses 2 GPUs)!.

So I postponed upgrading to a 2x Xeon workstation, as - observing the ratio of my CPU vs. GPU load - I concluded more GPU power is needed much more than CPU. I'm running now 2x GTX 1080 GPU (without an UI-only one), and am getting a solid 15 fps with all 66 blur nodes - quite a nice improvement over the 8-9 fps I was getting with just single 1080... And still, my 8-core i7 is not the bottleneck; its power would be best balanced with 3x GTX 1080 compute GPUs, I guess...

Piotr

Hi Piotr,
Thanks for sharing the results.

I re-read the results of Blake LaFarm:

viewtopic.php?f=21&t=52561&p=308349&hilit=candle+test#p308342

Between a Dual Xeon 10-Core 2.30GHz E5-2650 and an i7-5960X with 2 GTX1080: 2 fps differences for a 66 blur nodes. You are right to postpone the purchase.

Cheers, Jean-Claude

Piotr