Page 1 of 1

Performance on nMP?

Posted: Thu Feb 13, 2014 1:20 pm
by DavidDoyle
Wondering if anyone has any experience / performance figures from the new Mac Pro yet?

I'm doing noise reduction on a large number of clips on my Ivy Bridge Macbook Pro - performance is ok, the GPU is no help at all though.

In the future I will consider an nMP and it appears the D300/500/700 can be utilised. Some performance figures would be useful to give an idea on what can be achieved.

Also - does Neatvideo take advantage of the dual-GPU's yet?

Thanks,

David.

Posted: Fri Feb 14, 2014 11:50 am
by NVTeam
We have some initial results (measured in Neat Image rather than Neat Video but the speeds are more or less proportional there: divide them by 3 and you get approximate NV fps speeds for 1920x1080, rad=1, bitdepth=8 ) from two MacPro systems with D300 and D700:

1. D300
GPU only (ATI Radeon HD - FirePro D300 Compute Engine #1): 27.8 MPix/sec
GPU only (ATI Radeon HD - FirePro D300 Compute Engine #2): 29 MPix/sec
GPU only (ATI Radeon HD - FirePro D300 Compute Engine #1, ATI Radeon HD - FirePro D300 Compute Engine #2): 45 MPix/sec

2. D700
GPU only (ATI Radeon HD - FirePro D700 Compute Engine #1): 40.5 MPix/sec
GPU only (ATI Radeon HD - FirePro D700 Compute Engine #2): 41.7 MPix/sec
GPU only (ATI Radeon HD - FirePro D700 Compute Engine #1, ATI Radeon HD - FirePro D700 Compute Engine #2): 52 MPix/sec

D500 would be somewhere in between D300 and D700.

As you can see, dual GPUs do work faster than just one of them. The speeds do not add ideally of course. We will further optimize that in the future versions.

Vlad

Thank You

Posted: Fri Feb 14, 2014 1:53 pm
by DavidDoyle
That's excellent news - I presently get just under 6 fps (GPU makes negligible difference) so this could be nearly 3 times as fast just by using GPU's alone.

Were there any CPU+GPU figures available on the Mac Pro's you've used?

Thanks,

David.

Posted: Fri Feb 14, 2014 2:14 pm
by NVTeam
The best combined results from those machines were 46.3 and 66.6 MPix/sec correspondingly.

The D300 machine had 4+4 cores, and D700 - 8+8 cores.

Vlad

Posted: Sun Feb 16, 2014 11:49 pm
by Tim
Here are direct measurements from a new Mac Pro with 2x D300 using Neat Video 3.5, 1920x1080 p, 8 bit, radius 1:

CPU only (8 cores): 7.41 frames/sec
GPU only (ATI Radeon HD - FirePro D300 Compute Engine #2): 9.17 frames/sec
GPU only (ATI Radeon HD - FirePro D300 Compute Engine #1, ATI Radeon HD - FirePro D300 Compute Engine #2): 9.26 frames/sec
CPU (6 cores) and GPU (ATI Radeon HD - FirePro D300 Compute Engine #2): 10.6 frames/sec

Hope this helps

Posted: Tue Feb 18, 2014 12:00 pm
by DavidDoyle
Tim wrote:Here are direct measurements from a new Mac Pro with 2x D300 using Neat Video 3.5, 1920x1080 p, 8 bit, radius 1:

CPU only (8 cores): 7.41 frames/sec
GPU only (ATI Radeon HD - FirePro D300 Compute Engine #2): 9.17 frames/sec
GPU only (ATI Radeon HD - FirePro D300 Compute Engine #1, ATI Radeon HD - FirePro D300 Compute Engine #2): 9.26 frames/sec
CPU (6 cores) and GPU (ATI Radeon HD - FirePro D300 Compute Engine #2): 10.6 frames/sec

Hope this helps
Just to check on these - is there contention on Neat Video affecting the combined CPU and GPU processing in these figures? I get 6.21 frames/sec on my Ivy Bridge Macbook Pro with the same parameters. The GPU on the Macbook Pro provides no additional benefit at all.

Using Neat Image figures provided by Vlad provided quite an improvement in Mpix/sec using combined CPU + GPU which doesn't translate to Neat Video fps.

Extrapolating the M/Pix from Neat Image the nMP 8 core with D700 would give around 13 frames/sec from GPU's alone compared to the 9-10 with the D300.

Just trying to evaluate cost/benefit for the future - not looking to upgrade yet and appreciate there is quite a bit of optimisation still possible for all software in order to exploit the nMP.

Thanks,

David.

Posted: Tue Feb 18, 2014 12:40 pm
by NVTeam
It is quite possible that there is some difference between NI and NV in that part. NV certainly requires and likes more memory bandwidth (including GPU memory bandwidth) so if that component is in shortage (and it usually is) the combined efficiency of multiple GPUs may be lower than expected. Again, we will further optimize that in the future versions.

Vlad