AMD 3990x: 13.22 w/ dual Nvidia Quadro RTX 6000's: 75.24

Started by jhapeman, August 13, 2020, 06:22:30 PM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

jhapeman

The title says it all....I am waiting on the NVLink adapter to join my two cards together but I'm guessing that won't make a difference with the scene that the benchmark is rendering.  I built this box for rendering 360-degree jewelry renderings, it definitely is *fast*.

mafrieger

just wow! 75.24 seems to be the best score in this forum so far :-D

would be very interesting to see if NV-Link Bridge could optimize the scaling of 2 GPUs even further.
Since the score for 1 RTX 6000 is 42.27 (see https://forum.keyshot.com/index.php?topic=26507.0)
the ideal scaling would lead to  84.54. Your 75.24 is around 89% of this so there is still (a little) room for improvement.. ;-)

mafrieger

btw: would you mind benchmarking your CPU with SMT disabled, so that you have only one thread for every core (64C/64T)?

The idea behind this:
windows scheduler is still -- hmm lets call it "not optimized" for that many threads.
With this, "SMT off" which means less threads could in some cases lead to higher performance...

An other reason for higher performance with less threads could be lower thermal load and with this higher core clocks, which overcompensate the missing threads...

jhapeman

Quote from: mafrieger on August 25, 2020, 11:42:23 AM
just wow! 75.24 seems to be the best score in this forum so far :-D

would be very interesting to see if NV-Link Bridge could optimize the scaling of 2 GPUs even further.
Since the score for 1 RTX 6000 is 42.27 (see https://forum.keyshot.com/index.php?topic=26507.0)
the ideal scaling would lead to  84.54. Your 75.24 is around 89% of this so there is still (a little) room for improvement.. ;-)

Sorry for the delayed reply; I have the NV-Link Bridge installed and it has no impact at all on performance. It definitely doesn't scale linearly across two GPU's, but awfully close.  Not sure if that's a driver issue (Nvidia) or a Luxion issue, or some combination.

jhapeman

Quote from: mafrieger on August 25, 2020, 11:54:51 AM
btw: would you mind benchmarking your CPU with SMT disabled, so that you have only one thread for every core (64C/64T)?

The idea behind this:
windows scheduler is still -- hmm lets call it "not optimized" for that many threads.
With this, "SMT off" which means less threads could in some cases lead to higher performance...

An other reason for higher performance with less threads could be lower thermal load and with this higher core clocks, which overcompensate the missing threads...

I am using Windows 10 for Workstations where this issue has been addressed, so disabling SMT actually hurt overall performance when I did it.

mafrieger

#5
Quote from: jhapeman on October 11, 2020, 03:04:28 PM
Quote from: mafrieger on August 25, 2020, 11:42:23 AM
just wow! 75.24 seems to be the best score in this forum so far :-D

would be very interesting to see if NV-Link Bridge could optimize the scaling of 2 GPUs even further.
Since the score for 1 RTX 6000 is 42.27 (see https://forum.keyshot.com/index.php?topic=26507.0)
the ideal scaling would lead to  84.54. Your 75.24 is around 89% of this so there is still (a little) room for improvement.. ;-)

Sorry for the delayed reply; I have the NV-Link Bridge installed and it has no impact at all on performance. It definitely doesn't scale linearly across two GPU's, but awfully close.  Not sure if that's a driver issue (Nvidia) or a Luxion issue, or some combination.

interesting, thanks!

just to be sure, did you enable the usage of nvlink in the driver setup?
some hints could be found here:
https://www.pugetsystems.com/labs/support-hardware/How-to-Enable-and-Test-NVIDIA-NVLink-on-Quadro-and-GeForce-RTX-Cards-in-Windows-10-1266/

would be nice to hear how it goes...

btw: in the meantime some new drivers did arrive..

jhapeman

Quote from: mafrieger on October 12, 2020, 04:07:33 AM
interesting, thanks!

just to be sure, did you enable the usage of nvlink in the driver setup?
some hints could be found here:
https://www.pugetsystems.com/labs/support-hardware/How-to-Enable-and-Test-NVIDIA-NVLink-on-Quadro-and-GeForce-RTX-Cards-in-Windows-10-1266/

would be nice to hear how it goes...

btw: in the meantime some new drivers did arrive..

I had the NVLink bridge enabled, and SLI was fully functional, and I'm running the newest drivers.  I suspect that either Keyshot 9 doesn't take advantage of it, or the scenes I am rendering (mostly jewelry) are not nearly complex enough to require it (e.g., they don't require more memory than one card can supply, so therefore memory pooling isn't needed, etc.)