Holomark3 WIP - updates, wishes and discussions

Hi guys, I am working on Holomark3 and one of the new things I’ll test out is raytracing where I let Rhino work for x seconds and then I check how many passes it manages on this scene:

Currently I let it work for 60 seconds to get a descent score, does that sound OK to wait for a single test? Or phrased differently: How long is it OK for a benchmark to run? If I aim at letting the entire Holomark3 test take 5 minutes, does that sound OK? Can it be longer, or should I aim at shorter?

At 60 seconds I get 140 passes on a GTX1070.


There isn’t really a right or wrong to this. My feeling is that 60 seconds is the minimum you should let it run. Do GPUs perform diffrentely in earlier stages of the render process compared to later?

For the entire test I think that 5min is the minimum as well. 10min would be the max. Again, just my feeling.

1 Like

And that’s just what I am after, thanks!

Exciting to hear you’re working on Holomark 3!
Personal opinion here, no actual data to back up my thoughts.
I think that a 60 second GPU test would be the minimum for sure.
I could see a 2 minute GPU test and another 5 for the remainder of the tests.
When it comes to bench marking I’m happy letting a test run that little extra to try and get a more balanced / accurate result.


a little off topic but a neat render, raytraced is cycles?

1 Like

Surely is off topic, but I appreciate it! :smiley: And yes, Cycles is Raytraced. It’s important that the benchmark only uses built in functions IMO. That said, I will need to adapt it so it doesn’t downscale on laptops though. (Thanks for the off topic comment, I would not have thought about that right now if it wasn’t for you :wink: )

1 Like

Small update on the progress on Holomark3:

I just made a serial vs parallel projection of 40.000 points to surface test.

And this is the result of the test on my old dual 6 core xeon machine (12 cores, 24 threads)

serial = 15.86 sec
parallel = 1.34 sec
Speedgain= 11.82 x

1 Like

Refined the tool, made a better looking surface for the test and increased the initial points to 90.000:

Dual CPU workstation: (12 cores total, 24 threads)
Calculating base grid with 90000 points
serial = 41.65 sec
parallel = 3.43 sec
Speedgain= 12.14 X

Single CPU laptop: (4 cores total, 8 threads)
Calculating base grid with 90000 points
serial = 27.06 sec
parallel = 7.64 sec
Speedgain= 3.54 X

Just posted some images of the new benchmark file here:
(The motorcycle will be used for OpenGL testing and the glasses for Cycles)


Click the link to see some more pictures and some info on the render setup.

1 Like