ID 328063 - More efficient Multi-Threaded CPU support when rendering SphericalTransform nodes

Follow
  1. The Foundry SANDBOX
  2. Bug Tracker
  3. Nuke

Problem summary

When rendering scenes which include a SphericalTransform node, it is faster to use a PC setup with less CPU cores than a more powerful model.

Customers have reported the following when rendering larger scenes:
  • Dual 10-core hyper threaded CPU (=40 threads)  = 4 Minute render
  • Single 10-core CPU (=20 threads) = >1 minute
  • 6-core CPU (12 threads) = >40 seconds

 

I have tested this myself with the same CPU and found that decreasing the number of threads when rendering increases the speed.

 

Customer reported version
nuke.10.5v5

 

Customer reported platform
centos7

 

Steps to reproduce

1) Read your footage into Nuke   (I used 3 Frames - Lat Long map - 92mb each)

2) Attach your Read node to a SphericalTransform node

3) Setup your SphericalTransform  (For my footage, I used 'Lat Long map' to Cube)

4) Create a Write node and connect it to the SphericalTransform node

5) Render the footage using the Frame Server (Background Render Pre Nuke 11)

6) Open a 'Background Renders' tab, and save your render times.

7) Save your project

8) Head to "Preferences -> Threads/Processes -> export renders"

9) Select 'Customise renderer limits' and lower your theads.

10) Restart Nuke, open your saved file, follow steps 5 & 6

Result: Multi GPU renders is slower, as displayed below:

Speeds:
24 Threads:  09-11 seconds each
​

​

12 Threads: 7-10 seconds each
​
​

6 Threads: 5 seconds each
​
​
 

Workaround
Using less threads will speed up the SphericalTransform node render times, but it may slow down other nodes which are being rendered.
Thus it is best to not lower the Threads unless only rendering this specific node.

 

Reproduced by support
This bug has been reproduced in:
Nuke11.0v1 - Windows 7 - Mac10.12 - CentOS6.9
Nuke11.0v3 - Windows 7 - Mac10.12 - CentOS6.9
Nuke11.0v1 - Windows 7 - Mac10.12 - CentOS6.9
Nuke10.5v6 - Windows 7 - Mac10.12 - CentOS6.9
Nuke10.0v6 - Windows 7 - Mac10.12 - CentOS6.9
Nuke9.0v9 - Windows 7 - Mac10.12 - CentOS6.9
Nuke8.0v7 - Windows 7 - Mac10.12 - CentOS6.9
Nuke7.0v10 - Windows 7 - Mac10.12 - CentOS6.9

 

Earliest version tested
- This issue appears to be in all versions of the product

 

Expected behaviour
Customers would like to fully utilise their higher threaded CPUs when rendering VR & AR spherical transformations.

 

Actual behaviour

It is faster to render SphericalTransform nodes with less threads. 

    We're sorry to hear that

    Please tell us why