Forge does not multi-thread to take advantage of multiple cores on a
computer because you get faster overall throughput by running multiple
Forges and letting the operating system divide up the processing among
the available cores.
We have run timing tests on a correctly handled original negative
with a light dusting per frame, which is typical of most post
production scans. Very dirty frames will increase compute times,
especially if using motion estimation to detect dirt.
The machine we ran our timing tests on is a 2 X AMD Opteron
Dual-Core AMD Opteron(tm) 2212 (4 cores in total), clocked at 2GHz, 4GB
memory in total. The images are 2K (2048 x 1556) Northlight DPX scans,
all stored on the local disk. The figures are for average overall
throughputs when run on a sequence of 20 frames.
With pre-detected dirt (from an IR scan):
- Forge 1.1v3 32 bit 1 core - 9.0 seconds per frame
- Forge 1.1v3 32 bit 4 core - 2.7 seconds per frame
- Forge 2.0v1 32 bit 1 core - 7.0 seconds per frame
- Forge 2.0v1 32 bit 4 core - 2.0 seconds per frame
- Forge 2.0v1 64 bit 1 core - 5.6 seconds per frame
- Forge 2.0v1 64 bit 4 core - 1.7 seconds per frame
When using Forge's motion estimation to detect dirt:
- Forge 1.1v3 32 bit 1 core - 11.8 seconds per frame
- Forge 1.1v3 32 bit 4 core - 3.6 seconds per frame
- Forge 2.0v1 32 bit 1 core - 8.9 seconds per frame
- Forge 2.0v1 32 bit 4 core - 2.5 seconds per frame
- Forge 2.0v1 64 bit 1 core - 7.5 seconds per frame
- Forge 2.0v1 64 bit 4 core - 2.1 seconds per frame