I finally got a task to run and finish and validate on my GTX 1660 Ti (Win 7) machine. This task however ran 3 times longer when compared to my GTX 750 TI (Linux) machine. As the 1660 is much faster than my 750 I would expect that the run times would be correspondingly faster (not slower as it ended up being). [Note: this is comparing the run times of the newer double length tasks.]

I am wondering as well if there is a significant difference between Win 7 and Win 10 run times?

We don't have a windows 7 machine to test on, but it has been reported that windows 7 is consistently slower for MLDS CUDA than windows 10 even with the same hardware. Since its exactly the same code, I don't really know if there's anything I can do, as I think the issue is likely with pytorch or cuda itself only really being optimized/tuned for windows 10.

I know that's not a great answer, but its all I have at the moment.

Strangely enough, my "upgrade" from a GTX750Ti to a 970 also lead to a massive slowdown in throughput. However, the card is used on another motherboard and with another chipset. I'll run various test to investigate to why that is ....

I'm having the same problem with this host:
https://www.mlcathome.org/mlcathome/show_host_detail.php?hostid=11302
My GTX 1650s suddenly slowed down to around 10% and take 20 hours to finish a task.
My GTX 750ti can do it in 3 hrs on a 10 year old i7 in a Dell Optiplex.
Not sure why.

I wish R Zoltan was here.

---