Very interesting post from "vseven" over on the MilkyWay forum regarding running multiple MilkyWay work units:
https://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=3118&postid=67378#67378The TLDR; of it is that you get computation errors in MW work units if you run out of VRAM on your GPU, and he has observed that each WU will peak at using 1.8GB VRAM towards the end of finishing each work unit. So if you are running multiple WU at a time, divide your total VRAM by 1.8 to find out the max to run concurrently.
Examples,
- an 11GB 1080Ti should run a max of 6 concurrent WUs (11 divided by 1.8 equals 6.11)
- an 8GB Vega 64 chould run a max of 4 concurrent WUs (8 divided by 1.8 equals 4.44)