First off, Hi there! Interesting little project you have here, figured I'd hop in and give it a shot.

Next: I'd just point out that a warning on the site's home page or join page that this project is extremely heavy on system and GPU memory capacities would be highly appreciated.

The reason for this being that 2^64 CPU units flattened my 8c/16t 8GB Xeon system running on 14 of 16 threads within seconds of the unit starting, overwhelming main memory and the 4GB pagefile. 10^20 units were pulling 1600mb and 1470mb on my GTX 660 2GB and GT 440 OEM 1.5GB, respectively, and I'm pretty sure the units on the 1.5GB card were not agreeing with the memory capacity. The 1.5GB card has been disabled via cc_config flags for this project, and I've dropped the default Kernel size to 19.

I also have a question as to what this project likes more when it comes to GPUs: Single precision compute or double precision compute? I know Milkyway@home freaking loves DP compute, while SETI is content to putt along on SP compute.

Anyways, thanks.

1) 2^64 app needs a lot of memory per thread now, it can't be avoided at this stage of the search.
2) 10^20 is not supposed to run on GPUs with less than 2 GB memory. It's probably a bug that BOINC assigned it to your GT 440 OEM 1.5 GB.
3) 10^20 app uses only integer arithmetic, you can look at performance comparison here: https://sech.me/boinc/Amicable/gpu_list.php

P.S. 10^20 app does use 32-bit floats in a few places outside of main loops, but they take a negligible amount of total execution time. It can even run on GPUs that don't support double precision at all.

Alright, that does explain much for the 2^64 app and why it obliterated my computer. Lots of threads + not a whole lot of memory + memory-hog app = brutal on the memory. Whats the approximate per-thread memory useage for 2^64 anyways?

I gathered as much about the GPU Vram requirements when I was looking around on the forum to see what the possible issue was and how to fix it. Bit of a pity, but eh, I can live with it.

The bit about it being almost purely integer work is a nice bonus though. Being able to just chuck in any >/=2GB GPU and let it cook WUs is nice without having to worry about whether it prefers SP or DP compute, because DP-capable cards that actually do it at a decent ratio to SP are expensive.

Also, you will need 1 CPU per WU running on the GPU(s).
This is not typical of most GPU projects.