Why so many builds?

Different CUDA versions support different hardware (compute capabilites), and different driver versions support different CUDA versions. As such, we need multiple builds to support all the various hardware/software combinations.

GPUSieveSize and older GPUs:
Older GPUs (all GTX 1xxx and older, and any with less than 8GB RAM) may have trouble running with the current default GPUSieveSize setting of 2047. If you have problems, set it to 128 in mfaktc.ini and try increasing it (256, 512, 1024). Higher values give slightly better performance at risk of lag (or crash) if the GPU has insufficient RAM.

Use mfaktc with AutoPrimeNet.

Which build do I use?

The one that matches (or is the closest less-than-or-equal-to) the CUDA version supported by your driver.
e.g. Driver supports CUDA 12.1, use the build that targets CUDA 12.0.

To check your driver's CUDA version, run nvidia-smi, which reports the supported CUDA version on the
first line (on driver 410.72 and newer¹).
If your driver is older than that, consult this table.

Additional Linux builds by NStorm here.