Zde je nazor autora:
CUDA is efficient for algorithms, which allow a high level of
parallelism.
Unfortunately RAR algorithm does not belong to them.
If algorithm can use
only less than 10 threads efficiently,
it is better to run it on general CPU,
because one CPU core
is more powerful than one GPU core and typical CPU in
next years
will have 2 - 4, up to 8 cores.