Okay, the first GPU-enabled version is ready. Right now it only does the sieving on the GPU. Primality tests are still done on the CPU. The code hasn't been optimized at all yet; the performance is equivalent (+- a few %) on these setups:
1) Phenom x6 1055T, all 6 cores
2) Phenom x6 1055T, all 6 cores, plus a HD6990 using both cores
I don't think the current version is worth releasing. It still needs a bit of work, I want it to be twice as fast. I estimate there is room for a five-fold improvement, especially when I get the primality tests done on the GPU as well.
Tomorrow I will do tests using a slow CPU with lots of GPU power (Sempron 140 with 2x6990).
That's better than I expected - don't most projects like this have worse performance when they first start working than the equivalent? I remember valve talking about optimizing left 4 dead for linux and saying they were getting 2-3 fps on linux when it first started running compared to 200 fps on the same machine in windows.
Maybe it would help to have someone test this on a machine with an intel processor too - sometimes the proc brand can make a pretty big difference even when two procs are equivalent in performance overall.