07-11-2012, 12:29 AM
In fact, sha512unix is not very much faster on GPU as compared to CPU. If bruteforcing 5-char passwords is not practical on CPUs (your charset is too large?), then it's very likely it won't be practical on GPUs too.
The algorithm itself is moderately GPU-unfriendly because of the 64-bit operations that are not natively supported on GPUs and are emulated in software. But the biggest problem is that the GPR pressure is so high that you cannot utilize the hardware properly. The GPU just cannot run enough wavefronts due to limited resources and ALU units end up underutilized. There is also a paradox, under certain circumstances, sha512unix can be made faster by offloading part of the algorithm on the CPU rather than doing all of it on the GPU.
The algorithm itself is moderately GPU-unfriendly because of the 64-bit operations that are not natively supported on GPUs and are emulated in software. But the biggest problem is that the GPR pressure is so high that you cannot utilize the hardware properly. The GPU just cannot run enough wavefronts due to limited resources and ALU units end up underutilized. There is also a paradox, under certain circumstances, sha512unix can be made faster by offloading part of the algorithm on the CPU rather than doing all of it on the GPU.