03-24-2024, 06:29 PM
This speed difference is specific to a handful of modes/optimized kernels. You've unfortunately picked 2 of the few where it's impactful. What's going on from a technical standpoint is that if there is only 1 hash, we can cheat a bit in doing the math and skip incorrect candidates before we finish processing the full hash by checking specific blocks. This doesn't work for multiple targets because we would need to precompute the early skip stuff for each additional hash and that wouldn't work with how we do our matching currently and likely wouldn't give us the same benefit. The other factors that may play into your speed difference is the mask itself and how much work is available. The mask may be large _overall_ but the minimum divisions may still not be perfectly ideal. The benchmark runs a "real" mask attack under the hood, but with the mask ?b?b?b?b?b or a similar ideal workload usually. Anything less than that, even ?a?a?a?a?a?a?a or similar, will be ever so slightly less ideal and will incur small penalties in speed. This stacks up with other factors, like heat and/or power related clock speed throttling, small differences in the auto tuner from run to run, other stuff going on on the computer at the same time, small hiccups or delays in communication with the cards, differences between the silicon quality on individual GPUs, differences in GPU models and factory clock speed differences, and for some modes even the length of the plaintexts being processed. There's a lot of little factors that can take a few percent off the top and after you stack them all together, you may start to notice it but I wouldn't really worry about it too much. If you're within 10% or so of someone else running the same or a very similar attack, you're doing fine.