10-05-2023, 06:03 AM
(09-28-2023, 12:23 AM)Chick3nman Wrote: Unfortunately, that's not how any of that works. 80% of Hashcat's algorithms will be slower on even a fairly power FPGA than on a decent GPU, simply by nature of the limitations of modern FPGA platforms and the strengths of modern GPUs. FPGAs are not better at bcrypt because they are "optimized for bitcoin mining", that makes absolutely no sense at all. FPGAs are better at bcrypt because GPUs have limited L1 cache and each bcrypt thread needs 4KB of fast nearby memory, which many FPGAs have plenty of. I think you should probably take a closer look at the differences and especially the limitations of each platform before you jump into anything, you may be surprised. Modern GPUs are extremely good at parallel processing for fast algorithms due to their massive core counts and high clock speeds. FPGAs rarely come close due to the limitations on things like clock speeds and power. For many of the slower algorithms FPGAs might be more power efficient per unit of computation, but that's not going to be the case for the fast algorithms where they can't even keep up. And even for algorithms where they do end up more efficient, if they are ultimately much slower, which tends to be the case, they will not be a "better" option.
You could spend 9000$ on a U250 and another 3000$ on licensing to operate it(yes really, you have to do this per device), then spend probably a couple hundred thousand dollars on dev work to get every algorithm and attack mode implemented and you'd come out with great bcrypt speeds, and mediocre speeds in every other algorithm OR you could spend 1800$ on a 4090, have worse bcrypt speeds, but vastly superior speeds in every other algorithm, extreme flexibility, and nearly costless development and you'd maybe spend a few cents to a few bucks a month extra on power. This is why FPGAs are extremely rare for hashcat, they simply don't make sense for 99% of users. For the users that it DOES make sense for, they are already using them.
OK, I will state the goals of my project. I want a budget highspeed unit for a defcon killer.
There are over 200 companies making FPGA units for AI, I have had a chance to test only a few of them and you are 100% right My 10 4090's with custom firmware run 40000% faster than the same FPGA setup. Now I will say if my power was not an issue I would try a 100x FPGA setup, for a cost-for-cost test.
I am documenting the speed difference, from a basic built crunch 1-12 alpha number list, ver a bruteforce with the same variables. it's all about the disk space game now, highspeed SSD raids are sexy but not cheap.