FPGA for MD5: how fast can it go?

[oniric]

Hi all, I've just installed CUDA drivers for my notebook video card and with BarsWF CUDA bruteforcer I can check about 120M passwords per second. Now, I'm interested in raising this value and I'm thinking about FPGA. I know there are many models around, some of them are quite out of my budget but there are models from xilinx or altera that don't cost so much.

So my question is: how many passwords per second can such a "entry level" model check against an MD5 hash?

[gromeo]

go with the cuda and don't spend money on fpga unless you really need to. The fpga's cost a fortune. Are you sure you will regain your investment?

[oniric]

go with the cuda and don't spend money on fpga unless you really need to. The fpga's cost a fortune. Are you sure you will regain your investment?

Not at all I'm just saying that I've noticed there are FPGA for evey budget, like CYCLONE III series from Altera. I'm wondering how could they perform.

[tek911]

Not at all I'm just saying that I've noticed there are FPGA for evey budget, like CYCLONE III series from Altera. I'm wondering how could they perform.

I'm with gromeo. I chased the fpga path, and there are plenty of them out there. But in regard to whats already implemented code, (unless you want to get comfortable with another and implement yourself) you mainly looking at going with something like a pico e12 or something like that.

That being said I'm all for someone picking up one of the budget fpgas and running with tools for them. I'd be a happy guy, but since cuda hit i'm on the cuda wagon.

edit: when i mean already existing code, i'm talking table generation tools, winzip crack, etc. Most of that is written for pico fpgas.
http://forums.remote-exploit.org/showthread.php?t=10816

[thrasher12ga]

FWIW (you may already be aware of this) but work has been done using the PS3 console against MD5, with some pretty decent results. A few different projects are around for this, IIRC including a distributed platform. The code for PS3 MD5 is evolving and improving, apparently with much room for optimisation still there. They haven't been able to access the GPU yet, only the Cell SPU's. Even so, a figure of 80M-120M hashes/sec isn't bad.
Guggle "Rensselaer Polytechnic" and "Distributed Hash Cracker".. you could use HTTP to send the hash back to the office and use the hardware there to do the legwork.
OFC precomp/rainbow is probably an easier way to go unless the hash is salted/stretched.