Cryptohaze.com

Same as for Linux.

I may be missing a DLL or so - please test!

Also, multi-GPU is a bit untested right now - that box is busy with some other stuff.

Feedback plz.

(too big for the forum downloads)

http://www.cryptohaze.com/releases/CUDA ... alpha4.zip

Uploaded a new version with the cudart_64_32_16.dll that people seem to need.

looks great and works fine! the automatic configuration for threads and blocks is great!

Good to hear!

There is a bit of a known issue with networking right now - I'm not catching an exception I should be, so... if it crashes as a server when a client disconnects, it's a known issue.

It cracked my test MD4 hash, the speed was ~1900M p/s using autotune.
However, it did not close/shutdown after finding 1 hash out of the hashlist, and continued cracking.

Generated 100 [7;7] loweralpha-numeric plains and 900 of other charset, hashed it with NTLM() and fed to multiforcer. Multiforcer cracked precisely 100 hashes and outputted em right. GJ.

Generated 100 [9;9] and 100 [10;10] numeric plains, mixed em with crap, hashed with MD5(). Multiforcer did not miss a single hash. Stable in that department.

What does the zero copy option do?

Is there a way to "dump" parameters, which autotune picked as best ?

Win7 x64 with SP1
GTX 480 @ 800 Mhz

Rolf - thanks for the robust testing!

Rolf wrote:It cracked my test MD4 hash, the speed was ~1900M p/s using autotune.

Right now, autotune actually is ignored. I've replaced it with improved block/thread count based on general GPU size. I really need to prune old arguments out. It was in place because I didn't have good defaults, but at this point, since I have 3 sets of default block/thread counts, it's much closer to optimum.

Rolf wrote:However, it did not close/shutdown after finding 1 hash out of the hashlist, and continued cracking.

Known issue. There is an "--exit-after" option (unsure as to the exact parameter) that will stop after the specified number of hashes, but the current release doesn't check to see if you've cracked all the hashes. I've fixed this on password length change, but you raise the point that I should also check after each found hash & exit if everything has been found. That's a quick fix that will get in the next release.

Rolf wrote:Generated 100 [7;7] loweralpha-numeric plains and 900 of other charset, hashed it with NTLM() and fed to multiforcer. Multiforcer cracked precisely 100 hashes and outputted em right. GJ.

Generated 100 [9;9] and 100 [10;10] numeric plains, mixed em with crap, hashed with MD5(). Multiforcer did not miss a single hash. Stable in that department.

Awesome I've spent a lot of time trying to make sure it's 100% solid in actually finding all hashes. Glad to hear it passed your tests as well!

Rolf wrote:What does the zero copy option do?

Enables CUDA zero copy? Certain GPUs can access host memory directly (GTX200, Fermis, integrated chipsets). Zero copy uses this functionality. It doesn't have much benefit on the discrete GPUs, but for laptops that share host memory with the GPU, it allows you to simply have the GPU access the application memory space directly instead of copying from host memory to another chunk of host memory. I added it for curiosity, but it doesn't seem to have much of any performance impact.

Rolf wrote:Is there a way to "dump" parameters, which autotune picked as best ?

Since it doesn't actually do anything right now, no.

I am working on a wiki for the site that will contain a lot more information on the command lines/etc. I should be able to get it up this weekend - I've got a local dev copy I've been updating.

Thanks for the testing!

My pleasure.

I may be wrong, but Teslas may also have zero copy capability.

Rolf wrote:I may be wrong, but Teslas may also have zero copy capability.

You are correct.

I don't actually differentiate between Tesla & the core they're based on - the differences are no GPU output, more floating point power, and a lot more RAM.

A GTX580 & a Tesla based on the Fermi core have the same compute capability, and therefore the same features.

... plus, nobody actually uses Teslas for cracking unless they're on EC2.

Yeah, but NV artificially reduced GeForce DP FP performance to 1/4 of Tesla's, and artificially disabled crap like TCC mode, ECC capability for memory, etc etc.

I guess it's "nothing personal, just business".

just gave this a quick spin, I like the new interface was like 'woah' when it opened.

Getting around 1.3 Billion for NTLM single hash on GTX 470 with no extra switches.

GW