Some initial OpenCL/ATI numbers...

[Bitweasil]

ATI 6970: ~2000M MD5 links/sec.
GTX580 (CUDA): 1300M MD5 links/sec.

Progress...

And it works on CPUs, just is a bit slow.

I get around 21M links/sec on an old quad core AMD.

[Rolf]

Heh, nice!
GZ with getting OpenCL code working, wasnt too hard to port CUDA -> OpenCL, yes ?

I guess that's why ppl rarely implement GPU+CPU attacks - even the best CPU's performance is measly vs the best GPU.

[Bitweasil]

CUDA (runtime API) to OpenCL was a bit of a jump - if I were using the driver API it would be much faster to port.

CPU actually does have some value - it's nowhere near as fast as a discrete GPU, but it's comparable to integrated GPUs on laptops, and for rainbow tables, being able to leverage the GPUs and CPUs on a laptop will be very useful.

For instance, the early unibody Macbook Pros have a Core 2 Duo, a 9400M, and a 9600M. Each of them are slow, but combined I hope to see 100M links/sec or more for NTLM. That's a useful speed.

[Bitweasil]

Om nom nom.

Without BFI_INT but with bitalign:

6970: 2400M MD5 links/sec
6870: 1700M MD5 links/sec

Zomg.

For reference, a GTX580 ticks along at 1300M MD5 links/sec, and a 470 is around 800M MD5 links/sec.

"Slightly faster" is an understatement here...

[Bitweasil]

Looking like around 3500M NTLM/sec on the 6970...

*blinks*

[Bitweasil]

And around 9M NTLM/sec on the CPU of an Atom netbook with the Intel OpenCL library...