5970 slow?
#11
(04-19-2013, 10:49 AM)atom Wrote: I dont get what the problem is here. The numbers shown are the ones to expect.

Hi Atom! There's no problem really other than my misunderstanding. If you want to be picky, the problem may be information available about what to expect from specific algorithms. Reading available documentation and the forum extensively I couldn't figure out:
  • public benchmarks are almost always lite, plus ones are single hash anyway. Thus, best speeds reachable.
  • multiple hashes in plus incur in a tangible penalty. I did some tests for sha1: 2 hashes (~ 10%), 200k hashes (~ 20%), 500k hashes (~ 30%), 1M hashes (~ 40%).
  • salt + pass also incur in a tangible penalty. Looks ~ 7% for 200k hashes.

So being a newbie, I expected public benchmark performance and got ~ 25-30% less instead.

Just curious, why is salt + pass significantly slower than pass + salt ?
Could it be done faster for single salt situations?

Thank you!
#12
(04-19-2013, 10:12 AM)uzomoshi Wrote: This is my benchmark 825/1010. I had to set --gpu-accel=240 instead of 800 because every run was killed. I guess due to insufficient memory (I got 1Gb ram).

By default fglrx will only use 50% of available memory. But you can tell it to use all of the available memory:

Code:
export GPU_MAX_ALLOC_PERCENT=100
#13
(04-19-2013, 12:43 PM)uzomoshi Wrote: Just curious, why is salt + pass significantly slower than pass + salt ?
Could it be done faster for single salt situations?

No, that is because of the optimization techniques. w[0] must be variable to feed the gpu with work but in case its a salt, its static.