hashcat Forum › Misc › Hardware v
GPU memory considerations

Threaded Mode
GPU memory considerations
pragmatic Offline
Member
***
Posts: 105
Threads: 23
Joined: Apr 2012
#1
10-25-2012, 07:44 PM
Hello all,

So let me start off by saying that I'm well aware that the amount of GPU memory is essentially meaningless when it comes to performance - this is not yet another of those threads! Smile

That said, are there any implications of memory size one should consider when purchasing a modern card (say 3GB vs 6GB). From my understanding from other threads the memory does dictate how large / how many rules,hashes,dictionaries/masks/etc... one can load. Have memory sizes increased to the point where that no longer realistically matters, or are there still edge cases where it could be an issue?

Finally, how does memory size scale with multiple cards? If I have 4 3GB cards do I have 12 GB of global memory to work with or is there some bottleneck where we limited to the lowest amount of memory on a single card or something?

Thanks!
Find
Reply
atom Offline
Administrator
*******
Posts: 5,185
Threads: 230
Joined: Apr 2010
#2
10-26-2012, 10:03 AM
It doesnt matter because the OpenCL runtime limits a single GPU memory block to 512mb. In other words you will get out-of-memory Problems regardless of GPU memory size.
Website Find
Reply
Hash-IT Offline
Moderator
*****
Posts: 723
Threads: 85
Joined: Apr 2011
#3
10-26-2012, 11:33 AM
Well, thank you for asking this question pragmatic. I regularly run into out of memory problems when trying to add my rules and just assumed I needed a better graphics card with more RAM.

Looks like you saved me money and frustration !

It makes me wonder if OpenCL has such a limit why do they sell cards with more than 512MB RAM ?

atom, can you please tell us if this is just a bug or poor coding on behalf of OpenCL or is it something more complicated than that ?

I guess what I am asking is will OpenCL ever fix this problem ? Also are graphics card manufacturers misleading the public with more RAM when only a small amount of it can be addressed ?

Thank you.
Find
Reply
atom Offline
Administrator
*******
Posts: 5,185
Threads: 230
Joined: Apr 2010
#4
10-26-2012, 12:03 PM
Well I should say this is for AMD OpenCL runtime only. But the reasons behind? Only AMD can know Smile
Website Find
Reply
Hash-IT Offline
Moderator
*****
Posts: 723
Threads: 85
Joined: Apr 2011
#5
10-26-2012, 01:02 PM
Thank you for your reply atom.

So it sounds as if it could be fixed in the future. However if I remember correctly AMD are not really interested in our "little hash-cracking program" Sad

I will never forgive them for that comment !
Find
Reply
pragmatic Offline
Member
***
Posts: 105
Threads: 23
Joined: Apr 2012
#6
10-26-2012, 09:45 PM
(10-26-2012, 10:03 AM)atom Wrote: It doesnt matter because the OpenCL runtime limits a single GPU memory block to 512mb. In other words you will get out-of-memory Problems regardless of GPU memory size.

Does that mean that it's 512 MB per GPU, so 4 GPUs is 2 GB?
Find
Reply
atom Offline
Administrator
*******
Posts: 5,185
Threads: 230
Joined: Apr 2010
#7
10-27-2012, 08:28 AM
Yes, but it wont help since we need to copy the hashlists to each device on its own to have fast access.
Website Find
Reply