OpenCL optimisation for AES encryption, Hash and more

Status
Not open for further replies.

troun

Dabbler
Joined
Jul 13, 2013
Messages
33
Obviously the question is whenever FreeBSD (and Freenas) will benefits of OpenCL optimisations.
OpenCL show some interesting benefits in concrete topics like AES encryption, that could be used for SSH and/or disk Encryption.
http://developer.amd.com/resources/...articles-whitepapers/bulk-encryption-on-gpus/

We could even imagine accelerate the RAIDZ parity calculation, or NFS transactions as OpenCL show benefits in MD5 hash or checksum calculations, but I guess it may be more complicated (kernel and not application level?).

I particularly think to Brazos or Ion platforms.
 

Whattteva

Wizard
Joined
Mar 5, 2013
Messages
1,824
I think it probably would depend on if there are BSD drivers and API for those AMD GPU/IGP.
 

cyberjock

Inactive Account
Joined
Mar 25, 2012
Messages
19,526
I'd be scared to use OpenCL for parity calculations if it didn't use ECC memory. For integrated graphics, that's a possibility. But for discrete graphics, all of the standard desktop video cards don't use ECC memory.
 

troun

Dabbler
Joined
Jul 13, 2013
Messages
33
Whattteva, -> I though CUDA had been ported to FreeBSD somehow.
This is really too bad, in such application (crypto) it seems that we could expect factor x2 to x5.

If I believe this thread, it means that instead of 30MB/sec with an AMD E-350, you could get performance of much pricier and power-hungry Intel I5 or I7.


cyberjock, -> I guess for sensitive application you are right (which may not use Brazos proc) but for an home server, I'm not sure. I even think they are even checksum for data transfert between central RAM and VRAM.
Anyway, if at least encryption would be accelerated, it would be already a huge step...
 

Whattteva

Wizard
Joined
Mar 5, 2013
Messages
1,824
CUDA is Nvidia's API. Good luck getting that to run on AMD IGP, lol.
It is true though that running vector-mode algorithms on AMD IGP will typically be orders of magnitudes faster than running that similar code on an Intel chip (may be closer with their latest HD4000? someone correct me on this).
 
Status
Not open for further replies.
Top