Topic: HD5970 + HD6990 experience (Read 2056 times)

hi martok, thanks for your comment.

I did not know about the performance differences and I'll try to find out more.
I used only the latest 2.4 SDK (custom build pyopencl) so it's time for me to
compile diffent pyopencls with the diff SDK versions. I ll post some info soon.

I can just mentoin that I am a little bit disapointed about the HD6990/SDK 2.4
performance (I think there was a design change of the vector calculation 'core')
so I realy expected 2x speed (compared to 5970) (core/core).

After all I can also tell you that I am so sick of the way AMD threats their
customers. The opencl/cal forum is not moderated and is more a bad joke.
If the cards where not that fast compared to nvidia I would completly ignore
amd (gpus). for them is just a marketing feature not even supporting the
academic researches by just some damn open source - so we can all find out
our self. sorry for drifting out of topic but i just checked the amd forum - of
course there is no info about the different SDK speeds.

happy computing

PS. additional info to my hardware

I got 0 monitors attached
I did not use a crossfire cable

What catalyst version did you find works best with this setup? Also, which SDK version are you using. Afaik, the 5970 performs best with 2.1 but the 6990 with 2.3 or even 2.4. So mixing the two may not be ideal unless the miners have sorted this...

Hi forum,

I just want to tell you some experience with mining on a system. Just in case you are planing a system like that.
Here is the system relevant hardware list:

AMD Phenom(tm) II X6 1055T Processor
Chieftec APS-1000C (1000W / 900W for pciE)
OCZ Vertex2 E 2,5" SSD 60 GB
19" 4HU

XFX 2GB HD5970
XFX 4GB HD6990

overall price around 1600 EUR (January 2011)

I am using bitcoin-0.3.21 and m0mchil-poclbm-06cf386 on a Linux ubuntu 2.6.35-24-generic. I don't like ubuntu
but the installation of the AMD (GPU) SDK simply works. I lunch 4 times the poclbm.py in parallel because you
can not specify multiple devices. I think using dual gpu cards - nice and easy - is still not really supported by the
SDK - but from my coding experience with it - it works to start 4 kernels on 4 gpus but this will decrease the over
all computation speed - at least for me...

Just a quick note about overclocking the gpus. I decided not to do that simply because (I may be naive) I hope
the card will have a longer lifetime :-) I did some test with like 100-150mhz on the core and this increased the
fan noise "exp" :-)

Adapter 1 - ATI Radeon HD 5900 Series
Current Clocks :    550           1000
Current Peak :      725           1000
Configurable Peak Range : [550-1000]     [1000-1500]
GPU load :    100%

Adapter 2 - same

Adapter 3 - AMD Radeon HD 6990
Current Clocks :    830           1250
Current Peak :      830           1250
Configurable Peak Range : [500-1200]     [1250-1500]
GPU load :    99%

Adapter 4 - same


Speed results (running 48h+):

26/05/2011 X, 194437 khash/s < HD5970
26/05/2011 X, 194408 khash/s
26/05/2011 X, 340399 khash/s < HD6990
26/05/2011 X, 340302 khash/s


My experience:

I think this - "oh the complexity increases so far because of the gaming gpu kids" - is very overrated.
1. If you do not have a dedicated system for processing it is in my opinion impossible to use or (play with the system
2. My room temperature (20sm) increased like 5 degree (C) - air moisture down
3. It is impossible to sleep next to system - i am a nerd i am used to sleep next to case-less mb only systems (like 10 y ago) -
   but - if you are a computer guy - the noise of my system can be compared to a:
   - Vacuum cleaner full on
   - NetApp/EMC 4HU head (storage)
   - Old cisco or juniper cores
4. Maybe a 2nd system like that would fit in my room - but I thinks it is very impossible to have more
   than 2 or these systems in a room without climate control
5. The hashing speed of the 5970 started with ~250 mhash and decreased constant to 194mhash after like 10min - I am not sure
    for the cause - maybe I don't have enough power - don't know
6. if you are richy rich you can buy some custom water cooled extensions for the HD cards don't known like 500-600 EUR for both.

I did not plan this system for bitcoining but a dedicated opencl computation system... I am
working with cuda/cal/opencl sind 2-3 years mainly with crypto. I am trying to optimize BitcoinMiner.cl because in my opinion
you can have a big speed up by having kernel run times of 5-7 seconds each - of course this only works on a dedicated system
without needs for gfx refresh rates )

If you have questions please ask.

If you have tricks for increasing/optimizing the computation speed without overclocking please tell me

happy computing

(sorry for language difficulties not a native speaker)