Linux GPU Mining - Efficiency and Optimization

[humble]

I started mining in earnest last week with two cheap Ubuntu 10.10 machines, one equipped with a Radeon 5870 and one with a 6870, both running Diablo.

I'm getting 290M on the 5870 and about 220 on the 6870. According to the ratings at http://pastebin.com/AvymGnMJ this is almost 10% less than they should be capable of.

I have all of the compiz effects turned off and as little screen activity as possible going on (blank screen-saver). The only flag I have set is a worksize of 128. I've tried 64 and 256 but it didn't seem to make much difference. I tried the v flag once but it didn't seemed to be supported (threw an error of some kind).

Any suggestions/tips for how to squeeze more out of them?

[Essjay]

I'm getting ... about 220 on the 6870.

I'm not sure where the data in http://pastebin.com/AvymGnMJ came from. But I was running at about 220 on my 6870. (Using m0mchil's miner under Windows 7.)

If you want a quick boost, you can try overclocking it. I'm not sure what tools are available for Linux to do this. I would typically overclock my 6870 from 950MHz core to 1050MHz, and boost the fan speed to compensate. This was good for about 250Mhash/sec.

[theymos]

That data was compiled by ArtForz, who is using more efficient mining code. Those might also be theoretical maximums.

[Raulo]

Before I overclocked it, my 5850 got 244 Mhash/s with m0mchil's code so a bit better than this card in the table. I use SDK 2.1. SDK 2.3 was slower but I don't remember by how much. What SDK version do you use?

[humble]

I'm using version 2.3. Will look into overclocking and see what I can do there. Thanks.

[Raulo]

I'm using version 2.3. Will look into overclocking and see what I can do there. Thanks.

Optimize first with the standard clock. Use SDK 2.1, it is faster but probably not 10% faster so it will not explain everything. Did you look at all the options? Low -f gives a few percent speedup (but makes desktop usage impossible).