Your stale rates below are atrocious at over 3%!
My hashrate is waaaayy too low for my cards. I need some help getting them to work like they should!
Here's my hashrates
Here's my GPU Load
I have 2 machines running 2 HD6990 each with driver 11.5 and SDK 2.4 in Ubuntu 11.04
I run ./phoenix.py -u
http://user.worker@pool.com:port/ -k poclbm DEVICE=0 VECTORS AGGRESSION=11 FASTLOOP=false BFI_INT
I've tried different AGGRESSION values, tried with and without FASTLOOP and VECTORS and BFI_INT but it doesn't seem to get any better than this. Agression 13 makes the machines feel slow. Haven't tried higher.
CPU load is 100% when mining.
GPU Temperatures are around 80-90 degrees.
I should be able to get 350-375 MHash/sec per worker, right? What should I do to get this?
I don't think you need the FASTLOOP=false there. Also, I think you can push to AGGRESSION=12. I run my 6970 (one core of yours) @ 920MHz core clock and left the memory clock at the default 1375MHz ONLY because it is my workstation that I use for work and play even when mining on Windows 7 x64 Professional. Honestly, Phoenix 1.48 tends to cause too many stale shares compared to poclbm.py [since you will run via python directly]. With my 5850 cards, I use Phoenix 1.48 because the phatk kernel gives me more than a 3% boost and the lost of up to 1% additional to stale shares is acceptable. I typically see 0.3% - 0.5% with poclbm against deepbit.net and now tonight, the same with BTC Guild. Phoenix on the other hand, to the same pools, I average perhaps 1.2%-1.3% stale shares. I sometimes see even worse stale share issues with Phoenix when used with my 6970, so I simply use poclbm.
Phoenix phatk kernel only seems to help with the 58xx series [and maybe older], but actually is worse with the 69xx series in my experience [well, the 6970 anyway which means it will be the same for your 6990].
I would recommend trying poclbm with the following and maybe adjust your -f value down a little every now and then to find the best performance with solid stability; many people run with -f 1 and have no issues.
via the python compiler/interpreter:
python poclbm.py -d1 --host=btcguild.com --port=8332 --user=xxx_xxxx -pass=xxxxx -v -w 128 -f 10
In spite of what other people have said, I have only seen a small but significant decline in performance by using the default worksize [remove the -w switch and it will use 256 or just -w 256], so I set it to 128. My desktop runs with -f 30 and I see 393MH/s and my card runs at 73C-76C depending on the ambient temps and humidity (I NEVER let the temp go above 80C ... your decision how hot you are willing to run it for the long haul). I get he same rates with phoenix, but the stale shares is a deal breaker for me on the 6970. I should point out that I get 393MH/s with my 6970 at these settings [Aero running and two monitors on Windows 7]. Default for the 6970 is 880MHz core and 1375MHz clock and clearly the reduced that for the 6990.
I am not sure that you can clock the same rates on the 6990 [per core] as the 6970 due to the proximity of the GPUs on the board [they tend to reduce clock and/or voltage a little on the dual GPU cards (5970 and 6990 obviously). The 6970 is one hot card at full GPU usage (meaning you need to keep it cool and it is more difficult than other cards). Still, I believe my 920MHz core clock speed is probably quite conservative but that is because I can't really afford to risk the card on this workstation ... well, I guess I can if I am willing to throw my old MSI 570GTX OC back in which I loved but was a terrible mining card; I don't have another backup card for mining [or I would probably get it in a rig and mine with it now].
So, depending no the quality of your pool and the connection too it, you should get a solid 1% increase in performance by switching miners to poclbm. Look at overclocking the core clock some [little at a time, and do some checking on what works for others, but don't start at their rates ... each card is different given its own micro environment and may not behave well at the clock speeds that you find, but I am sure you should be able to gain some improvement; at least 10%. Reducing memory clock speed helps to reduce the heat as well, since mining doesn't require much from memory. I have read the rule of thumb is 35-40% of the core clock speed, but trying that on my 5850s has not proven to work for me and I have not tried it on my 6970 due to me using it as a work station (VPN, remote desktop, email, lots of services yada yada). BTW, I have found performance on Linux to be disappointing, but I think others can probably help you with that assuming there is a way to make the Linux drivers push the card as it does with Windows.
Experiment for a few days and baby sit it a bit when you start tinkering with the clock settings [each card may perform best with different timings] and check on stale/reject rates .. sometimes they are rejected as invalid which may mean your card is making mistakes if you push it too hard. I spent 9 hours with my first 5850 when I first set it up for mining as I slowly tweaked the clock speeds. Good thing too, since every now and again, it would just halt, but you couldn't necessarily tell by the fans [I could see if I was watching miner contributions at the pool website].
I hope I was of help. Please report back on what you come up with.
