Okay, several people in this thread asked me to do power tests, and I finally got around to it. Unlike my first Kill-A-Watt, this one hasn't melted... yet. I used a slightly older version, cause I'm in the middle of a rewrite for one hash.
For the people who just don't give a fuck about the details and want the results, here: 12.84% more power for 34.34% more hash overall. You can stop reading now.
For the rest of you, the system has a 270X, 280X, 290X, and 7950, all at stock voltages. All power draw measurements were done at the wall.
Power draw at idle - that is, after I booted it and let it sit for 5 - 10 minutes - is 123W.
Running the stock SGMiner 5 from GitHub (commit e481d67e59ad60edc69c026617219f8fae9d6c6e), the hashrate over all cards used was 16.60MH/s while pulling 740W from the wall.
Using the exact same miner code, cloned to a different folder to prevent mistakes, with my OpenCL, the hashrate was 22.30MH/s, using 835W in the process.
The configuration used for both tests follows - the GPUs, in order, are 270X, 290X, 7950, and 280X.
xintensity 128,64,128,128
worksize 64
engine clock 1155,1050,1155,1155
memclock 1500,1600,1500,1600
powertune 20,50,20,50
gpu-threads 2
gpu-fan 80
In closing, I'm glad people kept asking for this, because sitting down and measuring the average increase over most relevant AMD GPUs (Pitcairn, Tahiti, and Hawaii), has been slightly disappointing. My earlier estimates were based off of the hashrate of Hawaii (my 290X) only - seeing this average increase motivates me to do more of the time-consuming, difficult, aggressive optimizations rather than looking for more low-hanging fruit. Power use was also slightly higher than I expected - while I anticipated maybe 8% - 9%, it's actually nearly 13%. While that doesn't make the speed increase not worth it - nowhere near - it still could use some more experimentation and work.
Those power to hash figures aren't that great, you could do much better with a little underclocking...
