I originally posted this as a reply to a topic of a guy wondering why his 1500-watt PSU was overloading (rather, "why the computer just shuts off") with his quad-GPU rig. I figure it could be very useful to anyone running a serious mining operation, so I'm making this into a proper post...
For one thing, sucking up 1500 watts is about the equivalent of running a hair dryer constantly... or one of those box room heaters... or something obscenely power-hungry like that. I mean, we're talking like, "dude, you'd better check your power cord to the wall outlet and make sure it's a couple gauges heavier than normal" kind of crap. I'll bet you it gets warm, if not hot. So in light of that topic of conversation, you really need to make sure you've run the ELECTRIC BILL analysis on this kind of completely ludicrous rig... it's going to cost you almost triple-digits on your power bill, so you'd better be CERTAIN you're going to make that back in Bitcoin. Not to mention, if you use it indoors and you need to run air conditioning in that area, it's going to cost about 1.5x that amount in electricity to remove the heat produced by the PC. It's no joking matter. It's summer here and I run my mining rig literally, outside on the porch
(2nd floor = great anti-theft), to offset the cooling problem. And I'm still barely breaking even on the power bill, if that much.
That said, your GPUs are stupidly over-powered. Knock them back to pure stock and you should be fine. You also need to make sure:
1) Your CPU usage is ZERO PERCENT
while running miners. Any indication of CPU load is typically a rat's nest of API version, driver version, miner settings, etc., that results in any amount of CPU usage caused by a GPU miner. This was a long struggle for me to find the best way to solve. Basically it boils down to one thing: start with low aggression settings, and test each setting for 10-20 seconds to see which one starts causing CPU usage while you're seeing a "Mhash/sec" reading (actually running, that is). If it causes CPU usage, knock the aggression down to the last one you saw 0 at. And drivers, evidently there are some issues with Catalyst 11.7 and 11.8 (the latest), so you ought to move to either 11.6 (browse AMD's archive list), or find the 11.9 beta at guru3d.
2) Always keep your eyes on each GPU's temperature
. With that many GPUs, god knows how you fit them in the case. Fans don't move air when they're in a vacuum, so be sure you either have a handful of fans COORDINATED to push air from the front of the case through the back (pulling air in is always easier than pushing it out, it seems), or just leave your case side off and direct a blower fan at it or something. But I can tell you one thing: those GPUs were NEVER designed to operate at full-capacity, right next to each other, for an extended period of time. You've got to make sure you stay on top of their "health" to make sure they stay working properly.
. First of all, don't "just" overclock. Mining is a very unusual situation to put a GPU in, so it doesn't need the same uber-high clock speeds that 3D rendering usually requires. For example, on my 6770, setting the memory clock (default at 1200, I think) to 300MHz actually puts the GPU in some sort of timing mode that essentially super-charges its mining speed, but drops about 10 C off its running temperature. Maybe it shuts down some unused clock modules or something. Whatever it does, it lets me run the thing at 950MHz/300Mhz (core/mem) when its default is more like 850/1200, and bangs out an extra 30 Mhash/sec while keeping the GPU core cooler. Find the right settings, search around, and don't just "move the slider up". And FFS, put the voltage back to stock, that's a freakin' sweet way to nuke your cards in no time...
4) Clients and settings
. There's a HUUUUGGEEE difference that can be made between one client and another, and between one parameter and another. For example, my GPU goes from 211Mhash/sec to 160Mhash/sec when I just remove "worksize=128"; the kernel (currently using a modified phatk posted here somewhere, but very similar to the phatk2 provided with Phoenix) defaults to a worksize of 256 and my GPU only has a 128-bit memory bus (it doesn't know that). Performance goes through the floor. You may mess your pants when you see some 300-odd Mhash/sec, but what you don't know is that it could be straining itself with some bad settings and you could
be seeing 400 if you played around a bit. "BFI_INT" is pretty much a necessity. Play with "VECTORS"/"VECTORS2"/"VECTORS4" in combination with different "WORKSIZE=" (64, 128, 256) values - the "vectors" values mesh with the given "worksize" to produce each sort of work "chunk", and finding the right combination is key to a smooth-running system.
5) You don't need a new power supply
, just quit sucking power out of the wall like it's going out of style! Run the numbers - consider that 1500 watts continuous is what you're drawing (since we already know we're blowing past that as it is), run that as kilowatt-hours (kWh), which is approximately 1.5kW (long form, "1.5 kWh per hour"), and run that out to running 24/7 for a month. 732 hours in an average month, so we're talking... 1,098 kWh for your rig each month. And we plug price into it at about my average power rate, which is $0.15/kWh, and... *click click*... yeah, that'll run you $164.70/month in power to the box alone. Remember what I said about air-conditioning; when you factor in the $247.05 for cooling the house from the heat generated by the rig, that's $411.75 a month. 'Ya sweating yet?
I dunno, maybe I tldr'd here, spent 20 minutes writing this crap for one guy's rig, but dude, seriously, 2000 bucks? I was sweating spending 100 on my 6770 and had to split it with a friend
If it helps you any, there's an address in the sig that could help me buy a compliment to my single 6770!
edit: Oh, and if you want to know how much your miner is REALLY costing you, down to the cent... it's called a "KILL A WATT". Get one!! http://amzn.com/B000RGF29Q
(no, I don't make anything off this link, just convenience - go Google instead if you want.)
Any inaccuracies, explain them and I'd be more than happy to edit this. But I hope this brings a few common problems to light; I spent a huge chunk of time trying to squeeze extra performance out of my rig while watching the Kill-A-Watt meter, running the numbers, etc. Now I hope I can pass this on to someone else that can use it