650W should be no problem.  Even a small overvolt is a possibility but you will find that efficiency of the PSU will further amplify the diminshing returns of overvolting.  If you are running the cards at stock voltage then I'd say 650W is ideal.

Do be careful with overvolting though.  My dual 5850 system is currently drawing 304W from the wall for 724.3 MH/s from a 0.1V undervolt.  A 0.1V overvolt is going to send the power consumption far in the other direction and more might push the system over 650W.  I would not overvolt without a power meter so that I can be sure that the PSU is not being overworked.

I have a similar setup, my miner is mounted on a wooden board and is very close to where I sleep in a small room.  I very much do not like noise and generally tolerate much less than others but I've finally gotten my system to a point where heat is more of a concern than noise.  Before obtaining this miner all of my computers were solid-state.  I only attached HDDs using a USB harddrive enclosure so I could switch them off when not using them.  I've not had experience running a fan in my room 24-7 for nearly 5 years (I used to have an ordinary computer tower, not particularly noisy, but running it 24-7 drove me mad so I got rid of it and replaced it with a solid-state computer).

You can find pictures of my rig here.

I've attached Zalman VF3000A coolers to my Sapphire HD5850 Xtreme cards.  The coolers are designed for reference 58xx cards and my Xtreme cards are built much more like some of the 6xxx family so I had some difficulty mounting the coolers properly.  The fans on the coolers are 3-pin whereas the fans on the cards are 4-pin so I have to run the fans from the motherboard (if you want the cards to be able to control their fans then be aware of this if you decide to go down my route).  By using the enclosed Fanmates and turning them down to the minimum (which I think is about 50% fan speed) the coolers are pretty quiet (a fair bit quieter than most PSUs I'd say).

I replaced my 650W OCZ PSU with an 850W Scythe Couriki 2 PSU.  This isn't the most efficient or feature filled PSU but it's not bad and it's the best I know of for silence.  This power supply is a fair bit quieter than my two Zalman ZF3000A fans at minimum speed so if your PSU is giving you trouble then I recommend this.  3 5870s should be no bother for it (My 2x5850 system is currently drawing 304W at the wall).

The CPU fan is actually a little quieter than the PSU because it's running at minimum speed (It is rare that CPU usage exceeds 5% so the fan is never asked to go faster).  However, if I manually turn this fan up to full speed it completely drowns out the rest of the system so make sure these fans are at minimum speed.

I'm using a HDD and it clicks away regularly because I'm running bitcoind but I actually quite like the sound of the HDD so I'm not bothered.  You may want to run off of a USB or in RAM to eliminate this sound.

Last, but not least, it is a very good idea to undervolt your cards.  I took mine down by a full 0.1V (from 1.0875V to 0.9875V).  This reduced my stable hash rate from 830 MH/s to 724 MH/s, and this may seem like a lot to lose, but the drop in noise, heat, and power consumption is impressive and I highly recommend this.  The MH/J of my system is now very high (2.38 MH/J); I was barely over 2 before.

With all of this I can finally tolerate the noise level.  It would be nice if it was quieter but I can't think of a way of making this happen without installing water cooling, and I would have to be careful to get a quiet reservoir (I believe my system is quieter than some fully water-cooled systems already).

I'd be interested to learn how you fare with voltages this low or lower (I currently have little information to go on and am trying to improve my MH/J).

I'm still not very sure about minimising voltage but if you are trying to maximise MH/J then I'm almost certain that you need the card voltages below 1V for 5850s.  Of course, you may be willing to pay more than the cost of power for your Bitcoins and so it may make sense to increase the voltage.

When I posted before I was not aware that the VRMs on the 5850s handle voltage in multiples of 0.0125V (I'm not certain of this but fairly confident).  Thus, I was really running at 1.0125V/900/300.

My cards are currently running at 0.9875V/850/300 (352.7 MH/s +/- 0.1) and 0.9875/895/300 (371.6 MH/s +/- 0.2) for a total of (724.3 +/- 0.2).  My power consumption at the wall with the cards running is 304W +/- 2W.  When I close the mining software the cards fall back to 157MHz@0.95V and the power consumption drops to 111W +/- 1W.  I'm still running from my hard-drive and preventing it from going into any kind of power saving mode with bitcoind so there is a saving I'm not applying here too.

I've not had time to verify the stability of this setup but the cards did run all night at 850 and 890 respectively.  895 is a new venture.

Running a card at 0.9875V at 850 MHz yields similar MH/J as running it at 1.0125V at 900 MHz (taking system load into account, if you just count the cards then going up to 1.0125V and clocking up to 900 MHz actually lowers my MH/J).

Honestly, finding the correct voltage and clock rate for the best stable MH/J requires testing as every card is different.  I'll point out that my cards both seem to be pretty good at overclocking so you may find it difficult to even get above 800 MHz at 1V.  I'm using Linux, the cards are configured to give no signal, and I'm using Zalman ZF3000A coolers, and the ambiant temperature where I am is currently peaking at 21*C so I have some natural advantages when it comes to stability too.  My good card clocks to 1025 MHz stably at stock voltage and went for 3 hours at 1110 MHz at 1.25V (461.1 Mh/s) so it's not bad even compared to 5870s!  I'm using everything I learnt in competing for high hash rates to maximise my efficiency at low voltages.  PM me if you'd like more details.

My power supply situation is not optimal for minimising power consumption.

I'm using a Scythe Chouriki 2 850W power supply (80+ silver certified) and it claims to be about 89% efficient at a load of 300W, see http://www.techpowerup.com/reviews/Scythe/SPCR2-850P/5.html.  I was originally using a 650W OCZ supply of some kind but it was quite a bit louder than any other part of my system.

For the purposes of this, the psu is plugged directly into the wall.  In reality it is plugged into a 4-gang which is plugged into a 10 meter long 2-gang but I've put the power meter into the 4-gang and am only interested in the power draw at this point.  When I move I'll probably manage to arrange to have it plugged into the wall directly.

The voltage 'at the wall' is 243.6V.

Lol

I'm already at a single stick of 1GB ram.  I've tried toggling various elements in the bios but there was no apparent change in the power draw so I left them all enabled; I'm not using sata, usb, firewire, or on-board sound, but they are all enabled in the BIOS.  I'll try taking all of these off next time I go through a power cycle and see if it makes a difference when actually running the miner properly.

But I need my lights.  The power LED is particularly important.  I don't see how it's possible to tell whether or not the miner is switched on without the power LED.

Yes, the onboard graphics are disabled.  Thanks for pointing that out though.

I only dropped 7 watts when I parked my drive, perhaps I should try removing it altogether but this would take quite a bit of setting up so I'll leave it.

I tried underclocking the CPU and undervolting the CPU and RAM in the bios and found I instantly dropped 9 watts.  It's possible that when editing the BIOS that the chip and ram are in some kind of high performance mode and that the running miner would behave quite differently but it's caused no harm so I've left it.  The system is now quite a lot slower but, you're right, phoenix and bitcoind together only want 6% of the CPU at all times (used to be much less).

The motherboard is an M4A88T-V EVO.

Putting the bitcoin activity that I needed on a ramdisk and parking the drive reduces my power consumption by 7 watts (now down to 300).  It's a little inconvenient and configuring the system to boot from a usb or from the network doesn't seem to be too much hassle for 7 watts but it is a significant improvement that I didn't think of.

Could you give me a BTC address to donate to cicada?

Reducing the RAM clock and worksize lost me far too much hash rate to be viable.

Thanks for the help.

I've tried disabling the onboard devices but none of it makes even 1 watts difference for me so I left them all at default.  My system pulls 114 watts from the wall when I stop phoenix and am wondering if this 'idle' power consumption is excessive or not.

I do have an AMD Semperon 140 and it's underclocked and undervolted, I've also undervolted the RAM.

I have also worked very hard to maximise cooling and, at minimum fans, my cards are at 60*C and 46*C (added bonus, the miner is very quiet).

The two things I've not tried are doing away with internal drives (I'm currently using a 160GB Western Digital drive and this is being used constantly since I am also running bitcoind with my .bitcoin folder on the drive).  I will investigate parking this drive.

The other thing mentioned here I haven't tried is reducing the RAM speed on the cores significantly below 300 MHz and changing the worksize to 128.  I'll give this a go as it's easy to test.

My power meter finally arrived and so I finally have everything I need to try and maximise MH/J.

I've done some basic things like undervolting the CPU and GPUs slightly but would like to know more about maximising MH/J.  Just to be clear, this is not about maximising profit, and I would like to maintain a reasonable level of hashing speed (at least 75% of record hashing rates for the same cards).

Right now my miner (2 x Sapphire HD 5850 Xtreme) draws 307 Watts (at the wall) and manages 722 MH/s (so 2.35 MH/J).  Any ideas how I might improve this?

As usual, I will tip BTC for significant improvements.

I'm sorry, I've been assuming that the 5xxx cards and 6xxx cards use a different size process.  A quick look online suggests they both use a 40nm process but I can't be sure.  What I do know is that the Sapphire HD5850 Xtreme cards are quite a bit shorter than the reference 5850s and are built more like the 6xxx family.

I appreciate that I've been extremely lucky with my cards.  Perhaps the manufacturer has a lot to do with it.  I do know that when running a GUI and with stock cooling the maximum stable clocks of the cards were around 900 MHz but I didn't do any testing after removing the GUI and replacing the coolers so I can't say which of these two had the most impact.  A card that only goes to 870 MHz even with good cooling, and doesn't accept even a small overvoltage, sounds like a very bad card to me.  I assume that if you undervolt the card you lose stability at 870 MHz.

Ohm's law is "Voltage = Current * Resistance" but there are two caveats to mention here.

1) Current is not Power.  Check out Joule's Law to see that both heat and power are proportional to the square of the current.
2) As temperature increases, resistance increases, so current will grow slower than linearly as voltage is increased.

I agree that the maximum stable clock seems to grow logarithmically as voltage is increased (and so voltage must be increased exponentially to increase maximum stable clock linearly).  However, just as 1 + x is a good approximation for exp(x) for x close to 0, small changes in voltage close to stock voltage will affect maximum clock rate approximately linearly.

Also, I'm fairly sure that increasing the clock rate increases the power consumption linearly.  Combining this with the above, if you increase both your voltage and core clock by 10% you should expect the power consumption (of the core) to increase by 33.1% (a cubic relationship).  I think some people choose to do this is because, for them, 10% more MH/s is actually more valuable than 33.1% more power + shorter card life.  I'm sure some people do it because they think it's profitable when it is not for them.  Some people do it because they are overclockers and are actually slightly more interested in configuring, pushing, and comparing cards than the bitcoins.  These people will invest in after-market coolers or water cooling (not profitable), do copious testing (more profitable to just get reasonable settings and let the card run), and will post what they learn online (not profitable as it increases the network hashrate).

I think you can still get the Sapphire HD5850 Xtremes (I think they were released even after some of the 6xxx family!)  They cost about 130 GBP each.  I can't vouch for their power in gaming but they certainly fare well with the 6xxx family for mining.  Indeed, I believe the 5970 is the fastest mining card (I'm fairly sure it's capable of 1GH/s underwater and I don't think the 6990 can do that) so the 5xxx family is doing something right.

I see what you're saying.  I don't want to mislead people into thinking that overclocking to the extreme is a good idea financially and I agree that many people would be interested primarily in profit given the tools (not all but most).  Maximising profit is far from easy but my feeling is that a mild undervolt is best in this regard (accounting for the high value of BTC compared to power costs for GPU miners and the cost/depreciation of the cards themselves).

I personally rather like using mining for benchmarking cards simply because it shows off the OpenCL power that the cards have.  In the long term I will want to use my cards for tasks such as Folding@Home, Mandelbrot fractal generation, Strong Go/Weiqi/Baduk engine, and am not interested in 3D acceleration, video acceleration, desktop effects, or GUI at all.  Mining is a good benchmark for the former because it is a test of low memory/bandwidth raw OpenCL processing which I can run without a GUI and supports a rather cool crypto currency.  Naturally, for Folding@Home high efficiency is most important and I'll get to that but for Mandelbrot and Go it's all about performance.

Perhaps, it would be good to see what tricks people can play to lower their temperature delta.  However, this thread seems to be more along the lines of what temperatures are reasonable for 24-7 use of various cards.

Alright, I'll say this one more time.

Running your cards at high voltages or temperatures for extended periods will likely cause the cards to die prematurely.  These records should be taken as a measure of potential and can be useful for comparison but for most circumstances will not be profitable for long periods (considering power costs, card depriciation, or both).

If you are concerned only with maximum profit then this thread is not for you (unless you have no power or hardware costs and limited time) but you might hope to reap the benefits of the work done here in the long run.

When my power meter finally arrives I'll be undervolting my cards and looking for high MH/J scores then.  I'll make a thread for this when I'm ready and perhaps you'll find it more useful but, just as here, it will be for research and NOT recommended for maximum profit.

I do not deny the diminishing returns of increasing voltage but I don't understand why increasing voltage will increase power consumption exponentially and perhaps you can explain why?

If you increase the voltage (assuming no change in electrical resistance of the cards circuitry) then the current will increase in proportion (Voltage = Current * Resistance) and therefore the power consumption will increase quadratically (Power = Voltage * Current), not exponentially!  In real life the increased current will increases the temperature of the circuitry and therefore the electrical resistance so the growth of power with respect to voltage will actually be a little slower than quadratic.  Increasing clock rate will increase power consumption approximately linearly as far as I understand so if you increase both voltage and clock rate by 10% you might draw around 23% more power.

I simply cannot see how voltage and power can have an exponential relationship.

No, I'm talking about a 5850, in particular a Sapphire HD5850 Xtreme.  This particular card is built with the 40nm process and looks like a 6950 but has the essential structure of a standard 5850.  It will go to 1030 MHz on stock voltage provided I keep it cool enough and will run happily at that speed for days at a time without freezing.  To get more than 1030 MHz I had to increase the voltage and the increase from stock (1.0875V) to 1.25V was sufficient to make 1110 semi-stable (it only lasted 3 hours).  Indeed, the card went to 1130 MHz but was not stable (lasted only 90 seconds).

I'm sure the low temperature (about 55*C) and lack of a GUI made a difference to my maximum stable overclock.  In particular, I can't tell you whether or not mouse movement would have crashed the card because there was no mouse to move.  I'm sure if I was mining in a GUI environment I'd have to drop 50 MHz off of my clock to get stability, and another 40 MHz if the GPU was over 70*C.

I posted on the 5870 thread rather than here because they were interested in maximum speed whereas here we were more interested at maximum speed at 1GHz core.

Without changing the voltage it's hard to get a 5850 to 1GHz.  The most important factors seem to be system stability (don't drive a GUI and mine with the same card) and low temperature (certainly less than 60*C).

Just incase your inability to increase the voltage on your cards is a software issue, I'm able to change the voltage by using Catalyst 11.6 and AMDOverdriveCtrl (editing the .odrv profile for higher voltage, a GUI slider may not go beyond 1.0875V).  This is on Linux but I can't imagine it's harder on Windows.  No BIOS flashing or Trixx should be required.  Older versions of Catalyst will suffice if you only want to decrease your voltage.  I cannot make my voltage increase beyond 1.3V with this method.

The temperature where I am peaks at 21*C and when I leave my window open the temperatures of the cards drop to 64*C and 45*C respectively.  The conditions are very good for mining and I know I can push my cards harder (I took my good 5850 to 461.1 MH/s (1110 MHz) and it held it for about 3 hours at 56*C) but I'm more interested in achieving a near-silent miner.