Bitcoin Forum

Finally made the plunge on a Killawatt....what fun!

I don't know if this is unprecedented, but I was able to overclock AND undervolt 2 of my 7970's and drop the power consumption by a huge amount.

my settings:

Core voltage:  1060
Clock:  1125
Mem:  685

This was pretty stunning to me, but I am getting 1300+ MH/s @ 300+- watts.

4.33 MH/watt.

Now I realize this is not impressive when compared to massively undevolted and underclocked GPU's, but this is a massively OVERCLOCKED set-up!  As a bonus the temps droppe a good 5 degrees with the same fan speed.

Just throwing it out there.....I was pretty thrilled.  I will report back if it stays stable for 24 hours. 

 

That is pretty impressive work yochdog!!  :)   I can't wait to hear if remains stable,  gives the 5970 a run for its money as top dog.  ;D

That is a pretty good spot to hash at, are you air cooling? It will most likely be stable at that speed. Anything higher than 1125 MHz will increase wattage significantly, voltage will only need to be changed at ~1240MHz (~200 watts).  How did you decrease RAM clock, was it through driver?

Yes, air cooling with a small external fan blowing on them as well.

I decreased Mem clock using Afterburner beta 12.....it is the only tool I can get to clock it that low. 

Huh that is interesting, I could never get RAM below 1100, is 685 what Afterburner shows, does GPU-Z sensor readings indicate that same speed?

Something is most definitely wrong with your meter or your math, my friend.  At that clock/voltage, you should be drawing a bit over 200 watts per card (at the wall).  I have three 7970 rigs and they all draw more power at a lower core clock, lower voltage, and a lower memory clock.  To top it off, they don't have hard drives either.  Hell, my water cooled 7970 rig would draw more power than this.  In fact...  :D

I tried to get as close to that clock/voltage as possible (my cards aren't stable at that core/voltage combo).  With your same settings but 25 Mhz lower on the core clock, I've got 512 Watts total system power consumption with 2 cards.  Granted, this is on a machine with an i7 930, but this measurement was done with no CPU overclocking.  All components (including the GPUs) are water cooled.

Even if you're just trying to calculate the power draw of the GPUs themselves, your measurement is off somewhere.  My pump draws 9 watts.  My idle CPU, board, SSD, and a few power controlled fans aren't going to suck up 212 watts.

Hmm, maybe it is 2 separate rigs with 1 7970 each?

The oldest beta that was out when the 7970s were released would clock the memory lower.  They put a date limit in the code though, so it doesn't run anymore.  The newest beta will take the memclock lower than 685 when it's first installed (while it complains about needing a restart).  After a restart though, 685 is the lowest.  A BIOS flash is the surefire way to go.

I ran a WC 7970 and a 6850 together, both pull close to 500 watts under a heavy overclock, 300 watts is too low to be true for two 7970's. Sensor readings on 7970 show core around 210 watts with memory around 30 watts; thus actual cards are pulling in well over 200 watts. I get around 3.75 MH/W.

hmmmm, I will slap the Killawatt back on it and re-check the readings.

Perhaps I screwed something up, but I thought I was right on. 

Actually, I'm starting to wonder myself now.  Looking at the "VDDC In" stat in GPU-Z is showing a value closer to what you measured.  Because of AC/DC conversion in the PSU, we can tack on some extra Watts due to lost efficiency.

I have a 4x FC HBA in this machine right now and I forgot about it because I was only using it to test something the other day...  Apparently this thing is using a fairly absurd amount of power?  That aside, it doesn't really explain what's going on with my dedicated rig.

More testing.  BRB.  :D

*sigh*......

Alright, I went and re-measured the numbers.  I think what happened was I screwed up the idle wattage the first time around.  Here are the correct numbers:

System idle:  154
Mining with above settings:  512
Hash rate:  1340 MH/s

Soooo, I am getting 3.74 MH/watt.

I am bummed I was not getting what I thought....but still thrilled to shave subsantial power usage.  

I think I can get close to 4.00 MH/watt, but the system would have to be a dedicated miner.  Using other applications causes instability at lower voltages than the 1.08,

so like 180 watts a card. not bad.

Even at close to 4Mh/W it is real impressive.  :)  Nice

No. It is nonsense to compute the efficiency as 1340/(512-154) = 3.74 Mh/J...
For example, if the cards' idle power consumption would be worse and would be making the machine idle at 350 Watt, would you conclude it would raise efficiency to 1340/(512-350) = 8.27 Mh/J ? Of course not!

You are in fact getting 1340/512 = 2.62 Mh/J which is in line with what I have measured on my own HD 7970 machines.

It's not exactly nonsense, if he is trying to measure the efficiency of his cards as opposed to the efficiency of his entire system.  Now sure, to be perfectly fair he should remove the idle power usage of the video cards before subtracting the idle from load, but 7970 use very very little power in "long idle", less than 3W per card typically.

Also, he can hypothetically run more video cards on this system without impacting the idle power usage (other than the minor draw of an idle video card).  It's good to know the per-card efficiency to know if it would be worth doing so.

Indeed, the card's efficiency is also an interesting number, although a bit difficult to tell from a one measurement of idle/mining. The card's idle consumption may vary depending on how well the idle-features are working. The second card is probably in that negligible "long idle" state, but the primary may be consuming quite a bit if he is for example using two monitors or possibly overclocking with AB unofficial overclocking. At least that's the case for me, my 7970 fails to drop to 0.85V & 300/150 MHz often.

I suppose a better way to find out a card's efficiency in this case would be to compare the power draw when mining with primary card to the power draw when mining with both of them, assuming long idle is working well.

But yes, it's an interesting number, in fact probably more interesting to others than the efficiency of a quite power-hungry system. A dedicated mining rig with low power CPU and such with 4-6 7970s will make the cards' consumption dominant, instead of the system's consumption.

All I am trying to measure is the additional power consumption of the cards.  This is a machine that is always on, and has been always on for the last 3 years.  Thus, the idle power consumption is the baseline, as it would be consuming that regardless.  Determining the incremental energy consumption is all I am concerned about, and now have a good approximation of that number.

Now if I was building a dedicated mining rig from the ground up, of course the over-all system wattage would be the important number.  But in this case I am not dealing with a dedicated miner, thus it is not nonsense.

Another big factor people seem to ignore is how much power is drawn once the system has been mining for a while.   I get the sense most fire up their miner, wait maybe a minute, then check the kill a watt.  You really need to let the miner go for an hour or more to see the full effect.

My rigs use almost 10% more juice after getting up to temp (running stable for days) vs the first several minutes or even first hour of mining.

I also watch the amps more than the watts, since I don't want breakers tripping on unattended miners.   Our voltage tends to sag to below 110 during the day from a high of 118-120 each night.

What is your total $/MH (excluding electricity)?  3 sapphires + mobo, etc.

That's what I was wondering. $2 to mine 1 BTC is at current difficulty with your electricity rates.  I was more curious about the hashes as that is less specific to time and place.

So its (542 * 3 + 160 + 330) / 1986 = 1.065 $/MHs

Well whatever you run the system 24/7 at is what matters.  What do you run it at?

And this is precisely my point! This makes your numbers meaningful only to you, and meaningless to all of us on the forum (because most people here favor more efficient systems -- and a dedicated 2 x 7970 miner idles at 90W or so). You should have disclosed your unusually high baseline idle load in your first post.

Meaningless?  Are you serious?  Being able to compare apples to apples the power consumption of an overvlocked/undervolted 7970 to a 5970, ot 5870, or any other number of cards is meaningless?  What about those with existing farm of dedicated miners?  Might they find some meaning in stand-alone efficiency of cards?  Might they find some meaningful cost savings if they can upgrade to more efficieint GPU's?   

Fail on so many levels. 

Let me correct:

• 3.74 Mh/J is meaningless to us, because it is inflated by your inefficient baseline idle power of 150W (perhaps you have a high-power CPU, or your PSU's efficiency sharply drops at low loads, or you are running a graphics-intensive Windows Aero desktop, etc)
  
• 2.62 Mh/J is meaningful.
  

I disagree for exactly the same reasons.

2.62 Mh/J is meaningless as it's not comparing like-for-like, you're adding in the unknown variable of the system, which will be vastly different (as you noted) system to system.

3.74 Mh/J is meaningful, as other people, who wish to compare their values, can use this number by factoring out their own baseline system power.

I think both are useful.

One is nice for comparing cards, the other is nice for comparing systems.

I wanted to compare my 1.8GH rig to this 7970 rig, so overall power consumption was what I wanted.

Now if I wanted to look into swapping my 5970s for 7970s, then the per card consumption is more useful.

Your argument is invalid because 3.74 Mh/J is also influenced by unknown variables, such as the efficiency of power supplies which varies with load: http://www.anandtech.com/show/2624/3

Here is a thought experiment: yochdog's load/idle power draw is 512/154 Watt. He replaces his power supply with one that is just as efficient at high loads, but more efficient at low loads, changing his measuremnts to 512/130 Watt. Suddenly his mining efficiency went down from 1340/(512-154) = 3.74 Mh/J to 1340/(512-130) = 3.51 Mh/J ! Explain to me why using a formula in which efficiency becomes worse when using better hardware components is useful?

Of course, if everybody had clamp meters, the ultimate way to measure the efficiency of a card would be to measure current at the PCIe power connectors and PCIe slot, like I demonstrated a while ago: http://blog.zorinaq.com/?e=42

Holy bleeping shit. 

I try to put some useful (evidently not!) information out on the forum, and MRB comes along to piss all over it because anything but his prefered stats are meaningless.

Apologies to the forum.  I wasted everyones time with a useless post. 

From now on I will only post the total power consumption of my systems, and you are all on your own deciphering what is being used where.

GOOD LUCK!! 

+1.  People are going to have completely different motherboards, CPUs, hard drives (if any), power supplies, RAM, cooling, etc etc etc.  The most useful data is the power draw of the cards themselves.

We just need to get linux to allow downclocking memory beyond 150 Mhz  >:(

My 4x7970 rigs run about 1060 W (1040 is more correct with 2 Delta screamer fans not running).

2.5 Gh/s @ 1040 = So So...

Cards are running at 1050/900 and 625 Mh/s each...

Here is a fictitious system in idle/load conditions demonstrating my point that cards can draw more than what the kill-a-watt difference shows:

Idle: 130W at the wall, power supply outputs 95W (73% efficient) = 15W to the video card + 80W to the rest of the system
Load: 200W at the wall, power supply outputs 176W (88% efficient) = 96W to the video card + 80W to the rest of the system

yochdog would conclude from the kill-a-watt readings that the card under load is drawing at most an extra 70W (200-130), but in fact, it is drawing an extra 81W (96-15). This is why subtracting the idle from the load wattage at the wall is not as accurate as you all seem to think.

At the very least, if a kill-a-watt is all you have, I suggest you:
- publish your power supply specs to look up the energy efficiency curve at different loads
- measure "idle" condition with the card physically removed from the system

My first 7970 came with a number of issues while booting. I had it RMA'd, the new card that just arrived reaches 750 MH/s (1225MHz core, 1375 ram, 1.07 V, 168A) with sensor readings (HWiNFO64) indicate I am pulling 193 watts (Core + RAM).  My first card would already be over 220 watts at these settings, seems either software improved or hardware improved, I used the same drivers from the end of Feb, so probably hardware changed?

How accurate are the GPU/CPU sensor readings by GPU-Z or HWiNFO64?  Would using a multimeter carefully to measure PCIE and all other power connectors on system be more accurate than onboard sensors or KILL A WATT?

While I get your desire to have 100% accurate info, I don't fully agree that it is more useful.

1- virtually everyone is going to be running mining rigs on computer power supplies, which are never 100% efficient

2- knowing the actual real world power usage from the wall is very useful for calculating cost of mining, while knowing the exact power usage from the power supply is an interesting mental exercise it doesn't actually benefit anyone who is merely trying to do a cost/benefits analysis.

3- most miners use 80+ certified power supplies, if not 80+ silver or greater spec.  The difference between actual power usage and exact power usage is going to be less than you indicate.  For example, 80+ silver gets between 85% and 88% efficiency guaranteed, a 3% difference even on a 1000W scale is only 30W and not really a huge inaccuracy.

You forget that the 80 Plus standard does not guarantee efficiency below 20% load. Which is why I criticize idle measurements at the wall so much as they are likely to be below 20% load.

currently running 4x 7970s @ 925 / 150 @ 1.00v

850W at the wall. This system has the Big Bang Marshall and i7 2600k and 1300w CM Hybrid

I will OC them to 1125 and will let you guys know the power :D

close.

1125 / 150 @ 1.05V

1050W with 2600k @ 5.0Ghz

2700mh/sec

9.6 Amps

Not bad! I can have 2 of these rigs running in my dorm room without tripping the wire!

4 7970's at 1125 should be getting over 3300Mh

What... That would mean each 7970 would be getting 825mh... I think you meant 5...

This is probably just from leakage.  Each of the cards are going to have a GPU that is "binned" differently.  I've found on my rigs that as soon as I let the temps go much further than 68C, power consumption starts climbing pretty heavily (additional leakage from heat).  If I let the cards in one of my 4 card rigs run at 74-75C, power consumption goes up by almost 100W at the wall.  My 6970s/6870s/5870s/5970s don't do this until they get very hot (80C or above).

Your absolutely correct...  ooopps  sorry bout that.

My thoughts started with the PSU.  I have a 1250 seasonic pushing 4 of the cards, and then I seen the Mh number and thought he didn't have the updated code.

You mean 5.

I also have mine undervolted to run much cooler.

no i meant 4.   the rig has 5 cards, but the seasonic is only on 4 of them. I have another psu powering the 5th card of this mobo and a 5th card on another

10 7970's   2 seasonic 1250  1 seasonic 650

http://i1027.photobucket.com/albums/y335/jjiimm_64/10x7970%20rig/IMG_0536.jpg

This is interesting to note. So as processes get smaller and smaller, power usage will fluctuate more and more widely based on temperature. All the more reason to have good cooling.

Yep. Water FTW. About a 40 watt saving per card by going to water for my 7970's.

https://bitcointalk.org/index.php?topic=57410.msg683755#msg683755

It definitely seems that this is the case.  Even more interesting is the effect this causes when using fans.  If I let the fans go too far beyond the GPU's sweet spot, I actually have to use a similar or higher fan speed to keep the GPU *hotter* (because of the additional wattage).

You certainly hit a "point of no return" though where it costs more energy to spin the fans faster vs the energy you'd save from keeping the chip (and VRMs?) cooler.


The power savings are substantial, but it's certainly not 40 watts per card (unless you were running the cards very hot on air in the first place).  My water cooled 7970 machine saves about 27 watts (DC) per card (at 36-37c).  Keep in mind that some of the apparent power savings is simply from not having fans on the cards though.  Some of this power to remove the heat gets consumed by the pump and radiator fans.

They were running about 85c on air. So pretty warm. Went to about 40c on water. The fans do count as power savings in my case too. I already had the cpu watercooled and just add my gpu's to the same loop. No extra pumps, fans or anything.

Ah.  Yeah.  That's pretty hot.  :)  Dropping 40-50C will certainly save a good chunk of power.  :D