Topic: SAVE money - undervolting 7970 XFX, self made VBE7 bios = TABLE OF CONTENTS = - page 7. (Read 56601 times)

 

Yeah but the problem is, for the power meter measurement to be somewhat accurate I need to count it over several hours. I'm not going to turn off my miners for several hours while they produce 100+ dollars per day 

A cool idea, to use the powermeter of the whole flat. :-)

Yes, 3400/12, more than 200 W per card - sounds more realistic.

If you want to be able subtract your baseline, just go to all 12 cards,
and (g)pu (d)isable, and wait until they have cooled down.

And then read the meter.

 

Me too. And did my measurements.

In the low earnings realm it doesn't help.
In the high earnings realm it does help.

What does that mean? Please wait for the explaining posting.
(EDIT: Posting #103 gives a summary)

Four conversations at the same time *stresss*  :-)


Yes, that seems to be the way for the high-khash realm of the optimization:
TheStilt (BIOS-rewrite) + drakoin (undervolting with VBE7) + psw (optimized skrypt kernel)


Super. Congratulations. Happy to hear that there is energy saved now :-)

Heh, not much green initiative going on here in Russia. When I visit my home in Belgium i'll try and find one.

Only 1 card is watercooled, the other 2 are not. The watercooled one is ~15W below the others.

I have 9x7970 and 3x290 and if I count the kWh being drawn by the whole flat (but no appliances on) it seems everything together is around 3400W which is quite a lot for only 12 cards. But of course there is PSU inefficiency and I am using 4 mainboards in total as well. And my 27" screen is on most of the day which probably draws close to 100W as well. But it's not a very accurate way to count, since it leaves me guessing at the non-mining part of total power use.

Quick update here.

First:
I went the TheStilt-route. The threat mentioned earlier here (https://litecointalk.org/index.php?topic=12369.0) is interesting for it explains why there seems to be a sweet spot for (not power optimised) mining with respect to Core/Mem-clock. Short version: card makers take shortcuts in the biosses wrt memory timing settings, TheStilt fixes those. This means no more hunting for that golden Core/Memory clock pair. Just stick the memclock tot 1500 (or any of the following:
400MHz (0-400MHz)
800MHz (401-800MHz)
900MHz (801-900MHz)
1000MHz (901-1000MHz)
1125MHz (1001-1125MHz)
1250MHz (1126-1250MHz)
1375MHz (1251-1375MHz)
1500MHz (1376-1500MHz)
1625MHz (1501-1625MHz)
1750MHz (1626-VCO Max)

and timing of the memory will be optimal. This greatly reduces the search for good hash rates.

With theStilt bios, I applied Dracoin's fix of the voltages, which in turn enables me to under volt the card.

With C:1054, ML1500, V1.065 I have already a (whopping;)) 30W drop in power from the wall, and the same hash rates. Actually the hash rates are better, because the card doesn't seem to throttle now, it is constantly at ~730Khs whereas it used to dip well below that now and then.

I haven't explored much further yet, but the stabilising of the hash rate is already a major win.

Currently I'm having 2.8kHs/W (up from 2.5) and I'll see where I can get to when I pull down the memory clocks to the above exact numbers.

Regards,

Hertog

.

you are one lucky owner of a good GPU card which can read it out itself. Very cool.
Alright, good to know. Thanks.

online then.

Wow. Yes, it would be really interesting to see how your calculated is related to the real power consumption. 175-180W seems low, but if you are watercooling ...

Idea: If there is something like a "save energy" initiative nearby (or the green party :-) ), perhaps they can lend you a kill-a-watt. If it turns out to report the same values, you don't need to own one, anyways.

 

I have one ASUS 7970 which doesn't have these sensors available as well. But all my other cards (reference 7970) have them. I also had a Sapphire 7970 once with non-ref blue PCB which also gave me the info. Only the ASUS didn't.

I'd like to compare actual readings at the wall with what the GPU reports, but so far no-one has shown up who can see both





Yes, it's completely unlocked. I just mean... 1.025V is what you SET. Doesn't matter if you set it in trixx, MSI AB or the (modded) BIOS. You get less volts than what you set.




As far as I can tell there are more 7970s that can read it than those who can't, but I can't find a kill-a-watt around here in stores


I think so, yes. If I add up all VRM inputs from HWinfo64, I get around 175-180W input per card for 800 Kh. What they really pull, I don't know. But my Corsair 850W PSU does not mind to run 3 of these cards plus my overclocked 4770k, 4 HDDs, 10 fans and a D5 water pump at all.

You got me really excited there.  But unfortunately, it seems as if your cards have current / power measurement, mine have not. I would have noticed in GPU-z if there were a measurement. I tried now hwinfo64, but the same here:



What a pity, it would have been a brilliant option for a new generation of self-regulating mining programs, if all the GPUs reported back their power consumption.


That is strange.  A bit like my experience before I unlocked the voltage by flushing the bios. I chose a voltage in cgminer, but it did not go to that value.
Are you sure your voltage control is unlocked at all on that card?



But you are one lucky owner of a good GPU card which can read it out itself. Very cool.

Still, get a kill-a-watt if you really want to know. They are cheap.


Nice one. 805/150 = 5.37.  But that's only part of the card's consumption, right?

Thanks for sharing!

Yes, that's the way.
Flush a voltage X into the BIOS lower than the given stock voltage Y
(X
Good to hear it went well.  Happy optimizing!

Same here, I enjoyed reading your page, thanks!
Interesting the approach you took in the virtualization, good solution just flushing the BIOS with your wanted values.

BUT you took a BIOS for a card of a different vendor, didn't you?  Gigabyte on XFX?
IMHO, you don't need to go that far, just tweak your given XFX bios.


The tool you are suggesting "Radeon HD77xx/78xx/79xx BIOS Editor" by Dragonheart
answers in my case "Your card is non reference, voltage control is impossible".
But it is.  I  suggest to use VBE7 instead. https://duckduckgo.com/?q=vbe7+bios


Perhaps something in our thread here might actually help you with it, if you try again?

When crashing?  Booting?  Cgminer?    Raise the voltage a bit.

1) With a high voltage, find your CC/MC clockspeed optimum.
2) Then slowly lower the voltage until crash. Then reboot, and
3) use a voltage circa 0.025V or 0.035V above the crash threshold.

HTH
:-)

I have 5.1 euros per day for electricity (slightly less, but I rounded it up), I have 2269kH/s which nets me on average per 24h 17.8 euro/day with current ratings (average over last 17 days rounded down and not taking in to account that sometimes 1 of my machines wasn't mining meaning profits are actually higher). As you are so friendy and helpful i'll pm you the details of my pool, but it's no real big secret tbh...

I don't fully understand what you mean with "Measurement for 5 different settings". But I want to add, if electricity costs you as much as 40% of your earnings, that makes your earnings 2.5x higher than what your electricity costs, meaning that if u lower your best total hashrate by 10% to underclock/volt you need to have more than a 25% TOTAL decrease in power drawings or you actually lose money by undervolting/clocking. Which still seems a lot imo. The guy who lowered his powerdraw so that he could have 2 280x cards on a 450W supply, what are his hashing speeds? Because that is what matters. I see all of you in this forum talking about kH/W, but that doesn't matter. You need to compare earnings vs spendings and nothing more. Meaning when you lower your total hashrate 10% your total earnings will lower 10%. For every euro you earn less per day, you have to spend more than a euro less per day or it's less profitable.
And when you do your calculations, make sure you use TOTAL system power, because only your card's powerdraw is useless. When you pay your bills, you have to pay for total powerdraw, not just your cards. So lowering powerdraw by 25% for every 10% hashpower lost means lowering total power draw by 25%.

Well I was on the beta drivers of AMD, and they apparently gave me problems or perhaps I had driver corruption I don't know rly. When I undervolted my card, it helped to draw less power, but only up to a certain point, after that lowering the voltages more didn't seem to do much. Thing is I installed the latest non-beta drivers now and now when I undervolt more it actually goes down some, not all that much, but still some.

I am currently running on 1.04V for 1040Mhz core and 1500Mhz ram. The system that I am undervolting draws 83W with the gpu disabled and 300W with the card hashing away.
I have to say though, I can't play games with that vcore. It crashes, hashing is fine, but games need at least 1.15V but that's probably due to the fact that in games it runs at its full default speed of 1100Mhz instead of 1050MHz that I use for hashing (If I increase the MHz above 1050 for hashing my hashrate actually drops, same for the memory).

GPU-z, shows VRM input current (and input voltage) but it is separate for GPU and memory VRM so you need to add them up.
Also HWinfo64 will do, this one actually calculates power OUT based on VRM output current & voltage but we want to know input power which it does not calculate. So we still need to check what the input current is and multiply with the input voltage.

Curiously, input voltage is different a little from card to card, even though they are hooked to the same PSU.



Well 1.025V is the voltage setting. Either in afterburner or the card's BIOS. The actual voltage after Vdroop is 0.965. Again can be seen in GPU-z or HWinfo.

Here is a screenshot with HWinfo.




We see 9.4A @ 11.65V for the GPU and 3.875A at 11.47V for the memory. 109.5 + 44.5 = 154W input.
I've been using these readings since the beginning to optimize power consumption, since I don't have a kill-a-watt.


I've done the same for my R9 290, unfortunately that one is a bit unclear because the second VRM (which used to be memory) is showing low power usage but GPU VRM is showing relatively high usage, as if some of the memory is accounted for in the first one. Regardless, I've been able to tune the main power consuming VRM down to less than 150W input at 805 Kh/s. At default settings (833 kh/s) it was pulling around 230W from that VRM bank.

I enjoyed reading this thread, and found it interesting as I experienced many of the same trials, which I wrote about here:

http://zsprawl.com/iOS/2013/12/mining-bitcoins-in-esxi-using-an-xfx-7970/

Although I couldn't get undervolting to work at all. It just kept crashing on me.

@Dracoin, quick question, will this flashing make it possible to change the voltage through cgminer? You seem to imply such with statig that you flash once, and then (dynamically) alter voltages to find an optimum.

If not through cgminer, could you elaborate how this is done?

Thanks.

Edit:

Yes, via cgminer. It is finally able to change the voltages

.
* *  * * Intermission Announcement commercial * ** * **


.

The 21st of January was a very very important day.
In many aspects, but I want to mention one, which
MAKES ME VERY HAPPY :-) and I want to share it:

 I got tipped for the first time for my work here.

  YIPPIEH *Dance around the room* YIEHAH!

                                

... and how could it be different, it's 2014 now:
Wasn't bitcoins, no - I got tipped with DOGEcoins!

Thank you, community! Thank you cryptocoin world.


Please

     tip me if you can:

       BTC    1EyyjBMMHjMfx6M3Ngu4sn5M4QJ6HAtWFG
     [AUR]  AevVj2wQSsM2TeAn8oACZY51fzLU7JYSF3
     [BC]   BPz72JMFkw4ymWSDe2Cdcp5mBZKSXBuk9i
     [C2]   SAYYYy9e92xGVuWdgCi995R1kywxmkvBWn
     [DGB]  DPxNvaRZLoeQGZpAmRRae2AdCLWNiQnzqm
     [DOGE] DLE2HhboCsx5JbeLhRAGW7TQJze41dCgiK
     [EAC]  eaoPAMndxjTkvvFUUq5SuSxMVg6KGoEQYW
     [LTC]  LQt3Wky3b1v7xEaatVYw25n22hujKMa2yq
     [MAX]  mKyVKaKoRetNnEVZLoouogyS7iz7qj74uu
     [MINT] Mfxyu7TRCDxco5sGn8EV1hNqF2A7Ut9oo4
     [NOBL] 9XJfp3mUrrmmdus6HWPtmFftoEZ7vBeYfA
     [PRT]  PnKS4C8euVMzN4YZXEkYwyUXxpdGugsmQm
     [QRK]  QhfXtCetfQE6pnup59MW45au65vcMZYK88
     [VTC]  Vx2iYor2ir1Pq9QTALs7TRdE5kmW9aABD6
     ...then PM me so that I can THANK YOU  ;-)




Or let you rig run cgminer --scrypt a bit for me. Thx.


* *  * * Intermission Announcement commercial * ** * **

.

About the same here. 25.6 eurocent = 0.35$. And in the past weeks, there were days with 4 euros per day worth of coins with my >900 khash/s.

Thanks for your doubts. You are right, it was already a lot of effort, and we have to be 100% sure that it's worth it, and choose the most profitable, and still most ecological way.

You inspired me to redo all of the calculations.  See below, another posting.


3 times as much as electricity costs ... would mean 8.50 euros / MH / day  reliably, and on average.

Tell me your secrets, please.  Where do you mine??  Perhaps tell me by PM.


Yes. And the instructions, and tools are compiled in this thread. Not perfectly organized yet, but it's all here :-)

You can see yourself in that posting below. Measurements for 5 different settings.

Very good. That's the way forward!  If green energy coins could only be made more valuable somehow ...

Imagine the contrast - that just for this coin mania another Fukushima would be built ...

Congratulations. Now that you mention it, please explain the driver before and after situation to us, as it is power-related - perhaps it can be useful for someone else?



Worth it. I like the sound of that.

 

Whew, and you tell me that now *g*  :-)

What cool news, I didn't know that the machine knows it's own consumption.
I always thought it should know, but I didn't know that it does know :-)

How can I read out those amp sensors?

and even if they are totally off, as long as they are a monotonic function (more in means more out) they would give us an indication if the power consumption is getting better or worse.

WOW!

IMAGINE OUR MINING PROGRAMS KNOW HOW MUCH THEY EAT,
THEN THEY CAN AUTOMATICALLY OPTIMIZE THEMSELVES!

When this idea is going to get implemented, then I really want get tipped :-) from you, and
you, and yes, from you also *g*  Otherwise I do claim intellectual property rights on that idea.
 


Yes, here too. When I open a window, then the cold winter chill changes the whole atmosphere, sound, intensity, ...


What do you mean by "with 1.025V set 0.965 actual" ?

and:
"showing very very low in GPU-z. 135-140W input (DC)"
?

My GPU-z doesn't show any Watts :-(

Please tell me that my card still knows, and that we can find a way to read it out. I have been staring at this damn powermeter for hours by now, guessing the average Watts in a fluctuating random walk ...

cheers.

De nada.

Yes, in the end, undervolting turned out to be really easy. And indispensable if you want to save energy.
But in the beginning, I was as careful as you. Lots of reading and asking. That's weeks behind me now 

A quick start is here in posting #37.

That would be lovely, thanks. Please measure especially also your "before"-values.

Oh, and keep an eye on this thread. Quacko inspired me to redo all calculations. More soon.