Antminer S5 - Underclock - Undervolt - Best J/GH

[elrippo]

Problem with those & most 12V PSU's is that the adjustment range is too small, they only go down to 10.8V. Having 2 x 5V supplies give me the adjustment range I need, also the ones I bought were very chap compared to the ones you have linked to.

Rich

Agreed, but you can rate the Output current to get to your desired wattage.
I did some test, i have to post the pictures, with an S1
If i lower the Voltage the current will go up to maintaine the power needed, so if you drop the current you will also lower the wattage.

If i pull 12V and 12A i get 144Watt
If i pull 10V and 14,4A i also get 144Watt -> that´s what my S1 did
So if i pull 10,8V and 13,3A i get 144Watt or lower...

If this is not feasible, forget what i just wrote.

[RichBC]

Sorry I don't understand what you are suggesting here? The spec sheet for those PSU's says 10.8V to 13.5V. Wheras the 5V PSU's I am using have a spec of 4.3V to 5.8V, so 2 in series give 8.6V to 11.6V which is perfect for undervolting.

Rich

[elrippo]

Sorry I don't understand what you are suggesting here? The spec sheet for those PSU's says 10.8V to 13.5V. Wheras the 5V PSU's I am using have a spec of 4.3V to 5.8V, so 2 in series give 8.6V to 11.6V which is perfect for undervolting.

Rich

If i lower the voltage the current will go up to meet the consumption needed by the device.
So you can limit either the voltage or the current or both.

So i tested this on a S1 laying arround. If i give the S1 12V it draws 12,xxxA. If i limit the Voltage to 9,5V the current goes up to 15,5A to meet the draw of 144W by the device.
So if you limit the voltage to 10,8V the device is drawing 13,3A to meet the 144W needed, so you can adjust the draw by Volt, by Amps or both.
That´s why i opposed these PSU´s
If i am electrically completely wrong, please ignore my post 

[RichBC]

I still can't make any sense of what you are saying. The fundamental problem with the PSU's you have linked to is that the voltage will not adjust below 10.8V wheras I want to get down to 9V, simple as that...

Rich

[elrippo]

I still can't make any sense of what you are saying. The fundamental problem with the PSU's you have linked to is that the voltage will not adjust below 10.8V wheras I want to get down to 9V, simple as that...

Rich

Ok, i am fine with that 

[RichBC]

OK good. You also need to remember that you are referring to an S1, which uses DC-DC Buck converters to reduce the 12V input to the Core voltage needed by the BM1380 chips. If you reduce the 12V supply going to the converters, as you say, they will just take more current to compensate. Not much point in doing this as they then tend to be less efficient and take more power.

The S5 is completely different and does not use Buck converters. The BM1384 chips are connected in a string (Like Xmas tree lights) and the voltage for the core of the chips is reduced to what is needed by the series connection. 12V / 15 devices = 0.8V. So now if you reduce the 12V supply the voltage to each chip also reduces. 9V / 15 = 0.6V. Combine that with a frequency reduction selected for an acceptable HW error rate and you get the table of Hash & efficiencies I have shown.

I have made some measurements with the Meanwell 5V PSU's in series, here are the results.

Code:

MHz	V Core	V	A	W	GH	J/GH
200	0.69	10.35		260	665	0.391
175	0.66	9.9		210	575	0.365
150	0.65	9.7		175	500	0.350
125	0.64	9.6		147	415	0.354
100	0.63	9.45		119	330	0.361

The efficiency should be better at 100 & 125Mhz but is being held back by the less good Hash board, that needs a higher voltage, particularily as you get near 9.5V. I will make some measurements on the single better hash board at some point. However not bad.  

Rich

[elrippo]

OK good. You also need to remember that you are referring to an S1, which uses DC-DC Buck converters to reduce the 12V input to the Core voltage needed by the BM1380 chips. If you reduce the 12V supply going to the converters, as you say, they will just take more current to compensate. Not much point in doing this as they then tend to be less efficient and take more power.

The S5 is completely different and does not use Buck converters. The BM1384 chips are connected in a string (Like Xmas tree lights) and the voltage for the core of the chips is reduced to what is needed by the series connection. 12V / 15 devices = 0.8V. So now if you reduce the 12V supply the voltage to each chip also reduces. 9V / 15 = 0.6V. Combine that with a frequency reduction selected for an acceptable HW error rate and you get the table of Hash & efficiencies I have shown.

I have made some measurements with the Meanwell 5V PSU's in series, here are the results.

Code:

MHz	V Core	V	A	W	GH	J/GH
200	0.69	10.35		260	665	0.391
175	0.66	9.9		210	575	0.365
150	0.65	9.7		175	500	0.350
125	0.64	9.6		147	415	0.354
100	0.63	9.45		119	330	0.361

The efficiency should be better at 100 & 125Mhz but is being held back by the less good Hash board, that needs a higher voltage, particularily as you get near 9.5V. I will make some measurements on the single better hash board at some point. However not bad.  

Rich

Ok, got it 
I found one with 89% and an adjustable VoltPoti 

[RichBC]

Yes quite a nice supply except that it only goes from 6V to 9V which is not a lot of use.

A word of warning for anyone who has not read some of the earlier posts in the thread. People had tried undervolting the S5 from early on but had little success as it stopped hashing if you reduced the voltage much below 11V. The two S5 that I have are both V1.91 Hash Boards which have 4 oscillators spread round the board as opposed to one on all the earlier revisions.

It is highly possible but not proven that it is these additional oscillators on the V1.91 Hash board that enable undervolting.

Rich

[elrippo]

Yes quite a nice supply except that it only goes from 6V to 9V which is not a lot of use.

A word of warning for anyone who has not read some of the earlier posts in the thread. People had tried undervolting the S5 from early on but had little success as it stopped hashing if you reduced the voltage much below 11V. The two S5 that I have are both V1.91 Hash Boards which have 4 oscillators spread round the board as opposed to one on all the earlier revisions.

It is highly possible but not proven that it is these additional oscillators on the V1.91 Hash board that enable undervolting.

Rich

OK good. You also need to remember that you are referring to an S1, which uses DC-DC Buck converters to reduce the 12V input to the Core voltage needed by the BM1380 chips. If you reduce the 12V supply going to the converters, as you say, they will just take more current to compensate. Not much point in doing this as they then tend to be less efficient and take more power.

The S5 is completely different and does not use Buck converters. The BM1384 chips are connected in a string (Like Xmas tree lights) and the voltage for the core of the chips is reduced to what is needed by the series connection. 12V / 15 devices = 0.8V. So now if you reduce the 12V supply the voltage to each chip also reduces. 9V / 15 = 0.6V. Combine that with a frequency reduction selected for an acceptable HW error rate and you get the table of Hash & efficiencies I have shown.

I have made some measurements with the Meanwell 5V PSU's in series, here are the results.

Code:

MHz	V Core	V	A	W	GH	J/GH
200	0.69	10.35		260	665	0.391
175	0.66	9.9		210	575	0.365
150	0.65	9.7		175	500	0.350
125	0.64	9.6		147	415	0.354
100	0.63	9.45		119	330	0.361

The efficiency should be better at 100 & 125Mhz but is being held back by the less good Hash board, that needs a higher voltage, particularily as you get near 9.5V. I will make some measurements on the single better hash board at some point. However not bad.  

Rich

Ok, got it  
I found one with 89% and an adjustable VoltPoti  

I´ll talk to that guy if the Voltage goes up to 9,4 or 9,5. As i understood you were searching for a high efficient 9V power supply. I´ll let you know what the tolerances of these PSU´s are  

EDIT: Talked to the guy. He said, "usually" these PSU have a slight range over the stated one. It could be that this one goes up to 9,4 but that would be critical to find out.

[RichBC]

You are making this hard work, this is what I said earlier.

the 5V PSU's I am using have a spec of 4.3V to 5.8V, so 2 in series give 8.6V to 11.6V which is perfect for undervolting.

It's the ability to adjust the voltage in this range dependant on Difficulty & Electricity Price that is important. At any point in time the optimum setting for Voltage & Frequency changes and will change substantially when we have the halving next year.

Rich

[elrippo]

You are making this hard work, this is what I said earlier.

the 5V PSU's I am using have a spec of 4.3V to 5.8V, so 2 in series give 8.6V to 11.6V which is perfect for undervolting.

It's the ability to adjust the voltage in this range dependant on Difficulty & Electricity Price that is important. At any point in time the optimum setting for Voltage & Frequency changes and will change substantially when we have the halving next year.

Rich

I know what you said, i just was looking for one PSU in this range but i think your solution is the one to go 

[VirosaGITS]

I am curious as to what happen to the miner when ran at over 350 to 400mhz undervolted. It seem like the test results so far were only extended to up to stock and then downward. I'M guessing the lower voltage you go to, the upper(stable) clock limit goes down?

[RichBC]

It's very straightforward. The higher the Frequency the Higher the voltage needed to not get excessive HW errors. The Bitmain design point is for 12V and 350MHZ and they have been slightly conservative with the voltage such that even a worse case miner will have a low HW error rate.

Because of that most S5 running at 12V can be overclocked to 400MHz with an acceptable HW error rate. Conversely you can undervolt (11V) most S5 running at 350MHz and still get an acceptable HW error rate along with a J/GH efficiency improvement.

Rich

[VirosaGITS]

It's very straightforward. The higher the Frequency the Higher the voltage needed to not get excessive HW errors. The Bitmain design point is for 12V and 350MHZ and they have been slightly conservative with the voltage such that even a worse case miner will have a low HW error rate.

Because of that most S5 running at 12V can be overclocked to 400MHz with an acceptable HW error rate. Conversely you can undervolt (11V) most S5 running at 350MHz and still get an acceptable HW error rate along with a J/GH efficiency improvement.

Rich

Thought so, thanks. I'd be more looking in the vicinity of something like 11.6V for 375Mhz. And eventually downclock/downvolt them to a usage of 450w for 1~THs. That would allow me to power 3 Antminer S5 from 1 120v 15a breaker

[SLEI]

I got an S5 motherboard V1.3 and assembled a parallel DPS-600 power for it.
S5 didn't hash at 10,5V and I adjusted to 11V, I have now tested upto 300Mhz and got about 1000Gh/s with 0,004 error rate.
Family is sleeping more testing at daytime, fan speed was 2700 and 445W from wall with very small router at same plug.

With one DPS-600, 300 Mhz and unadjusted power with 12,4V idle voltage this same S5 took 550W from wall.

When the V1.9 S5s are made/sold?

[RichBC]

@SLEI that result is what I have seen from the early testing that has been done on the S5 with not much less than 11V being possible, however still a nice gain. Would be interesting to see if you can push the frequency a bit more at 11V?

I do not know when the V1.91 was sold. All I do know is that it's the latest board I have seen and the date on the PCB is 28-March-2015 so some time after that?

I have listed all the boards I can find along with a few notes in this post in the thread.

https://bitcointalk.org/index.php?topic=1151460.msg12164839#msg12164839

As I have said in that post my best theory on why the V1.91 will undervolt is the multiple oscillators, however that is still unproven , but is the only significant difference. Oscillators could probably be retrofitted to earlier boards but would be a tricky job.


Rich

[QuintLeo]

"Ran some tests today on the DPS-600 PSU with a voltage adjust pot.(-12V to the Voltage input pin, Pins 5 to 9)  Bottom line not much use with an S5. "

One is not enought but two is. Quick search from net gave me a price of over 50$ per Mean Well 5V supply and You need at least two of them. Efficience is about same and You can get pair of DPS-600s with wires at 40$.   
17$ per DPS-600 shipped; RC servo extensions for on-off pins, trimpots and current share; molex-PCI-E connectors for power feed. 

DPS-600 is easy to hack You don't need to open it.

 Better efficiency on the Meanwells, if you grab the right ones.
 Cost vs. efficiency is usually a tradeoff in any power supply situation.

So i tested this on a S1 laying arround.

 S1 has onboard voltage regulation, no need to use an external "undervolt capable" power supply, and defeats the purpose anyway.
 S5 is a string design with NO onboard voltage regulation, you HAVE to undervolt them externally.


 While the 12V Meanwells usually spec to go down to 10.8, they'll usually go a LITTLE lower - not to 9V though.

[SLEI]

S5 V 1.3 seems to be very moody, no start when cold at morning 300Mhz , volts were 10,9.

11,3V were enought for std 350Mhz hashspeed 1100 530W, fan 3700 and loud as std.

Now early at evening 10,7V is enought for 293,75 Mhz 930Gh/s and 415W 2400rpm fan.

[RichBC]

I guess if you are near the limit of something just working or not, then you only need to change another variable like temperature a little and that's enough to throw it either way. Another factor is that the voltage needed to start hashing is slightly higher than the voltage needed to keep hashing. Anytime you have a system that is hashing if you have adjusted the voltage down since it started you need to reboot to check if it will actually start at that voltage.

I have been making measurements on individual S5 Hash boards and there is some variance between them, will post up the results when I have finished.

Rich

[philipma1957]

S5 V 1.3 seems to be very moody, no start when cold at morning 300Mhz , volts were 10,9.

11,3V were enought for std 350Mhz hashspeed 1100 530W, fan 3700 and loud as std.

Now early at evening 10,7V is enought for 293,75 Mhz 930Gh/s and 415W 2400rpm fan.

the 5 volt meanwells are a bad choice.  all of them  are not efficient.


 this one is good  my goal is 2 of these at 10.2 volts or 10.5 volts  running an s-7

if I can drop from 12 to 10.2  that is 15%      since no plat psu does better the 94%  this one is 88%

   so I am better then 11% power improvement. over a high end plat

http://www.trcelectronics.com/View/Mean-Well/HRP-600-12.shtml

https://www.trcelectronics.com/ecomm/pdf/hrp600.pdf