Topic: Antminer S5 - Underclock - Undervolt - Best J/GH - page 14. (Read 31123 times)

Yes there will be some voltages & signals that can be checked, but not quite ready for that yet as I am still feeling my way round the circuit. Measurements are complicated by the string design with no common ground, with the ground for each pair of BM1384 up the chain being VCC from the previous stage. Also means you cannot just isolate a stage which would be a way of tracking down a problem.

It could be as simple as poor heatsink compound on a chip, but I suspect it will not be that easy. I will fully measure the performance of the 4 hash boards I have, while thinking about how to find differences between the boards. If as I think is the case I have 2 very good hash boards I will put them into a single S5 Module and set it Hashing  while I investigate the other two.

I will get there and had expected this to be a tricky process, the surprise was a level of instant success. 

Rich

Very well done. I guess it is not possible to check the voltage on a chip by chip basis to see if there's some issue somewhere?

Have been busy this Morning, in amongst all the BM1385 excitement, and have made some progress on the 2nd S5 with V1.91 Hash Board, no heatsinks.

I disconnected in turn each of the Hash Boards and remade some of the measurements. The first board continued to exhibit the long voltage knee between a large number of hardware errors & no errors that I explained Yesterday.

The second board on it's own behaves very much like the first S5 with the heatsinks, in fact it had very good performance, the best I have seen so far. That said I had suspected that one of the boards in the first unit performed better, the figures reflect the performance of the lower performer,  but never checked it out.

So here are some sample measurements from the "good" hash board in the second unit.


Remember this is only 1/2 an S5 so Amps, Watts & GH are half previous tables, other figures can be directly compared.

So not quite up to BM1385 performance but at 150MHz & 9 Volts gives 248GH with a (record breaking) J/GH of 0.269 This was with no HW errors and also remember that the Fan & Controller board current is included even though it is driving a single Hash Board. Power is, as in all the other measurements, calculated  from the supply, not at the wall.

So now need to go and understand what is holding back the other Hash Board in the 2nd S5? First examination shows it to be a slightly different board. Both are 1.91 with the 5 Oscillator modules.

The First S5, with heatsinks, both boards are marked V1.91 - GYS0150 - 2015-03-05

The "good board in the second S5 is marked V1.91 - GYS0107 - 2015-03-28

The "less good" board in the second S5 is marked V1.91 - GYS0199 - 22-15

Whether any of the markings are significant I do not know?

I will have to set to and make some measurements, but my gut feel remains that there is a chip which is for some reason under-performing?


Rich

should do so since he was getting   200 watts and 665 gh on two boards.

My electrical knowledge is decent not great but I think since this all comes from a 1 rail server psu 2 bucks a rail should be okay.  

do not have to go there.  since I can send pull the fuse from each converter replace with a low amp tie both pos and both neg and run just a fan.

no magic smoke and we would be good to  go.




cooling fan allows for more power handling

My point was that the buck converter that Phillip mention was only capable of 15A. At 12V that would be 180W which fit well within the connector limits. IF however, you need more than 15A for the blade, then you would likely need to employ a 2nd buck, which would most naturally use the 2nd PCIe socket. If one buck can support an entire blade, then no issue.  

don't have an s-5 on hand I have coins on order to buy one and freaking bitmaintech sold out the s-5

I will cross that bridge when the coins come on friday from coinbase.

I became interested in this because at .35 watts I can run gear at my summer price rate of 16.9 cents.

All my gear was sold in May  2015 except for sidehack's usb sticks and 1 u2

Why not run a single board for now?

they will do the job I have used them before.  I used the 10 amp version and it pushed 9 volts and 93 watts 24/7 for weeks.  I know the 10 amp version works well and it did approach 95% efficiency as it was listed.

the 15 amp version is rated 98%  under ideal conditions..  basically dropping from 12 to 10 or 12 to 9 is not like dropping from 24 to 5 .

Also if you use a fan with them they will even work better.

Since I plan to go down to 9.5-10 volts  and use around 100 -105 watts a rail  I am 106 of 150 which is  under 80%   the safe  24/7/365 is to derate be 20%  so under 120 watts and I will be good to go.

Well that is the good new the bad news is amazon sent me the wrong order today  they sent 1 15 amp piece and a 3 amp piece.

So I am exchanging the  3 amp for a 15 amp piece.

Even if all this works out at 15 bucks a rail it cost too much

30 an s-5

and I think 18 x 15 or 270 for an s-5+  

for now I am dead in the water until I get the second one replaced.

I have this server to test with

http://www.gigampz.com/store/p21/Gigampz_80Plus_Platinum_Efficiency%2C_1200_Watt_Kit.html

All i can add to that 1 connector per board thing is that users report running with only 1 connector at stock settings without any problem, as its within the 300w spec per connector i believe.

At least i can say doggie is very vocal about this and if you take a look at the S1 or even S3 to an extent they run at 250w and 220w watt per connector. So an undervolted S5 as per the graph start at that margin and then only go lower.

So unless the connector has some weird mechanic where they don't work because A go up for V to go down, they should be fine since its ultimately lower watt than the max in any case.

On the buck converters, as VirosaGITS says and particularly as we are undervolting I think 1 per board would do. However I cannot see myself risking one an an S5 hash board. I also have my doubts if they will meet the efficiency claims. I agree that potentially they are a neat solution, but feel that I / We / Someone will need to find a way of reducing the voltage on a Server PSU, if we are not just going to throw away everything we have gained. As I said I have a couple of different ones here which I will look at when I get time.

The controller board takes it's power from one hash board, definitely some of that comes from the separate 4 pin connector. I am unsure if it also takes any voltages from the 18 Pin connector?

Rich

The second S5 arrived this Afternoon, there is good news not so good news.

Its's a V1.91 but without heatsinks, has the 5 clock modules fitted, but has a different silk screen on the PCB. I have not had the cover off yet to have a close look.

The good news is that it will boot at 125MHz & 9.25V like the first unit. The not so good news is that the HW error rate at those settings runs at about .02% whereas the first one was .0005%, although I did not log it. The error rate itself is however not the main difference. On the first unit there was a smooth and quite rapid transition from a lot of errors, through a few errors, to acceptable / no errors as you inched the voltage up. At the lower voltages it typically spanned 0.25V.

On the 2nd unit I have to increase from 9.25V to 10.0V to get to that acceptable/ error free point. The same characteristic applies at higher frequency / voltage but the additional voltage needed decreases. At 250Mhz it increases from 10.4V to 10.7V. I will look at it more carefully and make proper measurements in the Morning.

So I guess we should not be surprised there are variances, problem now is to find why and can anything be done about it? One thing I will do on both units is to repeat the measurements with just one Hash board connected at a time to see how each of the individual boards compare?

I have this gut feeling that there is a chip in the second unit causing the problem, but if that is the case it will be difficult to find. It is a pity that HW errors are not reported on a per chip basis. 

For comparison on the points I have checked Unit 1 & Unit 2


All the above provisional as it's been done rather quickly.



Rich

Hmm, unless i missed something. Why use 2 connectors? Just 1 connector should be fine?

Putting aside the costs for the moment, I would urge caution in working with the step-down bucks. It seems to me that you'll need to use one buck per PCIe socket. That would only get you 15A for the one socket, and I wonder if that will supply enough current for the blade (i.e. 150W @ 10V). The bigger hazard then might be in thinking that you could use a second buck for the other PCIe socket on the same S5 hashing blade. As I understand the S5 hashing blade, Bitmain cautions you NOT to use two different power supplies (or rails) to the same blade, since it's unlikely they will have exactly the same value on the 12V lines, and the two supplies will "fight" over the result, and it might not end well. If you use two buck's, then you effectively have two different rails going into the same hashing blade, even if they originated from the same power supply.

I am NOT an electrical engineer by trade, and it may well be possible to get the buck voltages close enough that it won't matter. I am certain if you need more than one buck per hashing blade it will be important to get them very close in their voltage.

If you can get somebody like sidehack to weigh in on this and tell me I am full of sh*t, that would be swell. I just didn't want something to get damaged (e.g. an S5 blade, a buck) without considering the possible ramifications of using adjustable step down buck converters.

One thing I have always wondered: Where does the controller get it's power from? One specific hash blade or what? That might also factor into this as well.

Best of luck, Phillip!

yeah the step down bucks came today..

  so lets say they work and really do 15 amps at 10 volts and 95% = All best case numbers!

that is 15 bucks a pcie cable at least 2 per s-5 costly

I am just being a bit cautious at the moment. Don't want people to blame me if this all comes crashing down. I have another V1.91, without heatsinks coming later Today so will be able to confirm that my first one was not a fluke.

I had first attempt Today at turning down the Volts on one of My Server PSU's an Intel DPS-1200TB A. An excellent Platinum PSU at 12V. No pots to twiddle inside it so I had a go at using the Sense voltage to make an adjustment. Goes up to 13.4V but could only get it down from the standard 12.2V to 12.02V. At that point there seems to a limit built into it.

I feel that to get the best out of the undervolting we are going to have to find a way of adjusting a Server PSU voltage, as any added converter module is going to throw away a lot of the gain. I have a couple of other Server PSU's to try so will see where I get with them?


Rich

Don't let skepticism ruin your mood. I think you'll make a bunch of friends if an affordable mod come out of this. Skepticism is a necessity to keep your sanity, here. Its always like this in public places where money is involved. Anyways;

Its a bit of a shame we can't do this mod for 0.001cent like you can with the S1 soft mod. I'd be interested to test this for fun if i/we find a convenient/dirt cheap way to test it out on separate units. Otherwise for now i'm far from being able to draw a benefit from doing this currently.

Beside maybe making the unit nearly completely silent. Which could make quite a few individual's wives very happy.

I fear i can't really contribute currently, as all i have available is a Multimeter and a 1.91 antminer from the same Batch as you.

Possible good news on the S5+ design. Have had a look at the low res images of the Hash board front & back and as far as I can see there are multiple clock modules and unless my eyes deceive me I can see the level shifting diode missing from every interstage clock coupling.

So now the IF's  IF I am right on this then there is a clock module for every triple, 16 in total. So IF this is what makes the S5 undervolt it may well carry through to the S5+

Rich

I did some quick math on it.  the step downs I found  cost 15 each    9 boards with 3 connectors = 27 stepdowns at 15 is 405 not acceptable to spend that much.

But to test this I can use  1 board run it at freq 200 get   power numbers.     

With 12 volts each board should use  close to 385 watts at freq 325

Once i see what the one board does at freq 200 (my guess is 237 watts as  200/325 x 385 = 237)


then run the one board with 2 step downs which are rated for 150 + 150 = 300 watts   if I go down to 10 volts  10/12 x 237 = 197.5 watts a board

so the full thing  would be 9 x 197.5= 1777 watts  and hash at   4.75 th      or .375 watts a gh.






If all of the above is true  the s-5+ just got better.

My gut feeling, For what that's worth,  is that it will probably be ok....   You will of course need a lot of amps at the reduced supply voltage, so we will need to crack modifying server PSU's to run at a lower voltage. The good news is that because they have extended the chain length to 16 all of the voltages in the table will be Plus about 0.7V for the same J/GH. So in the order 10V to 11V

Rich

yeah but if true it makes the  s-5++ worth quite a bit more.

Start out with whatever your power/sound/heat will bear.  Then clock differently as your prices for power change.

In my case power is 16.9 cents in the summer  add more for ac.  Not that I use ac to cool it but it will warm my house a bit   so I am in the 18 cent a kwatt for power right now.

If it works on an s-5+  I can down clock until my winter power rates come in  13 cents minus a 3 cent benefit for the heat. 10 cent power.

It allows me the s-5+ in house.

 If it does not work  I have a place to host it.  I ordered the coins from coinbase  they come on friday.