Topic: Antminer S5 - Underclock - Undervolt - Best J/GH - page 4. (Read 31123 times)

Mine came from the used batch, but any miners after a X date (listed somewhere in the thread) are v1.91. The problem is probably that at some point there was a mixed batch and also primarily the fact that they are also no longer for sale.

So your best bet might be to find sellers of the used hardware and ask them if their S5's has v1.91 PCBs.

I may have asked this and forgotten, but which "batch" did the v1.91 boards come from? I've got the capital to pick up another one to experiment with and I want to make sure I end up with the right one, given many eBay sellers don't know their board version numbers or haven't responded to my inquiries.

I wouldn't expect nor ask you to explain the whole thing to me, just pointing me in the right direction to get started, which you've done already with that other thread. I'll let you all know when the scope shows up and I can start screwing around with it, and once a sale of some more BTC finishes transferring out, I should be acquiring that V 1.91 I mentioned before.

Thanks for the links and info.

Ok I can't spell it all out but...

Have a read through this thread, this post with pictures of the boards and a few notes is useful.

https://bitcointalksearch.org/topic/m.12164839

Download the BM1384 data sheet, shows the pinout and other details
https://bitmaintech.com/support.htm?pid=00720141220132903742H2RLy6C40630

The LDO regulators that supply the voltage fpr the IO Voltage & the PLL are key. They are marked 4VK4 & R18 for the later stages.
Do a google search on LN1134  and SPX5205 for data sheets on these.

Find your way round the board. As a clue the bottom stage in the chain is at 0V and is directly below the 18 Pin data connector. The chain then goes round anticlockwise with the top stage being at 12V.

Lot's more I could say but that's enough for now.   Just be careful when making measurements remember the ground for each stage up is the +ve for the previous stage. Finally set the clock right down to 100MHz to start with reduces the consumption & heat when working on the board.

Rich

Somewhat like starting capacitors on motors (although current vs voltage here), just need it to get it going, then you can reduce the voltage?

If you want to point out a few places for me to start and what to look for, since this will be the first time I've ever dug into these things, that would be a definite bonus. Halloween makes a damned disaster of downtown Madison, so I'll definitely be stuck at home doing nothing that night and I'll gladly take some time digging in.

Also a quick run-down on the specifics of these hash boards, what certain sections/components are, etc. would definitely speed things up on my end, since I'm mainly an analog guy (I do amplifier repair, mostly)

Yes well worth doing. I had a quick look but came to no conclusions. Just remeber that you need to move the ground probe to the ground on each stage as you move up the chain. Also only have one measuring device connected at a time to avoid the risk of shorting stages with different grounds.

As well as the signals the voltages available for the LDO needs to be checked. A quick bit of maths I did indicated that the LD0 gets a bit close to it's limit at lower voltages. Another "feature" that complicates things is that the voltage needed to start hashing, at lower voltages, is higher than the voltage to maintain hashing.

Rich

I'd think in a system like this that certain things, such as the signal and controls have very little tolerance, and given the undervolting, maybe the signals aren't moving around at a high enough voltage to be picked up by other parts (reducing voltage possibly reducing output level of some given component).

What about hooking up a non 1.91 to a scope just to see any differences in inter-component signals, that been done yet? I've got a new scope showing up soon here and if you want me to check some things out I can, since I've still got that older version one here.

It's on my list of things to do.   Part of the hold up is that my two working S5's are V1.91 which can be quite successfully undervolted. I have a V1.5 but it was a failed unit I bought of ebay. I have been taking BM1384 chips off the completely failed board to repair the better board and currently have 26 of the 30 chips working.

However a couple of things are hampering progress. First is that I only have a hot air gun and it's not really up to these QFN chips, so I will invest in a proper infra-red rework station. The second problem is that I am always uncertain when I have removed a chip if it is a good chip, which combined with the lack of a rework station makes progress slow. I have also ordered a qfn test socket and plan on making a little test jig for chips so that I only use good chips in the repair.

Anyway when I get sorted and the board is repaired I plan on making some careful measurements as the voltage is reduced, as well as exploring the value / part played by the multiple oscillators and the alternate LDO Regulator power scheme used in the V1.91...

The other interesting snippet to throw into the pot is that in playing with the R1 which has a single BM1384 the Core Voltage that it will reliably hash at at a given frequency is quite a bit lower than the S5, so this is probably because of the simplicity of a single chip, which does not rely on chained signals or voltages, but it is a clue that signal integrity is key to an undervolted S5?

Rich

Has anyone dug deeper into these things, like seeing how the power or signal chains look on a scope at lower voltages?

I just tried 250W @ 10.4V and the voltage dropped down to 10.258. I didn't give it enough credit, maybe because of the crappy uni-t 61E I'm using. Also the fequencies from 325 and above to 350MHz have been tested for only 10 minutes (voltage drop was around 0.3V and much more important the ~43 amps @ 350MHz).

My best revenue is between 250-275MHz at this difficulty, exchangerate and costs for electricity, so there is no need for the sweet spot yet and one converter for each S5 is good enough for me.

Well, there is still some room for finetuning...

Nice set of results and quite consistent with mine. I can't start up with less than 9V although once hashing I can then wind down to about 8.4V although it's a bit prone to then losing one of the boards.

I have connected the 40A BC to one of the S5 instead of the 2 x 5V Meanwell Supplies. It's set to 10V running at 250MHz and I am seeing an on load voltage change of 0.2V whereas it was 0.1V with the Meanwell supplies. I will also try some higher Frequencies, but need to hook up another 40A BC to the other S5 first as I am running them on a single controller board.

Rich

Yes, I got them too. It looks like the brass terminals are made of copper (or are copper-like material, could be also an copper/zinc alloy =bronze -you never know) and the package has been shipped from singapore.

I did some more tests @200Mhz with lower voltages.

black = 40A buck converter (so far there are no issues after 12h)
blue = 2x 15A buck converters (no issues after 1 month @ 24/7)

@100MHz	270GHs	@08.93V***	96 Watts @wall	0.3555 J/GH	HW 0.08%
@100MHz	328GHs	@09.80V	118Watts @wall	0.3598 J/GH	HW 0.0060%
@125MHz	285GHs	@08.93V***	119Watts @wall	0.4175 J/GH	HW 0.12%
@125MHz	371GHs	@09.05V	123Watts @wall	0.3315 J/GH	HW 0.071%
@125MHz	401GHs	@09.80V	144Watts @wall	0.3591 J/GH	HW 0.011%
@150MHz	437GHs	@09.05V	147Watts @wall	0.3364 J/GH	HW 0.11%
@150MHz	487GHs	@09.80V	170Watts @wall	0.3491 J/GH	HW 0.03%
@175MHz	546GHs	@09.80V	196Watts @wall	0.3590 J/GH	HW 0.06%
@200MHz	632GHs	@09.80V	221Watts @wall	0.3497 J/GH	HW 0.06%
@200MHz	663GHs	@10.001V	245Watts @wall	0.3695 J/GH
@200MHz	657GHs	@10.80V	267Watts @wall	0.406 J/GH	HW 0.0027%
@250MHz	810GHs	@10.05V	305Watts @Wall	0.3766 J/GH
@275MHz	859GHs	@10.50V	349Watts @Wall	0.4063 J/GH	HW 0.07%
@300MHz	897GHs*	@10.50V	378Watts @Wall	0.423 J/GH	HW 0.15%
@300MHz	949GHs**	@10.80V	397Watts @Wall	0.4183 J/GH	HW 0.06%
@300MHz	972GHs	@11.10V****	428Watts @Wall	0.4403 J/GH	HW 0.008%
@312.5MHz	949GHs	@10.80V	414Watts @Wall	0.4362 J/GH	HW 0.10%
@325MHz	966GHs	@10.80V	429Watts @Wall	0.4430 J/GH	HW 0.14%
@325MHz	1047GHs	@11.1V	465Watts @Wall	0.4441 J/GH	HW 0.028%
@337.5MHz	1071GHs	@11.1V	482Watts @Wall	0.4496 J/GH	HW 0.047%
@350MHz	1090GHs	@11.1V	502Watts @Wall	0.4606 J/GH	HW 0.096%

*One board dropped after 30 minutes, it was around 912GHs before that happend
**I doublechecked that
***One board dropped or didn't even start
**** Voltage drops under load by 0.1V (11.2 without load), Amps @36.25
Sweet spot so far

more tests at lower/higher volts soon

Edit: lower frequencies and voltages -> 9.05V seems to work fine and stable

Yes I don't like running with much more than 0.05% HW error and aim for between 0.01% to 0.02% at which point I seem the get the theoretical hash rate. I usually find at 0.1% and above the hash rate drops below what you expect, although as I have said before at power up I sometimes get quite a lot of HW errors so you have to let that rate work through. Remeber you can calculate the expected Hash Rate by multiplying MHz x 3.3

I have also just taken delivery of a couple of 40A Buck Converters, had a quick check and they seem ok but have not connected to the S5 yet. Are yours the Black metal case potted ones with 3 Brass terminals and the gold heatsink?

Been playing with the R1 so just letting the S5 which is on the Meanwell 2 x 5V PSU's run at the moment, but will probably try the 40A units Tomorrow.

Rich

The 40a buckconverter arrived yesterday. I've ramped up the voltage a bit:

@275MHz	859GHs	@10.50V	349Watts @Wall	0.4063 J/GH	HW 0.07%
@300MHz	897GHs*	@10.50V	378Watts @Wall	0.423 J/GH	HW 0.15%
@300MHz	949GHs**	@10.80V	397Watts @Wall	0.4188 J/GH	HW 0.06%
@312.5MHz	949GHs	@10.80V	414Watts @Wall	0.4362 J/GH	HW 0.10%
@325MHz	966GHs	@10.80V	429Watts @Wall	0.4430 J/GH	HW 0.14%

*one board dropped after 30 minutes, it was around 912GHs before that happend

**I doublechecked that, it's valid. The second run (doing it now) is @ 955.5 GHs (avg. after 33 minutes). Its nice to see, that even a slightly higher voltage can increase your efficiency as well. My guess is, that a lower HW error rate might be the reason for that -even if it is negligible @0.15. So finetuning it is an eminent factor.

Yes agreed I have a spreadsheet and at any point in time you have to adjust the Frequency for maximum profit. However even at the standard clock frequency of 350MHz I can still undervolt to 11.1V for greater efficiency and profit.

My electricity in the UK is $0.15 and at current difficulty I am now running at 300MHz and 10.7V

Rich

So i played around with some numbers to the result of undervolt of the S5 to profit to see what kind of numbers we could expect in this difficulty range.



So as we can see depending on your costs undervolting the S5 could end up costing you profit.  So I would be careful about when and if undervolting will result in greater profits.  In my case it would not be until difficulty rises 30% and at that point then applying the mod would add benefit.  Just a word of caution that not all things come out in theory as they do when you compare the data on paper.

Correct. I bought half a S5, which is why i'm trying to find a working S5 blade to slot it in, it would be pretty simple, but so far no luck.

It look like the boards, at factory setting do perform nearly exactly the same however, but i do not know why when "we" push them down or up you start seeing differences as i'm not very knowledgeable about electronic fabrication.


It could, i'm not sure about the variation in temp, i think it just depend on where the sensor is on the board, if its closer to intake, it will read lower even though overall the board is at the same temperature as the other.

However i noticed that if you don't put two fans, one of the board is actually fairly hotter than the other, and if you put them on the side, the top board seem to heat up more, probably because of thermodynamics... i mean the how air likes to go up.

Ok here's what I have been running with for the last couple of Weeks.



It's a single S5 Controller with 4 Hash Boards connected. 250MHz with the best two Boards undervolted to 10V & the less good pair at 10.3V. Am using the pairs of 5V Meanwell PSU's to power it. Hashes at 1650GH/S and takes 615W at the wall giving 0.373J/GH.

It's just in my target HW error rate at 0.222% but more importantly the actual Hash rate matches the Theoretical Hash rate. The boards are arranged in the systems by how good they are and you can see that the best two run at 35 & 37 Degrees and the worst two at 41Degrees. Will probably run with this setup until I replace it with individual Buck Converters for each board, assuming they give a better result?

Rich

Yes interesting numbers, why id 5S3 running at half speed, is it just a single  board?

 I have come to the conclusion that there is a lot of variance between Systems and also between individual boards. I have running at hope two S5 at 250Mhz, one system undervolted to 10V the other running at 10.3V. I have put the best 2 boards in the first system. There is however quite a variance in the HW Error rate & the board temperature. I will post up the numbers when I get home.

Rich

Stock. I don't have undervolt or overvolt capabilities, this show the difference between the boards. The 0.006% is over a long period of time, so its temps. Even though it does take a while to stabilize the HWE% over time, it does go down as well, not just start up. The S4 is like that. HWE tend to go down a bit after start. But that just seem to me like most ASIC like to start with some HW.

Here is the header, not that it will give much more information;


This just show that v1.91 has some capability to stabilize the chip's voltage slightly, which may be what the oscillator were intended to do when put in place. Enabling us to feed it down/over volted current is probably a side effect.

The input voltage is the same for #1 and #3 (and #2 is effectively the same too but #1 #3 has the same source and same reading) yet they have different HWE%. The chips are the same, controllers are the same, etc. You know better than me what that leave as reasons.