Topic: Antminer S5 - Underclock - Undervolt - Best J/GH - page 5. (Read 31090 times)

Yes in general I would expect performance to be better when cold than hot. I have however as detailed in my earlier post sometimes seen a high error rate when first switched on that then reduces. This may or not be temperature related?

Not sure that I fully understand the table above? What was the voltage for the tests? Would be good to have some column headings on the table?


Rich

I want to add more observations between the different boards versions;

#1 = v1.91
#2 = v1.3
#3 = v1.5



All the 3 different board like the cold. v1.91 is showing 0%, its at 11 HW atm. v1.3 and v1.5 behave the same way, i don't know the differences between the boards but it does not seem to be performance related. The v1.91 does perform marginally better.

#1 and #3 is powered by the same, brand new EVGA GS, PSU.
#2 is on a EVGA G2 1k

Difference of v1.3 and v1.5 at these temps vs hot 60c~ is 0.0035%~ vs 0.006%~

Yeah, I just didn't think they were THAT temperamental. Sometimes I can't go below 12.25V or it complains. Why 60 seconds?

I only see that feature when I am very undervolted down at levels below 8.9V, however I suspect what you are seeing is the problem of undervolting a board other than V1.91?

As I see that  VirosaGITS has also just suggested. 


Rich

Well this would be because pre v1.91 boards does not have proper volt variation support. RichBC found out thats probably what the oscillator are for.

As such you'll only be able to do very slight under/overvolting.

The biggest issue I've been encountering so far is that when undervolted it will run and hash for 60 seconds, then at that one minute mark, it will start bitching and stop hashing.

Good that is working for you with the V1.3 Board. That is a nice range on the supply can you post up the make & model? As  VirosaGITS says it's all about the Hash Rate you can reliably achieve rather than the HW error rate. That said for long term running I aim for between 0.01% - 0.02% I find at these levels I achieve the expected Hash rate. A few other Hints & Tips I have found along the way. 

The expected hash rate for a S5 is 3.3 x Frequency in MHz, here are a few examples.


When you are impatiently waiting for the Hash Rate to stabilise it seems to act like a damped oscillation. If you watch it rise & fall if you note the high & a low of the latest cycle, add them together and divide by 2 it gives you a good idea where it will end up.

When pushing the undervolting I sometimes find that initially the HW error rate is quite high, sometimes as much as 3% or 4% however do not immediately right it off as I find that it will sometimes drop down to acceptable levels.

Measure each board individually as you may find that one board is better than the other. I have two S5 and have ended up swapping the best boards into one system as I can then get the best out of each system. In my final configuration I plan on using Buck converters and may use one for each hash board to enable them to be individually optimised.

When you have some solid results for the V1.3 please post them up. I have a V1.5 Board but bought the system as faulty so at the moment have no measurements.

Rich

Well, at this point I'm playing with that voltage adjustment pot like a teenager with... well yeah. I'm getting some impressive numbers (at least to me running with my old mess of a supply) and I'll eventually find a few good points to put together for data on mine, a V 1.3 board.

1% HWE is huge for Antminers, at stock loads i get under 0.0001% error rate. But 1% is still just 1% of your hashrate wasted. Its not much if you're looking at a 30% efficiency gain.

But with a good PSU, such as a EVGA G2/GS, seasonics, other high quality OEM, you should'nt really go above 0.1% even if you push it hard and i'd say thats rather minimal. I don't know about the stability of the servers PSU, sadly. But even if its not too clean power, again 1% isint too bad its 14GH at 1.4THS wasted.

That other supply of mine ran stable at 425, just shy of 1.4TH, but at an expensive 870W since it was running out at 12.65V. For some reason I figured a high % of HWE would actually degrade performance. I did have issues with hash boards not firing up under 11V, but I'll keep playing with it with that knowledge now.

Whatever. If you get 1% HW and you get a better J/GH or better performance, it doesnt matter. Its just that when you go too low or too high voltage for the freq, you will lose performance unrelated to HW.

If you try to go down, just go down to whats stable and the best J/GH, if you go up, you might be able to stabilize 400-425HZ, it doesnt seem possible with stock.

With a good stable PSU, i get very little HWE% however. Under 0.01%

My power supply showed up and I've been playing with it a bit. Shit, you weren't kidding about the older ones not liking heavy undervolting. At what HWE% should I consider something not plausible as a combination of MHz VS Voltage? My supply can go from 9.3V up to 13.something (although for obvious reasons I'm not going to go higher than my other supply, at 12.65V)

Yes it's a great pity that you cannot adjust the frequency on the fly. One thing I will try is to feed the chips with an external oscillator, they are initially driven with 25MHz which is then converted in a PLL with dividers etc to the final frequency used by the chips. There will be some range that the frequency can be changed over and things will still work, whether it is wide enough to be useful I do not know? However worth a try as that would enable on the fly frequency change.

My other idea on making the load Dynamic is to use an S3 and to switch on or off the individual Buck converters hence adding or removing stages. What I do not know is if you cut a stage & then put it back is if it starts hashing again, however it will be a quick experiment to find out.

What is really needed is some custom firmware with a nice external interface and all of this would then be very simple.  

Not too bothered about ROI, but nice to know that it's quite quick, as it's a fun project & should continue to have some use, mining or otherwise for some Years. The depressing thing is that I think just using the electricity generated to power other things in the house and save on electricity cost is better value than using it to mine Bitcoin.  


Rich

I like your idea with the Arduino... with that it would be possible to change the voltage. Its just too bad that we have to restart the cgminer to change the fequency of the asics and losing one minute for each adjustment. It would be nice to change the asic speed on the fly - like sending a halt state to the asics then change the voltage (up or down), then the speed as well and a continue/resume command to the asics. Hm... sounds too sophisticated, I remember my 386er days as a child and those turbo buttons, I guess that could work.

To the solar panels: Well, the numbers are good, but not as good as supposed to be. With 0.370 KW I should get ~ 400- 480 KW per year (the average output for solar panels here is 1100KW - 1300KW for a 1KWp rated solar panel). The panels are facing to the south, but tweaking the angle to the sun for wintertime can improve a lot.

Edit: Just found a calculator and some avg. specs for panels in the UK, so according to this, you can ROI them within 1-2 years as well. (£92 per year for 500Wp sounds terrific)

Yes a guy had just installed a 4MW Farm and was selling off some panels that had cosmetic damage. I was late to the party otherwise I would have had a few more. Paid 50GBP each, just minor marks so perfect for the job. They are Gintech 255W each.

The sensible / easy thing to do would just be to plug them up to a Grid Tie Micro Inverter and just plug them into the house supply, but where's the fun in that. 

So the outline plan is an Arduino, some Buck Converters, some Hall Effect Current Monitors, a small battery and then some way of Dynamically adjusting the Mining Load. Voltage to the S5 is easy and will just be a Buck Converter with a Digital Pot IC controlled by the Arduino, but I need to adjust the frequency as well?

Anyway I am getting ahead of myself, just the planning and collecting parts stage of the project at the moment...

Interesting numbers on what you can generate with a couple of panels, weather here in the UK will not be that good. I did separately also get hold of a 700W MPPT Grid Tie inverter for 85GBP which I will plug them up to when I am not playing. 

Rich

You got a couple of 255W panels for a 100 quid? All together or per panel ? Thats around $155 and a very good price for one 255W panel.

I'm also using solar panels since april 15-> managed to get 2x185W for ~$140.

I'm paying $0.23 per KWh during daytime and $0.13 in the nighttime. So its quite worth it to use some solar panels - on a sunny day they can generate up to 1.8 - 1.5 KW per day and even on a cloudy day those panels can give you 0.3-0.5 KW. With feeding the electricity directly into the outlet with an solar inverter is the most efficient way ~95%. Of course the energy is eaten up by the miner immediately -so nothing goes into the grid (you need a permission for that in most counties). The panels gererated >200KW in the summer season (@.23 thats 46 bucks) so a ROI by the end of 2016 at least mid 2017 is quite possible.

I see, thats unfortunate, would of made sense since you'd basically be selling the electricity back to the grid. Please post here with the link when you start that thread. Raising my electricity cap with solar panels would be interesting.

I am usually the first to say treat Solar as a completely different exercise than Mining. However I don't want a full installation with feed in tariffs etc. I want something I can play with to prove it can be done in a simple cost effective way, so that I can continue to mine whatever happens to Difficulty, Halving s, BTC price etc.. As I say I am just pulling together ideas, and collecting parts at the moment. I picked up a couple of 255W panels for £100 last Weekend.

I would like to do it completely without batteries however some sort or reservoir is going to be needed, particularly with the UK Weather... My idea is that it will mine only when there is power available and the miner will dynamically change the hash rate to consume the available power.

Anyway just an idea at the moment. I will start a thread when I get going.

Sending the meter backwards is a great idea, unfortunately meters these days, even the mechanical ones have a backstop built in to prevent this.


Rich

From what i understand batteries are expensive and they have a short lifetime? Most people who wanted a 24/7 solar setup ended up seeing that the cost to maintain it wasnt much cheaper than the electricity cost.

Could you instead just feed the electricity back into your electricity meter to make it go in reverse at day and suck it up at night? Assuming you feed back twice what you use at day, it would effectively nullify your usage at night?

Yes a nice result. For me with my 15c electricity some form of undervolting is essential as otherwise it is costing me to run an S5. I will be ok until the Halving, but then will have to come up with another plan to keep mining.

I am happy to spend money on PSU and Buck Converters etc as unlike the Miners they are generic and can be used in other projects.  With a bit of luck if I move the S5's on at the right time I will get most of my money back.

Not sure what my next move will be, perhaps the S7 will come down in price or I am gathering the bits for a small Solar Mining experiment. I want to try and do it with only a small battery with the mining load being dynamic to reflect the power available. Just an idea at the moment, will start a new thread when I get started properly.

Still want to explore what are the factors that enable the V1.91 to be undervolted and other revisions of the boards not? I feel there may be a bit more to squeeze out of an S5 by fully understanding the effect that changing not just the core voltage but the other IO and PLL voltage, plus the interstage signal / level shifters etc. At the moment it is not clear which of the voltages or signals is the one that causes Hashing to stop?

Rich