Topic: Hacking The KNC Firmware: Overclocking - page 62. (Read 144362 times)

This is my best. I've played the whole day with it.... 

Correct.  As shown you could go way more but I'm not looking to take too many risks.. at least for now :-)

The best results are tuning each die for low HW error.
Goal of <2% HW per die (under 1% for happy die)
You need bfgminer to find this info!

I set the voltages in Advanced Tuning to default.
I bump the SPI clock up to 256k.
I do this before I run the custom start script.
I ignore Volts & Amps and just look at Watt.
(I plan to return to volt testing for the slow die later)

The following example would lower die 0 on module 0 and the rest remain stock.

Code:

# ------
# Insert DIRECTLY below the original clock line in the COPY OF THE FILE
# KEEP THE ORIGINAL LINE and add a section for each module/die you want to change
# 6 possible ASIC modules 4 die each
# $p = plug # on the mainboard (count from zero)
# $c = die #
#*KnCMiner on the Advanced Tuning web page counts from 1 !(but not in the code)
#*They count die from 1 also
#*Subtract 1 from the ASIC # on the web page to use in the .sh file
################
if [ $p -eq 0 ]
then
       if [ $c -eq 0 ]
       then
               cmd=$(printf "0x86,0x%02X,0x01,0xC1" $c)
       fi
       if [ $c -eq 1 ]
       then
               cmd=$(printf "0x86,0x%02X,0x01,0xD1" $c)
       fi
       if [ $c -eq 2 ]
       then
               cmd=$(printf "0x86,0x%02X,0x01,0xD1" $c)
       fi
       if [ $c -eq 3 ]
       then
               cmd=$(printf "0x86,0x%02X,0x01,0xD1" $c)
       fi
fi
# ------

I am running an open case.  With the issues I had in the beginning I've just made it habit.  I have a workstation set up in my garage and I run my miner down there since the temp is considerably lower right now.  I work on PC's a lot and instead of buying expensive canned air to clean them all the time I invested in an air tank, hose, and blow nozzle ($50) that I can fill up any  place with an air pump. Works great to keep my miner/fans clean in the garage too. I blow them off about once a week to keep residue / dust off.  I have actually been paying close attention to the temps since I applied the mod.  At first I had 2 ASIC's that were getting close to 70, the other 2 are good to go and run between 48 - 56 give or take.  From what I've read 70 is no worries.  I had a box of PC case fans (nothing special just plane jane case fans) I had ordered some time back and plugged 2 of them up to my power supply and positioned them at about a 45 degree angle on the sides of the ASIC boards that blow air on top and underneath of the boards.  Surprisingly  this took the temperature down a pretty good amount. Honestly most of the heat seems to be coming from the voltage regulators vs the ASIC chips.  I have to say that I have been lucky with my DIE's too though. Pretty happy with the output right now.  When it gets too expensive to run this thing I will go balls to the wall just for fun to see what I can make happen, as I may be able to provide info to people getting the new KNC products.  If I fry it then oh well as I'm planning on taking the miner out to a field and blowing holes in it with my AR-15 when no longer needed.  Although I will remove the nice heat syncs if they still work OK.  Long story short.. I am running a Cooler Master Silent Pro Gold 1000w PS and my outputs are as follows:

SSH

Mining Status

CGMiner Status   Running (pid=9584)
Last Checked   Tue Dec 17 15:36:27 UTC 2013
Avg. Hash Rate   633 Gh/s
WU   8174
Difficulty Accepted   2688638
HW Status

ASIC slot #1   39.0 ℃
ASIC slot #2   -
ASIC slot #3   56.5 ℃
ASIC slot #4   39.5 ℃
ASIC slot #5   34.5 ℃
ASIC slot #6   -

I mine with Eligius now and you can have a look at my stats:

http://eligius.st/~wizkid057/newstats/userstats.php/1K5sYZvJK42k8WBRYHcW3SxhCRZNdKYmAL

Hope this helps someone and thank you once again to all that put this out there!

So, my 8 VRM boards did not blow up. Tried 1F1 and 201, 201 did not make the unit go faster apparently but it was a very brief test. I bailout fearing something was wrong.

I did not see the amps go up like crazy in either.

They range from 42.7 to 46 with 1F1. Temps below 60 with extra fans and a an attempt (not a very good one) at a custom case to increase air density. Cold room. Remember that the temps are measured in a little IC in the board and not directly in the ASICS. This can be misleading especially if you have an open case.

Thanks tolip, taco, gravito and Fpgaminer! Hope a lot of 8 VRMs now have the courage to try it also. After 15 mn my average is a bit higher than 600 from 550.

For those hesitant on the method I just edited using vi (like The-skeptic wrote just before me) but tolips sed replace command is probably more user friendly (just one copy paste).  


edit: a short incursion into 201 and 211 seems to lead to poorer results in my unit (maybe others) due to increased HW but maybe these go down after an extended time when cores get turned off (similar to what someone already wrote), but something that I did not risk.

The sed command changes one register and gives you a mild 7% boost in speed. In order to see the fastest speeds you need to read though the hieroglyphics on this thread and figure out what they are saying. when I figured it out, I am getting almost 15% over factory speeds.

The main thing you need to do is ssh into your miner and use a text editor to edit /etc/init.d/cgminer.sh

try:
vi /etc/init.d/cgminer.sh


Use the cursor keys on your keyboard to scroll down till you see this code, the important parts you will change are in bold:


When you see someone say "Try 201" (which gives about 12% speed increase) you would edit the code to look like:


Hit escape when finished. Then hit the : key, then x to save.

relaunch cgminer with the command line:
/etc/init.d/cgminer.sh restart

Don't blame me if shit blows up. I'm just one newbie helping another.


I personally use 211, but have extra fans inside the case for extra cooling, just to be safe. A list of codes and corresponding speeds are listed several times in this thread, I suggest trying several until you find one that gives you a good boost in speed without going over 40-42watts per die, since they are rated at 40W per VRM.

Good luck.

and all you did was issue this:

sed sBD1BF1B /config/zzz.sh ; /config/zzz.sh restart

??

Thanks for sharing! Day 1? 8 VRMs?

what temps are you seeing??  added fans or just open case ?

I second that.  I'm getting between 575 and 625 on my day 1 shipment miner (which arrived on day 5) .  The Beaglebone board went south on mine the second week I had it and I lost 2 weeks due to incompetence by KNC.  This is truly amazing and I thank you so very much for sharing the info.  Still wish I had my 2 weeks back when the difficulty was still 86,933,018, but this takes a little pain away from the wound and I salute you for it.  Curious if any of you guys bit on the new miner?  If so I look forward to seeing what you can accomplish.  I didn't bite on that one myself after getting burned the first time... but will be watching.  Many thanks!


- .m.w. -

I would like to thank the folks who've done the legwork here. I have never been a unix fellow, but I have been around pcs long enough to remember having to manually edit batch files and .sys files to get programs to load properly, so this brought some of that primal computer beast back in me.

I have a Jupiter that originally started life as a Saturn with two bad die 0 boards. KNC sucks at customer support and expects me to send both hashing boards back and lose a week to 10 days worth of hashing which will take months to mine back. I was lucky enough to pick up two good hashing boards during their 2 minute blow out and those are hashing nicely at full speed, giving me a jupiter with 14 good dies out of 16 total. With .99-tune and no overclocking I see 494-496GHs reported at CGminer with a 2.3-2.6% HW error rating and 485-495 reported at BTCGuild.

With just the sed command line I was boosted to 527 GHS with about 2.1% HWE. After reading the cryptic messages earlier in tis thread, I finally was able to figure out where to edit the cgminer.sh file to give me a really nice overclock. A 211 setting has me at 567 GHs and a nice low 1.65% HWE at CGMiner. BTCGuild reports 558 to 564 Ghs. My local KNC stats page shows that the enabled dies are each consuming 39-43W each, which is right at their stated 40W.

Regardless, my hash rate is up higher than my non-overclocked healthy Jupiter and the HWE rate is about 1% lower, so I m very happy.

My verdict: Awesome. Increases hashrate while lowering hardware error rates, and appears to actually run lower HWE than the stock clock settings. I would recommend replacing the case fans with something more vigorous to keep the rear ASICs cool, or add some additional fans inside the case to increase airflow, since this definitely brings the temps up 5-7 degrees. But otherwise, rock on! A 14.75 increase in hash rate for free deserves a tip!

-D

Here is a partial list of clock speed and value(s)
(it goes on in both directions)
?= untested


I have hunted for one more column, not found yet...

My error, thx for the correction.

0.99 Tuning is correct. (0.99 test refers to same thing)

Everyone keeps referring to 0.99 test. Is that the same as 0.99 Tuning and does that matter?

I think you're right, if you could get one year OC'd you could be better off then 2 years not OC'd.  This could be the case if all companies follow through with what's predicted. No?

I wish there were some extensive research done on how OCing ASICs affects their longevity.  It's the one reason I won't overclock any of my gear...

But if you look @ it from a mining calculator perspectve, I guess you could argue that you want as much hash power as early as possible, as opposed to as long possible with average hash rate... hmmm
This goes against my electronic-principles!

You guys using the Eligius Pool? Eligius.st?

TRAGEDY

2 days in a row my miner waits for me to sleep and stops one die.
(I am inventing fiction to explain my die knowing I am awake)
Speed at pool still better than stock with one die off.
Different die each time.
First time I had to POWER CYCLE 2 times to get it going again!
It was an intense few minutes.
If you have ever had a PC not boot due to dead disk you know the feeling.
Second time I just restarted with lower clock on affected die and all OK.
(no reboot necessary)
I lower die speed one 'notch' and continue.

I use 99 test.
I leave Advanced Tuning at DEFAULT voltages.
I change SPI clock to 256k (default is probably fine)
(no ryme or reason for this one but it looks good on the scope)
SPI clock too fast and HW increase.

I have every die running different speed.
Sometimes I can change the clock on a single die on the fly.
Sometimes when I try to change on the fly it stops everything.

I have a notion that the ALL KnCMiner ASIC's do not like much more than 40 W per die.
Posted VRM info from Nov miners shows all close to 20W also in line with notion.
I can run a die faster/hotter, but not all die on ASIC fast/hot.
I keep AVERAGE power on ASIC close to 40 W per die and it works well.
I do not run 2 in a row fast, I skip one to keep total power lower.

The clock and the power are directly related.
1D1 750 38W (stock)
3B2 750??? (might not work but there is a pattern to follow)
3C2 762 38W
3D2 775 38W
3E2 untested (787???)
3F2 untested (800???)
1F1 800 39W
211 850??? (does not work but is next in 'pattern')
825 39.5 W
850 41W
875 43W
950 49W (not stable long term)

If you have a sleeping die you might try
1C1 700?? (just a guess)
392 725??? (just a guess)
3A2 737??? (just a guess)
Something is better than nothing!

I'm guessing the VCO runs between 1-2GHz based on observations with bfgminer and miner.php

WARNING some clock combinations will never work and may not be good to try.
WARNING some clock combinations draw LOTS of POWER
DO NOT go hunting with 8 VRM's unless willing to sacrifice a die or ASIC to the cause!
You have been warned!


If you are not using bfgminer you are flailing in the dark.
You will never know the status/speed of individual die.

If you want to use custom miner.php page (needs bfgminer)
My modified myminer.php looks like this -> http://codepad.org/YLQBbJEm

"Dangerous toys are fun but you might get hurt" Vash the Stampede

Enjoy


EDIT if your HW error is above 1-2% you can probably do better.
When a die is happy HW error are low.
If HW error is high a lower clock(lower HW) will be faster at the pool!
You need bfgminer to know details.

this is precisely what I want to avoid, i.e. cook my ASIC chips.

And that's the reason why I'm asking, since after a cursory look to the thread it seems to me that nobody have tried those tweaks on a october 8 VRMs ASIC board.

Thanks for posting that.