Topic: Hacking The KNC Firmware: Overclocking - page 64. (Read 144362 times)

Hello,
All this tweaks that you're testing, which firmware require or are they ready for all firmware versions?.
+1 for your work.
Thanks in adavance.

This is my 'slow' one.
first ASIC die 0,2,3 are running 1F1
die 1 is running 3E2

Second ASIC is all 1F1 every die (he is a happy camper)

I have neglected to follow my own recommendation.
I have some die that like a little more power.
I am still under 40W per die
ASIC module temp is steady.
HW errors are low.


This miner started life at my house with 20-30% HW errors on 0.95
0.96 got the HW errors down to 2-4%
Fiddle with the clocks and it is better still
I bumped the SPI clock up a bit and got lower HW error.
256000Hz is what I have it set to now.
Too high >300K and I get lots of HW error again.

If you change anything on the 'Advanced' page and 'Apply' you need to restart with modified .sh again.

YMMV

While I was asleep bfgminer quit and cgminer took over at default speed.

Below is how I do different clock on different die and module.

ASIC plugs on mainboard are numbered 0,1,2,3,4,5
$p =2 stands for my first ASIC plugged into the third plug and counting up from zero it is 2.
$c = 1 is second die, program counts them from zero (ex 0,1,2,3)

$p =5 is my second ASIC plugged into 6th plug on mainboard.
Adjust according to your miner setup.
It is admittedly sloppy coding but it works, just a test.

I have Saturn, for a Jupiter you need 4 $p sections _IF_ you want different clocks per ASIC.


My next focus will be running only 1/2 the cores on one die and trying higher clocks.
I hope to determine if the limit is power or noise related on 4 VRM modules.

the first 1 is in the previous byte

01     D1    (probably already tried without knowing)

Don't add 3 hex digits where there are only 2 originally.
There is only room for 8 bits. (4 bits = 1 hex digit = 1 nibble)
For reference a hex to bin cheat sheet

0=0000
1=0001
2=0010
3=0011
4=0100
5=0101
6=0110
7=0111
8=1000
9=1001
A=1010
B=1011
C=1100
D=1101
E=1110
F=1111


ex: 0xD1 same as  11010001
the 0x stands for hexadecimal

Sorry but the persons that have the hardware cut me out of the datacenter vnc remote so I can`t try
I`ve seen that the  bertmode v4 is modifying the cgminer.sh file
Best
T.

can you paste your file ?
maybe my modification ware not correct
I did not try 1f1  just f1
changed the d1 with f1 , c1 , e1

tried e1 , f1 , c1  with no actual result .
hash rate lower then d1

@gravito
can you post your file
Best
T.

Thanks Tolip, that was fairly simple.

I'm doing the heat thing, my miner looks pretty cozy lately (literally), so I don't wanna disturb it.

Next disruption I have, I'll give it a go and report my results here. If I use the method regularly,

I will donate. Thanx once again.

It seems many are interested but few familiar with the tools.

DANGER WILL ROBINSON
standard 'parachute rigger line of sight warranty' applies to this post.

One way to test, if you are desperate and have a sleeping die.

First have a funeral for your warranty. They don't live forever
Think twice about following along what any fool on the intertubes says.

load and start 0.99-test
restore to factory defaults

then paste this line into the miner SSH command line and hit enter.

"sed sBD1BC1B /config/zzz.sh ; /config/zzz.sh restart"

Do not include the double quotes.
Should only work with 0.99-test.
Not recommended if you have 8 VRM 'running'.

You can try replacing the "C" in the line above with anything from 1-9 or A-F
F is my favorite.

To go back to normal

"/etc/init.d/cgminer.sh restart"
or
make any change in the Advanced tab and 'apply'
or
just cycle power.

Compare the new file with old to see what sed did
To tune a miner IMneverHO ya really need bfgminer minimum to find clock speed and HW on each core.

If this is not enough info to get some going or not and report results I give up!

Hope it helps!

Awesome work tolip, bondus, fpgaminer et al.! Woudl totally use that resell link if I hadn't ordered already.

Still trying to decypher a few things here but if I understood correctly, October 8 VRM owners stand to gain a lot if they play with these variables AND are able to switch on all the VRMs. Any clue on how to do that? Factory 0.90 firmware?  

I wish KNC gave you guys a BTC or three and had you suggest a preliminary fix for the die0 problem.

I suppose its not in their best interests, because we're not dumping money on them with a shovel.  

Thank you for sharing your work, fpgaminer and tolip_wen  

@  thorvald

I modify the one included with 99 test. If VI is the obstacle  install nano.
(possibly 'opkg install nano' )

@ gravito
yes. yes.

Voltages mostly stock while hunting.  750 is stock, 800 possibly with 1F1.
I have 2 can and 2 can't so far. 2 I had to hunt for best.
I use custom myminer.php linked to in this post -> https://bitcointalksearch.org/topic/m.3714194
Then you can scroll thru the big picture of clocks and HW errors etc.

Have not verified the clock reset from the Advanced page, I will next downtime.
Easy guess is I was incorrect initially and assume reset until verified by someone.

A PLL link -->> http://ece.wpi.edu/analog/resources/PLLTutorialISSCC2004.pdf
I hope we can find the sweet spots and prove mathmatically.

Enjoy


 

Ha..  I guess you missed my holy war with 'orama on this and how he kept telling everyone that the extra 4 didn't matter until finally they coughed up that they had a shortage..  just wish they were up front with it so we knew what we were running into at the time.

Can you post a example of the cgminer.sh you use
T.

They are your toys do as you please.
The part is a 40A part.
Just because KnC could not get them to work in pairs, and chose to run 1/2 as many but run them over spec, and managed to make most of them work is not my definition of working 'WELL'.

Do you think the folks with a sleeping die for a month consider them to be working 'WELL'?

They like to run cool IMneverHO and those who advocate otherwise do not understand why hot has an effect, or that there are better ways to solve the issue instead of whipping them into submission.

I don't question the HOT results, I question that whole approach to the problem.

YMMV

nonsense!!!  KNC shipped our miners to run WELL above those meager limits  

I probably mis spoke when I stated that starting a stalled clock can be done easily with the Apply button on the Advanced tuning suite page.
I now think it just resets clock to defaults and the results I saw were misinterpreted.

No howto at this time from me.
It is easy and dangerous!
I don't teach small children to mix bleach and ammonia either.

Who am I kidding, clouds of poison gas are fun!!!
I'll bury an easy clue or two in the text below.

YOU WILL VOID YOUR WARRANTY
Don't cry to me or KnCMiner if YOU do something stupid, own it fool!!!

Minimum requirement is the test suite for Volt Amp Watt info.
Highly recommended is bfgminer, you can see the speed of each core.
Also recommended is miner.php, lots of core info on a big webpage.

A good place to start might be
https://bitcointalksearch.org/topic/m.3714194

I think the variables are N M and P  in that order (standard PLL vars)
This can be tested but my tolerance for downtime is limited.

The address of the variables to change is already in this thread.
The file containing the address is identified in the thread.

look for 3 hex items in a row (not same byte)
Leave the first and last set to 1 when starting.

First pass add one to the middle item and test.
Add 2 to middle item and you might have a 150 Gh/s per module machine.
(did not work on one of my machines)

If you add 1 to the third item also add to the first to compensate.
Not a 1:1 ratio
All variables are interrelated a PLL is a second order system.

Some combinations will do wierd things.
I am not responsibile if you let the smoke out.

For awakening a sleeping die try subtracting from the middle item.
try 1C1, or possibly 5D3

For speed try 1F1, 1E1, or 3E2

You should never exceed 47A and 37W indicated in the Advanced page of the test suite.
It can be done at low power, if you draw more try another combo.

If it were not for sleeping die folks I would never have said a peep.
KnCMiner did not seem interested, when I inquired, in just trying to fix the dead die using clocks.
I offered to detail a simple test as proof of concept.
The fix for them may include a bonus for others, oh well, they deserve 4 running die per module.
Perhaps an extra 9% can take some of the sting out of the wound.

EDIT I forgot my shameless plug for donations.
I give it away but you can contribute if you choose.

I spent a lot to find and document the details and pass them on, downtime is money.
You can help fund this research if interested.
Just point your miner to one of my workers for a few minutes.

For BTCGuild worker = tolip_ZsearchFund (PPS, best for short donations)
For GHash.io  worker = tolip.anything (replace anything with your name if you like)

You can also use my reseller ID if you purchase a Neptune.

https://www.kncminer.com/?resellerid=206


Enjoy!

Any updates to this ? I'm sure lots of people are watching and standing by

Just to be clear.

On die clock is normally ~749.975 MHz   (<<-- Mega Hz)
Can get ~790-800 reliably.
~0.8 Volts ~47A ~37W per die. IF I KEEP EM COOL!

from 144 Gh/s per module to around 150 Gh/s.
or 144 Gh/s and 0.5% HW errors (with 3 non functioning core) at even lower power.

It's about 9% gain at the pool. Does not negate even one diff increase, but it helps.

If it wakes up a die 33% + 9%  gain.