Topic: Hacking The KNC Firmware: Overclocking - page 38. (Read 144377 times)

mekadeka, did you measure it at the pool? If you didn't have time for that, which of the numbers in cgminer did you use as a reference?

cheers!

Temen, sent you a modest something.

Updates for folks OC'ing October Saturns in particular (because it seems there are differences):

I tried 265 on a machine running 0.99.1-tune. This seems to have netted me the best hashing results so far, FWIW.
EDIT: At about 340ghs right now with low pool rejects. Best I could do with 211 (the next best) was around 311.

Temen or anyone else, why would 265 work particularly well? Can anything be extrapolated from this particular clocking? I've tried a slew of them, and this FW (0.99.1-tune) are the best for me so far. I did not make any tuning changes, nor did I use wizkid's cgminer, nor do I have bertmod in place right now.

I know that 265 itself is faster than 211, but that's not all of it. I tried a bunch faster than 211 (221, 231, 241, 251, 305), and none worked as well as 265. I don't know for certain if 0.99.1-tune was on each of those, however, which is what I'm using right now.

Can anything be told from the 265 clock setting? Can I set it a logical multiple up from here? (Sorry, I really do not know anything at all about OC'ing. Apologies.)


Thanks!

sorry 

you didn't read the whole post

Well thxs to Nil's help on IRC and his thread http://forum.kncminer.com/forum/main-category/main-forum/27841-0-99-x-tuning-die-tuning

I'm now able to see per die hw error n apply more volts to that trouble maker!! woot..and now playing around with volts and watching graph i guess we were seeing things cause i got one mod that doesn't have much error and hasn't moved since playing around with others

and yes adding to many volts to a die will increase hw error as well...that is why i was having such a hard time with 251...cause i just kept adding more n more...all is good now .5% hw and 720gh/s on 251

I see the same. :/

I wasn't able to try different clocks in different boards because of that. Not able to  pass 201 (Oct Jupiter, all boards, cgminer, BFGminer gives me much lower hashrate for some reason, tune 0.99.1, 0.0542V on all dies).

october jupiter ?

This is the same in my experience and is why I OC but don't change the volts.
I couldn't get a good feel with the volts,
as even the way they react changes over an indefinable period of time.
It's like trying to mix paints and not understanding the underlying color theory.

@mekadeka: I think its bertmod 0.4. I have november 0.99.1-E and it seems to wrk with it. Is there newer revidion hm... =)

I'm having a hard time getting 251 running good...had 241 running at 700gh/s with .5% hw errors...so i tried reseting volts to default and playing around with it...I'm hoping i can eventually get it running good...starting to notice 2 vrms like to drop down in amps like temen but a restart on oc'in will get them to play well...dunno if that means they are giving out to the stress or what but i hope they don't go will update after more playing...i'm on an oct. jup just incase anyone forgot...ya i don't think oct. jups will be able to oc as high as nov..not even close

I've also noticed if you play around with the volts on one module...it'll affect all modules..dunno if its just my unit or what but if i change the volts on one module...I get different hashrate/hw errors on all modules...or maybe i'm just seeing things idk hehe

which version of bertmod are you using?

I don't know anything about clock-speeds as I dont have any advanced tuning on november, just bertmod (which is by no means bad, its great tool to see vrm amp and other stuff). I wish i had some command to set voltage manually on asics. I think it has to be 12cset command and bus and address and commad bytes etc. combination. It writes number 3 on the line on the start to busses 3 4 5 6 7 8 and their respective vrms on part wher it says "clear faults on megadlynx". That number 3 is fault clearance on written to correct address (which seems to be 0x11...0x18 but dont have the guts to try it =) Dont know anything about coding so perhaps I leave it be..

VOUT setting would be PMBUS address 0x21 I think but where it is located here dunno.

---

3.8.1 Set TCK/SK Divisor (FT2232D)
0x86,
0xValueL,
0xValueH
This will set the clock divisor. The TCK/SK always has a duty cycle of 50%, except between commands where it will remain in its initial state. The initial state is set using the Set Data Bits Low Byte command (0x80). For example, to use it in JTAG mode you would issue:-

---

Thats excerpt from kncminer github some command processor manual.

There is also register which divides the clock speed by 5 as default. I wonder if kncminer just disables this division with neptune 

'
Thanks, Temen. I'll definitely throw something your way for giving me a hand.

I did try the wizkid cgminer. The result appears to be faster hashing  poolside but with pool reporting significantly more rejections . My errors have always been pretty low. Right now they're reporting about 6-7% after 15 minutes of testing. Former hash on the machine was 311, now4. showing 317 in wkcgminer. Pool rate is all over the place, right now at 304. I'll let it run a little more.

Temen, how should one go about testing clock speeds?

There is something quite different about the November Saturns, IMHO.

To me, it seems the answer has to be yes, only November. I have an October unit. Last night I tested a number of different variables used on Temen's version. The biggest fault in my test is that I could not let each test run for more than about 6-10 minutes. This might or might not be long enough to determine if a given configuration will *eventually* become a faster hash rate. I have not tried Temen's full configuration yet (including bertmod), but I did do Temen's speed 305.

So far, for my October machine, the fastest has been cgminer 3.9.0 and 211. I also threw on three additional case fans placed in the middle of the case (just standing there, not affixed right now) and blowing directly on the chips. This drastically reduced my temps from mid/high 60s to mid 40s.

Things I tried with my October unit that didn't seem to be faster (but tests might not have been long enough):

1. Tried Nov-only FW 0.99.1-e with 305. Just a shot in the dark that didn't do much.
2. Tried 251, 231 with I think 99 and 99-tuning. I don't think these were notably faster.
3. Tried 99-tuning with 241 and 221. I think 221 looked promised, but not run long enough.

I tried a variety of voltages in 2 steps, but nothing seemed faster. TBH, I am clueless and nearly ever OCer knows better, so I'm just a blind tester. I don't understanding tuning at all.

I think three things so far:

1. A true testing matrix trying a full variable set trying FWs, clock speeds, and custom cgminers, and voltage, spi frequency, and spi voltage is needed.
2. A determination concerning length of testing time is needed.
3. I'm probably missing 3 more things.

Same question

do i understand it right. 305 ist only for november units?

Can't you see what failed with bertmod? Maybe you can't do anything about it without advanced tuning but perhaps see if it's temperature related (FAULT 4, etc.)

@mekadeka: wizkid you can get like this, instructions were on kncforum posted by bitbuster, didint find the link but here are instructions, you need to ssh to the miner first and grab control of the miner with screen -dr:

q
wget http://vpn.wizkid057.com/nas/cgminer-binary-wizdev20131222
chmod +x cgminer-binary-wizdev20131222
screen -dm ./cgminer-binary-wizdev20131222 -c /config/cgminer.conf
screen -d -r

My experience is that when applied same setting this version gets better hashrate on console and little bit less current consumption BUT on the poolside It's always less than with cgminer 3.9.0. Dont know if its issue with pool or what. On the other hand I have seen with 3.9.0 on the poolside even more hash-speed than it shows. Like if cgminer shows 459gh/s, poolside it sometimes as high as 488gh/s. This is with 3.9.0. Of course there is some variation.

For me setting 305 seems to be absolute STABLE maximum. With higher settings either one die or one vrm fails. Dont know which it is because don't have advanced tuning.

I fiddled some more with this clock divider. Setting 606 works for me, it gets around 450gh/s. 777 works somehow. The most control is with this xx5 range. With this you can set everything in small steps.

Waiting for donations xD

Temen, how does one put wizkid's version of cgminer in place? I looked around for an answer, but I am unclear.
Thank you.

+1
Yesterday I cleaned one of my Jupiters and temps fell 10 degrees.
Inside it looked like a dust bomb has exploded. I used a painting brush with long and soft bristles to wipe everything, even between the fins of the cooler and the fan blades while sucking everything with the vacuum cleaner at the same time. It took me around 30 minutes for a good deep clean.
I have to clean all the others soon, but I hate to turn them off.