Hacking The KNC Firmware: Overclocking - page 27.

crashoveride54902

hero member

Activity: 784

Merit: 504

Dream become broken often

Quote from: captain-v on February 23, 2014, 12:16:43 AM

Quote from: Powell on February 21, 2014, 06:53:37 PM

Got my second October 4VRM Jupiter in and on first bootup worked fine for a few hours. I updated to 1.0 to test and I noticed when I turned off SPI Auto this:

I've tried the hit it with SPI 3.3 and max voltage. Let it run a few hours like that. I have had them drop when OC to high but this I can't seem to rescue.

There is it at max voltage on that one. I've swapped back to .99.1 , .99.2 and 1.0 and can't seem to get it to jump back up.

I've tried running 231 and 221 but sadly nothing helped kick that one up even. Sad

Any other suggestions so I can put this Jupiter to a happy # Wink

I've been having the same problem until today on my O S+1. Now SPI @3.3V and die1 with -0.0806V and 675MHz made it work. I was getting higher rates with 231 around 450GH/s. Now with all other dies @775MHz total close to 425GH/s with less errors.... Also with all other dies at -0.1465V, power dropped from 560W (w/ Stock values) to 460W at the wall.

Is there a way to 231 the other ASICs and leave this slepy die1 on ASIC alone @675MHz?

yes read the thread and find the msg where it talks about making every die different oc

jelin1984

legendary

Activity: 2408

Merit: 1004

You think leave it as it is

Hashing at 640 gh
Code 211

And firmware 1.00

elasticband

legendary

Activity: 1036

Merit: 1000

Nighty Night Don't Let The Trolls Bite Nom Nom Nom

okay it is not as back and white as just yes or no things can go up down left right or just straight sideways Roll Eyes

jelin1984

legendary

Activity: 2408

Merit: 1004

Hahah

Not all the vrm at 54 only one or two at 54 amper

The other is at 43-48 amp
Is ok or too high

elasticband

legendary

Activity: 1036

Merit: 1000

Nighty Night Don't Let The Trolls Bite Nom Nom Nom

Quote from: jelin1984 on February 23, 2014, 10:14:00 AM

Which is the safe amper for running?

Is better to lower 725 MHz each vrm?
From default 775mhz?

yes, no or maybe.....

jelin1984

legendary

Activity: 2408

Merit: 1004

Which is the safe amper for running?

Is better to lower 725 MHz each vrm?
From default 775mhz?

elasticband

legendary

Activity: 1036

Merit: 1000

Nighty Night Don't Let The Trolls Bite Nom Nom Nom

Quote from: jelin1984 on February 23, 2014, 09:24:46 AM

is ok 54a per vrm or is it
too high? Huh

??

yes. stop it running.

jelin1984

legendary

Activity: 2408

Merit: 1004

is ok 54a per vrm or is it
too high? Huh

??

elasticband

legendary

Activity: 1036

Merit: 1000

Nighty Night Don't Let The Trolls Bite Nom Nom Nom

about 4 days later i am still running 291/810w/790GH/1.5%HW

edit: i would say i would try going higher, but the 200w per board does worry me, but seems okay so far.

captain-v

newbie

Activity: 3

Merit: 0

Quote from: Powell on February 21, 2014, 06:53:37 PM

Got my second October 4VRM Jupiter in and on first bootup worked fine for a few hours. I updated to 1.0 to test and I noticed when I turned off SPI Auto this:

http://i.imgur.com/YskroHj.jpg

I've tried the hit it with SPI 3.3 and max voltage. Let it run a few hours like that. I have had them drop when OC to high but this I can't seem to rescue.

http://i.imgur.com/Fcg0xqb.jpg

There is it at max voltage on that one. I've swapped back to .99.1 , .99.2 and 1.0 and can't seem to get it to jump back up.

I've tried running 231 and 221 but sadly nothing helped kick that one up even. Sad

Any other suggestions so I can put this Jupiter to a happy # Wink

I've been having the same problem until today on my O S+1. Now SPI @3.3V and die1 with -0.0806V and 675MHz made it work. I was getting higher rates with 231 around 450GH/s. Now with all other dies @775MHz total close to 425GH/s with less errors.... Also with all other dies at -0.1465V, power dropped from 560W (w/ Stock values) to 460W at the wall.

Is there a way to 231 the other ASICs and leave this slepy die1 on ASIC alone @675MHz?

tolip_wen

sr. member

Activity: 386

Merit: 250

Quote from: padrino on February 22, 2014, 09:50:54 PM

Quote from: tolip_wen on February 22, 2014, 08:48:32 PM

Quote from: padrino on February 22, 2014, 02:41:53 PM

I might have overlooked something obvious but can someone tell me the math behind the HEX to MHz conversion, for example what is the equation that gets 211 equal to 850Mhz?

*approximate equation

take the first 2 digits the '21'
convert hex to decimal to get 33
multiply 33 x 25MHz and ya get 825MHZ
(add 25MHz to your answer)
The last digit is the divisor, it's 1 so we are done.

YMMV

Thanks for the details, that helps some but the use of a 5 on the end confuses it some.. I'm wrapping up a custom firmware that opens the frequency settings however I need to set the registers manually as KNC is disallowing it in their binary tool. I have it all working just trying to get a stragithforward, yet reliable equation to flip the MHz to the proper hex but it's not looking good yet.

EDIT What I originally posted below is/was an inprogress draft, forkin kbd went south

Just guessing...
5 does not compute, or does it?
What I had assumed was the ASIC only cares about the last 2 bits.

Code:

Hexadecimal     Binary
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

If this is a correct guess 1,5,9,D all have the same last 2 bits.

I did briefly hunt for a third bit functionality but did not persue it due to knowing there is a limit on the clock that is divided.
Divide by 1 has the most potential at the top end.
Divide by 2 if you think you can get 12.5MHz more out of it.

A 3rd bit might equate to another divide by 2.

850MHz would be 874 if the last bit can be zero
or maybe
850MHz would be 875 if the last bit needs to be set

It seems in practice the clock potential far exceeds the power/thermal potential so nobody has reported a clock failure as the limiting factor.

Standard parachute rigger 'Line of Sight Guarantee' applies!
No idea what the ASIC thinks just documented reverse engineering and pattern matching.

EDIT #2
One thing to keep in mind.
Not sure how it applies to 28nm but, the clock itself uses power and slower is often more efficient.
A div by one may leave more for the hash cores to consume from the budget.

YMMV

padrino

legendary

Activity: 1428

Merit: 1000

https://www.bitworks.io

Quote from: tolip_wen on February 22, 2014, 08:48:32 PM

Quote from: padrino on February 22, 2014, 02:41:53 PM

I might have overlooked something obvious but can someone tell me the math behind the HEX to MHz conversion, for example what is the equation that gets 211 equal to 850Mhz?

*approximate equation

take the first 2 digits the '21'
convert hex to decimal to get 33
multiply 33 x 25MHz and ya get 825MHZ
(add 25MHz to your answer)
The last digit is the divisor, it's 1 so we are done.

YMMV

Thanks for the details, that helps some but the use of a 5 on the end confuses it some.. I'm wrapping up a custom firmware that opens the frequency settings however I need to set the registers manually as KNC is disallowing it in their binary tool. I have it all working just trying to get a stragithforward, yet reliable equation to flip the MHz to the proper hex but it's not looking good yet.

tolip_wen

sr. member

Activity: 386

Merit: 250

TRAGEDY!

I cooked a blue PCI extender.
One corner of the white connector on the end of the blue extender is now black and crunchy.

(puts on hater hat)
This is typically caused by a poor crimp or stretched out female socket on the crimped part.
(along with too small a wire gauge used, no 'margin on margin' on that part!)
Why such a shitty part on an expensive product?
(removes hat)

EDIT: in typical hater fashion I spake before I had all the data.
Post mortem indicates damage at both ends of the PSU cable and one end of the blue extension.
Determining the initial point of failure is unlikely and to point only to the KnCMiner extension is unfair.
To make that statement while overclocking is evidence of operator error.

Yes, I know better. (learning thru pain is what I do best!)
This box (a Saturn) was not too radical, it's still air cooled.
(about 180 WATT at the wall for this module)

Time to install second 6 pin PCIe power connector on each PCB.
Thx KnC for the option!

Ya get what ya play for!

All HW other than the extender cable survives

If you are reading this and still using the blue extenders you have an option!
Learn from others mistakes, or not!

YMMV

tolip_wen

sr. member

Activity: 386

Merit: 250

Quote from: padrino on February 22, 2014, 02:41:53 PM

I might have overlooked something obvious but can someone tell me the math behind the HEX to MHz conversion, for example what is the equation that gets 211 equal to 850Mhz?

*approximate equation

take the first 2 digits the '21'
convert hex to decimal to get 33
multiply 33 x 25MHz and ya get 825MHZ
(add 25MHz to your answer)
The last digit is the divisor, it's 1 so we are done.

YMMV

padrino

legendary

Activity: 1428

Merit: 1000

https://www.bitworks.io

Quote from: j03 on February 22, 2014, 03:15:57 PM

Quote from: padrino on February 22, 2014, 02:41:53 PM

I might have overlooked something obvious but can someone tell me the math behind the HEX to MHz conversion, for example what is the equation that gets 211 equal to 850Mhz?

This may answer your question:

https://bitcointalksearch.org/topic/m.5108870

Thanks for pointing it out, I had looked at that before posting, I wish it did but following the logic in that post I can see how he is stepping up but not how one converts between frequencies and the hex setting.

j03

member

Activity: 62

Merit: 10

Quote from: padrino on February 22, 2014, 02:41:53 PM

I might have overlooked something obvious but can someone tell me the math behind the HEX to MHz conversion, for example what is the equation that gets 211 equal to 850Mhz?

This may answer your question:

https://bitcointalksearch.org/topic/m.5108870

padrino

legendary

Activity: 1428

Merit: 1000

https://www.bitworks.io

I might have overlooked something obvious but can someone tell me the math behind the HEX to MHz conversion, for example what is the equation that gets 211 equal to 850Mhz?

crashoveride54902

hero member

Activity: 784

Merit: 504

Dream become broken often

Quote from: sickpig on February 21, 2014, 08:21:49 PM

Quote from: elasticband on February 21, 2014, 07:58:06 PM

Quote from: bondus on February 16, 2014, 09:17:19 AM

I have inspected what the November firmware has actually changed.

What I can see it has changed a value in one of the registers in the VRMs. It has set the "ON" bit in the OPERATIONS register,changing what conditions are needed to turn the VRMs on.
This change will be permanent even if rebooted or power-cycled. It can be changed using i2c commands:

I suggest stopping cgminer before doing this to limit the power when the operation is done.

Reading the datasheet for the VRMs I can see that when using them in pairs (as we do when all are turned on) there are some requirements on how they should be turned on. This method does not do that!

I have heard that hizzing sound as well. Some boards do that. They did that on 0.90 firmware too. No idea why.

I do not know if this is safe or not, it might be unsafe and burn your boards.

I was not referring to this particular script/functionality, i.e. enabling all VRMs.

Bondus said that 1.00 do not apply voltage settings to all the VRMs even if they are all enabled, this is bad because the VRMs are supposed to work in couple by desing. Having different voltage settings on coupled VRMs could demage the HW.

ya he changed something to that script to make it work with 1.0

but i wonder why he didn't post it here yet?

sickpig

legendary

Activity: 1260

Merit: 1008

Quote from: elasticband on February 21, 2014, 07:58:06 PM

Quote from: bondus on February 16, 2014, 09:17:19 AM

I have inspected what the November firmware has actually changed.

What I can see it has changed a value in one of the registers in the VRMs. It has set the "ON" bit in the OPERATIONS register,changing what conditions are needed to turn the VRMs on.
This change will be permanent even if rebooted or power-cycled. It can be changed using i2c commands:

Turn all VRMs on:

#!/bin/sh
# Stop SPI poller
i2cset -y 2 0x71 2 0

for b in 3 4 5 6 7 8 ; do
   for d in 0 1 2 3 4 5 6 7 ; do
   i2cset -y $b 0x1$d 1 0x80 >/dev/null 2>&1
   done
done

Restore to factory settings:

#!/bin/sh
# Stop SPI poller
i2cset -y 2 0x71 2 0

for b in 3 4 5 6 7 8 ; do
   for d in 0 1 2 3 4 5 6 7 ; do
   i2cset -y $b 0x1$d 1 0x0 >/dev/null 2>&1
   done
done

I suggest stopping cgminer before doing this to limit the power when the operation is done.

Reading the datasheet for the VRMs I can see that when using them in pairs (as we do when all are turned on) there are some requirements on how they should be turned on. This method does not do that!

I have heard that hizzing sound as well. Some boards do that. They did that on 0.90 firmware too. No idea why.

I do not know if this is safe or not, it might be unsafe and burn your boards.

I was not referring to this particular script/functionality, i.e. enabling all VRMs.

Bondus said that 1.00 do not apply voltage settings to all the VRMs even if they are all enabled, this is bad because the VRMs are supposed to work in couple by desing. Having different voltage settings on coupled VRMs could demage the HW.

sickpig

legendary

Activity: 1260

Merit: 1008

Quote from: crashoveride54902 on February 21, 2014, 07:54:12 PM

Quote from: sickpig on February 21, 2014, 04:38:31 AM

Quote from: crashoveride54902 on February 20, 2014, 11:30:13 PM

Quote from: bondus on February 20, 2014, 06:31:06 PM

A warning:

For you brave souls that have enabled all 8 VRMS (early Octobers) I saw today that the 1.0.0 firmware does not apply the voltage adjustment to all 8 VRMs. It only applies it to the 4 that are officially enabled. The old FW applied this adjustment to all VRMs.

The VRMs work in pairs and if the voltage adjustment (VOUT_TRIM register) of the 2 VRMs in a pair is different things will go bad.

If you still want to use all 8 VRMs I suggest you read and understand the datasheet of the VRM.
The VRMs are very resilient to abuse, but they have limits.

Other than that FW 1.0.0 works very good. HW errors have dropped a lot for me. Good chips are close to 0% errors (2% before), bad chips still close to 2%.

THanks to his help via IRC I was able to update fw and still use my 2 8vrm boards...and my hw error droped down to .11% from .25% yay hopefully it stays that low as i just restarted...Thxs bondus

do you care to share bondus help?

i would post his script if he gave me permission too...he's not on IRC so i can't ask him Sad

sorry

don't worry. I appriciated your attitude, and obviously it would be amazing if bondus grant you the permission to share such a script Tongue

Topic: Hacking The KNC Firmware: Overclocking - page 27. (Read 144343 times)