Hacking The KNC Firmware: Overclocking

[Biffa]

thanks ! Don't you think that jupiter clocked on 325 mhz is too dangerous for a long run ?

Not dangerous, but it does void your warranty, which is why I would NEVER overclock my KnC hardware, but I have a purely scientific interest in these find people investigating the inner workings of the hardware.

But if I was, theoretically, going to overclock my KnC hardware then the requirements of getting as many coins as quickly as possible would probably override my concerns about danger or warranty becuase by the time they actually started to be a concern, well, with the difficulty increases we have had recently, the equipment would be mining dust by then really.

[1714219028]

1714219028

[mekadeka]

you have to minus the percentage of errors and rejected shares from your hash rate to get hash rate that pool should show.

eg.  if your hash rate is 300GH/s and you have 2% errors and 1% rejected then 300-9 = 291GH/s

Thanks. Ok, where do I find errors? The rejects are .13%. Also, I'm not following your formula. Your example is 2% errors and 1% rejects (3%) subtracted from the hashrate (300). That's 297, not 291.

That said, if I followed your formula, I'd be less than 1% plus presumably 30% errors (326 - .13% - [30% errors]. That seems incorrect?

[Biffa]

you have to minus the percentage of errors and rejected shares from your hash rate to get hash rate that pool should show.

eg.  if your hash rate is 300GH/s and you have 2% errors and 1% rejected then 300-9 = 291GH/s

Thanks. Ok, where do I find errors? The rejects are .13%. Also, I'm not following your formula. Your example is 2% errors and 1% rejects (3%) subtracted from the hashrate (300). That's 297, not 291.

That said, if I followed your formula, I'd be less than 1% plus presumably 30% errors (326 - .13% - [30% errors]. That seems incorrect?

You would have to get the information from cgminer on the device.

So either SSH onto the box or use the API and a third part App, or use Bertmod and it will show the Hardware errors (HW) on the web page.

[mekadeka]

You would have to get the information from cgminer on the device.

So either SSH onto the box or use the API and a third part App, or use Bertmod and it will show the Hardware errors (HW) on the web page.

Thank you. I'm able to SSH in, but I'm not clear on how to view errors once there. The "screen -r" view is this (keep in mind I just went down to 231, so the hashrate is reporting this stepped down as opposed to my original comments):

 cgminer version 3.9.0 - Started: [2014-01-24 15:25:51]
--------------------------------------------------------------------------------
 (5s):304.4G (avg):311.1Gh/s | A:159374  R:2242  HW:4403  WU:4346.0/m
 ST: 2  SS: 0  NB: 8  LW: 176865  GF: 0  RF: 1
 Connected to stratum-lb-usa48.btcguild.com diff 128 with stratum as user qweqweqweqweqwe_qweqweqweqweqwe34
 Block: 8c5d58f4...  Diff:2.19G  Started: [16:02:36]  Best share: 231K
--------------------------------------------------------------------------------

 [P]ool management [ S ]ettings [ D ]isplay options [Q]uit
 KnC 0:                | 305.7G/311.4Gh/s | A:159502 R:2242 HW:4407 WU:4350.0/m

--------------------------------------------------------------------------------

[Biffa]

R: is rejected, and HW: is hardware errors. A: is accepted

[mekadeka]

R: is rejected, and HW: is hardware errors. A: is accepted

A:159374
R:2242
= 1.406% rejections

How are HW errors quantified? How do I make sense of the 4403?

Thank you.

[Biffa]

R: is rejected, and HW: is hardware errors. A: is accepted

A:159374
R:2242
= 1.406% rejections

How are HW errors quantified? How do I make sense of the 4403?

Thank you.

In the same way as rejections.

Out of the Accepted, x% were rejected and x% were hardware errors.

[merv77]

cgminer version 3.9.0 - Started: [2014-01-24 15:25:51]
--------------------------------------------------------------------------------
 (5s):304.4G (avg):311.1Gh/s | A:159374  R:2242  HW:4403  WU:4346.0/m
 ST: 2  SS: 0  NB: 8  LW: 176865  GF: 0  RF: 1
 Connected to stratum-lb-usa48.btcguild.com diff 128 with stratum as user qweqweqweqweqwe_qweqweqweqweqwe34
 Block: 8c5d58f4...  Diff:2.19G  Started: [16:02:36]  Best share: 231K
--------------------------------------------------------------------------------

 [P]ool management [ S ]ettings [ D ]isplay options [Q]uit
 KnC 0:                | 305.7G/311.4Gh/s | A:159502 R:2242 HW:4407 WU:4350.0/m

--------------------------------------------------------------------------------

Accepted:159374  Rejects:2242  HWHardWare Errors:4403
2242/159374*100=1.4% rejects
4403/159374*100=2.76% errors

1.4+2.76=4.16% total

304*4.16/100=12.65GH/s

304-12.65=291.35GH/s

[mekadeka]

In the same way as rejections.
Out of the Accepted, x% were rejected and x% were hardware errors.

Ah, thanks.
So of my miner's total work:

1.406% are rejections
2.76% are HW errors
=4.166% of empty work.

The above is based off my current 211 configuration, but this difference would not cover the much wider gap of machine reporting versus poolside of my previous 251. So, I think the test should be extended to 231 and then 251. Perhaps I'll try temus's latest configuration, which is different altogether.

What is WU, by the way?
Thank you.

[Biffa]

In the same way as rejections.
Out of the Accepted, x% were rejected and x% were hardware errors.

Ah, thanks.
So of my miner's total work:

1.406% are rejections
2.76% are HW errors
=4.166% of empty work.

The above is based off my current 211 configuration, but this difference would not cover the much wider gap of machine reporting versus poolside of my previous 251. So, I think the test should be extended to 231 and then 251. Perhaps I'll try temus's latest configuration, which is different altogether.

What is WU, by the way?
Thank you.

From https://github.com/ckolivas/cgminer/blob/master/README

Code:

The output line shows the following:
(5s):1713.6 (avg):1707.8 Mh/s | A:729 R:8 HW:0 WU:22.53/m

Each column is as follows:
5s: A 5 second exponentially decaying average hash rate
avg: An all time average hash rate
A: The total difficulty of Accepted shares
R: The total difficulty of Rejected shares
HW: The number of HardWare errors
WU: The Work Utility defined as the number of diff1 shares work / minute
     (accepted or rejected).

 BAS 1: max 67C 3.27V | 62.29G/62.19Gh/s | A:140813 R:256 HW:2860 WU: 852.0/m

Each column is as follows:
Temperature (if supported)
Fanspeed (if supported)
A 5 second exponentially decaying average hash rate
An all time average hash rate
The total difficulty of accepted shares
The total difficulty of rejected shares
The number of hardware erorrs
The work utility defined as the number of diff1 shares work / minute

The cgminer status line shows:
 ST: 1 SS: 0 NB: 1 LW: 8 GF: 1 RF: 1

ST is STaged work items (ready to use).
SS is Stale Shares discarded (detected and not submitted so don't count as rejects)
NB is New Blocks detected on the network
LW is Locally generated Work items
GF is Getwork Fail Occasions (server slow to provide work)
RF is Remote Fail occasions (server slow to accept work)

The block display shows:
Block: 0074c5e482e34a506d2a051a... Started: [17:17:22] Best share: 2.71K

This shows a short stretch of the current block, when the new block started,
and the all time best difficulty share you've found since starting cgminer
this time.

[mekadeka]

Accepted:159374  Rejects:2242  HWHardWare Errors:4403
2242/159374*100=1.4% rejects
4403/159374*100=2.76% errors

1.4+2.76=4.16% total

304*4.16/100=12.65GH/s

304-12.65=291.35GH/s

Ah, we were doing the same thing at the same time. The second part of the figuring is correct (but I didn't include it), but remember that this is for my current settings at 211, to which I reverted this AM after the big discrepancy I was finding at 251. I'll try again and test. Thanks!

[tolip_wen]

Ideas on cooling

I put air obstructions in my case so all air must pass the ASIC and VRM's.

For Sat find/make a cardboard box that fits between front (towards BBB) ASIC modules.
For Jup also find/make 2 smaller boxes to fill up outside edges.

Also the air that passes over the aluminum plate is not doing much cooling I direct it towards the VRM's.

Painters masking tape or similar wrapped around the heat pipes below the aluminum radiator will direct more air to the VRM's.

You could shove a wadded paper towel in there for testing.
Beware cardboard/paper burns, might not be ideal for long term usage in a box with a history of sparks.

Folks who drop fans to bar might consider blocking top(no fan area) of aluminum radiator also.

Aluminum tape/foil might be a better choice but then beware risk of shorts, keep it neat!

Make sure the cables do not block airflow also.

Oct. GE VRM's are designed for airflow from ASIC side to edge of board side.
KnCMiner is not ideal for OC airflow but it is what it is.
I'm glad they chose shortest lowV/highA copper runs making OC possible!

YMMV

[redbluebg]

Here is how I have moved my fans, their are strapped with zip lock to the PCIe cables, temperature dropped with almost 20 degrees, now running under 40 degrees overclocked to 211, getting 485 GH/s. 231 and 251 are not working very good for my miner

https://i.imgur.com/yXdBhzT.jpg
https://i.imgur.com/aXkBuLb.jpg

[sickpig]

Ideas on cooling

I put air obstructions in my case so all air must pass the ASIC and VRM's.

For Sat find/make a cardboard box that fits between front (towards BBB) ASIC modules.
For Jup also find/make 2 smaller boxes to fill up outside edges.

Also the air that passes over the aluminum plate is not doing much cooling I direct it towards the VRM's.

Painters masking tape or similar wrapped around the heat pipes below the aluminum radiator will direct more air to the VRM's.

You could shove a wadded paper towel in there for testing.
Beware cardboard/paper burns, might not be ideal for long term usage in a box with a history of sparks.

Folks who drop fans to bar might consider blocking top(no fan area) of aluminum radiator also.

Aluminum tape/foil might be a better choice but then beware risk of shorts, keep it neat!

Make sure the cables do not block airflow also.

Oct. GE VRM's are designed for airflow from ASIC side to edge of board side.
KnCMiner is not ideal for OC airflow but it is what it is.
I'm glad they chose shortest lowV/highA copper runs making OC possible!

YMMV

Care to share same pics?

[DPoS]

Here is how I have moved my fans, their are strapped with zip lock to the PCIe cables, temperature dropped with almost 20 degrees, now running under 40 degrees overclocked to 211, getting 485 GH/s. 231 and 251 are not working very good for my miner

not sure on those temp readings.. I saw without the shroud you get less drop and tilting down kept more heat on the boards

are you 20 degrees less than stock config?


you might also be pointing directly onto the temp sensor.  I forget where that is but if it is working for you don't move it !

[temen]

@pedrosoft: Closed case. I have repositioned those "case" fans to blow to asics boards (leaning in front of asic fan), heatsinks etc. on vrmr's. Read my previous mails=) Now I'm having 471 GH/s with 2 modules. I have a feeling that the asic doesnt even get so hot, voltage regulator are what heats the whole thing.

Sometimes vrm:s start to dropt their amps to 24 amp, once the whole asic 0 was down to 24 amps, and hashing 415ghs. Stopped the miner, waited for vrms to cool and started again.

If you want to overclock, you cant start the miner on hot VRMs. You'll just get fault 96/97. VRMs needs to be cool when starting the miner when overclocking! If they are not your miner will settle for lower hashrate for rest of the session or those VRM-faults.

[redbluebg]

Here is how I have moved my fans, their are strapped with zip lock to the PCIe cables, temperature dropped with almost 20 degrees, now running under 40 degrees overclocked to 211, getting 485 GH/s. 231 and 251 are not working very good for my miner

not sure on those temp readings.. I saw without the shroud you get less drop and tilting down kept more heat on the boards

are you 20 degrees less than stock config?


you might also be pointing directly onto the temp sensor.  I forget where that is but if it is working for you don't move it !

It is minus 20 degree from overclocked, before my temps were around 57-60 now there are 37 to 40, both times with the same config 211. now the whole box feels cooler, the vrms are cooler, and the whole radiator is cooler, I think the readings are real.

[DPoS]

It is minus 20 degree from overclocked, before my temps were around 57-60 now there are 37 to 40, both times with the same config 211. now the whole box feels cooler, the vrms are cooler, and the whole radiator is cooler, I think the readings are real.

nice, when I did something similar it didn't seem to do well without the shrouds is why I put them on the crossbar
maybe you have it at a better angle than I did

mine run pretty cool as well but 20c is a big drop

[captain-v]

Here is how I have moved my fans, their are strapped with zip lock to the PCIe cables, temperature dropped with almost 20 degrees, now running under 40 degrees overclocked to 211, getting 485 GH/s. 231 and 251 are not working very good for my miner

https://i.imgur.com/yXdBhzT.jpg
https://i.imgur.com/aXkBuLb.jpg

Below are the images from OCT 4 VRMs with fans blowing about 45 Degree to the boards like Redbluebg doing. Running 231 with stock settings...

https://i.imgur.com/fLwkotI.jpg
https://i.imgur.com/jFLMBXt.jpg

Based on these images I am not sure if adding heatsinks to VRMs be really beneficial. There is a danger of reducing the air flow in areas really needs cooling.

Does anyone know where is the temperature sensor for the ASICs?






1BFkmxwdjH4PM8X2YUK8HTuN1nnQn4bwct

[redbluebg]

Here is how I have moved my fans, their are strapped with zip lock to the PCIe cables, temperature dropped with almost 20 degrees, now running under 40 degrees overclocked to 211, getting 485 GH/s. 231 and 251 are not working very good for my miner

https://i.imgur.com/yXdBhzT.jpg
https://i.imgur.com/aXkBuLb.jpg

Below are the images from OCT 4 VRMs with fans blowing about 45 Degree to the boards like Redbluebg doing. Running 231 with stock settings...

https://i.imgur.com/fLwkotI.jpg
https://i.imgur.com/jFLMBXt.jpg

Based on these images I am not sure if adding heatsinks to VRMs be really beneficial. There is a danger of reducing the air flow in areas really needs cooling.

Does anyone know where is the temperature sensor for the ASICs?






1BFkmxwdjH4PM8X2YUK8HTuN1nnQn4bwct

What temps do you get on the chip itself? From the web interface and from your readings?  And what hash rate do you get per board?
If there is a better way to cool the down side of the VRMs, we'd be able to achieve better performance, any ideas?