Topic: Hacking The KNC Firmware: Overclocking - page 9. (Read 144359 times)

I believe is the difference in your construction appliance.

If you had a thermal camera, you could see it.

Or measure (beam measure?) the heatsink and the vrm temps of all of them.
If you see great difference between vrm/heatsink , then there is bad thermal contact.

I am sick of KNC, need so much babysitting.

I'm now ready with modifying all my boxes (see pictures previous posts), and running on 500 Mhz (at periods shows 3.5 Ths but most of the time 3.4).

But I noticed something odd with the temps. Now that the metal plate is removed and each VRM has it's own heatsink, the very strange thing is that the right-side VRM's (1 to 4) are 10 full degrees hotter than the left-side (this is on all 5 boxes).  There is no significant difference in Volt, Amp and Watt.... so for some reason the temp between VRM 4 and 5 are 10 degrees apart... which is very strange as those two are just in front of main Fan ??
Anyone has an idea ?

I noticed the BBB 'seems/looks to me' maxed out.

Bandaids considered (some tried) to date,
nix lighttpd (WIMP interface is overated anyways if not needing VRM info)
run cgminer without screen
Might gain a single % @ CPU

OC the bone 10%.
Source for clock proggy
https://github.com/mindbeat/bclock

I think it is for BB White not BB Black it's old code.
Must verify values in bclock.h with BBB ARM datasheet (not done yet)

I compiled and ran it, set it to 1104000Hz, got all excited, looked at output from
'cat /sys/bus/cpu/devices/cpu0/cpufreq/cpuinfo_cur_freq' and lost excitement.

/sys/bus/cpu/devices/cpu0/cpufreq/cpuinfo_cur_freq could be stale and it worked, no idea.

Looked at load with top and still maxed out.
Did [spi1] and cgminer just gobble it up???

I kinda feel like slacking/playing so I may just measure BBB clock remotely with SDR dongle and see if it changes.
(evil semi-tangent monster strikes again)

I thought I was golden but honestly did not test long enough.
After reboot and for longer period default BBB clock approaches same percieved results.

I have no idea if the ASICs are gobbling up the MIPS.
My hunch is output from cgminer and the parade of not goodness scrolling by.
It would not be my first useless tangent BTW.
At some  point I'll throw a scope on the ribbon signals and look, no promise when though.

Lots has changed with the ASIC so we got the fun of discovery to anticipate!

Welp, thats where I'm at.

Still workin on cooling, spacing pitstops for highest UPTIME though as it takes a lot of GH/s improvement to improve on just lettin it run to begin with.

Thx folks for sharing early tests and results re: cooling

YMMV


P.S. To compile 'bclock' on the Nep:
I used the first part of the build script for bfgminer to install transient build environment on BBB.
Bfgminer build script is in README.ASIC in bfg source tree.
Had to change makefile to 'CC = arm-angstrom-linux-gnueabi-gcc'
BEWARE
'Status' webpage will show miner stopped. it is running, just a web page SNAFU due to build environment.
Your hashrate will drop ~20% during install of build environment, 10 minutes or so.
N'Joy

EDIT
SDR dongle ineffective to deduce CPU clock, possibly due to AC fans and nekkid ASIC PCBs nearby.
I have measured ASIC clocks with this method reliably in the past FWIW.
I am leaving it up for a while, it looks like more, temps have moved up a tad. Not linear 10% improvement.

EDIT # 2
I wouldn't swear to it but I think KnC already OC'd the BBB by exactly 10% in the kernel.
I definately found 1.104 GHz with the antenna next to CPU and did not see 1GHz.
The program to manipulate the stock CPU freqs is missing and the files I looked at may just be related unremoved remnants.

I also had a bit of excitement, the smoke escaped from a 2kW PSU (probably operator error)
resistance across DC outputs now 0 Ohm so more than a fuse ...
One spare PSU just paid for itself.

Too much fun!

As of late I've settled on 0x361 in my firmware, Ive have great luck with all boards running at 250Gh/s each and all units I've put together with 6 boards maintain 1.5Th/s.

The guide I wrote that has been distributed to many of the donators explains the fan configuration needed to keep them cool, it's a minimal investment for a huge difference in temperate, on the stock fans you can't go more than 750-800Gh/s before things start shutting down for thermal reasons..

also what watt from wall
for nov and october jupiter

at 940gh and 680  gh hahsing speed?

I think I have an easy (to fabricate) way to watercool the VRMs.

Aluminum plates on the VRMs and a copper U attached to the plates with JB Weld.
(JB weld is strong high temp CONDUCTIVE  epoxy, also OK thermal properties)

I'll never get the investment in cooling back, but radiators and WBs are reusable forever if water is treated correctly.
Gonna look into cheap car radiators for the next nep, possibly even used.

Can anyone think of something else that is kinda easy and not involve 15 @ USD$100 ram waterblocks.

Luckily pit stops only involve a 20% reduction in hashrate with nep.

I DID get a nep so I could meddle with the workings after all

For those who may be wondering WTF FOR, easy way to move the HEAT elsewhere via hose.
The ASIC are quite happy with the air cooled heatsinks, just add fans if needed.

The VRMs are a bad dream though.
A good ASIC surrounded by mediocrity.


YMMV

You can try setting a lower voltage like -0.0366 and watch if some cores throw hardware errors. If you get hw errs and disabled cores, try increasing the voltage one step at a time. Lower voltage gives you lower power draw and lower vrm temperatures. This guy achieved to gain 100 watts with that: https://bitcointalk.org/index.php?topic=313978.msg5796532%23msg5796532 (read post #898)

If you have good cooling, you can even try to go up with the frequency. I have some dies running with 0x361 (this is max setting) on my nov jupiter without errors.

335 is a wierd clock setting, did you try 331?

The last digit is a divisor and you are probably not dividing a clock by 5 and still reach 900MHz on a 28nm ASIC.
The PLL won't go that fast.
But if the field is only 2 bits wide 5 and 1 would be the same.
2 bit divisor is typical of many PLL.

YMMV

for nov. unit? 0.000 for volts and 0x335 for the oc...make sure you have extra cooling as the vrms will get hot

what option i must put at padrino firmware for have it 900 gh hashing speed?

 AFAIK the modules work exactly the same way but the 50a model has three extra settings on the PM Bus to control dynamic loop compensation which is not available on the lower rated modules.

well that sucks and it would have to depend on the 50a model having the same layout as the 40a model wouldn't it? i don't see why they would be different but ya never know

 You have no choice but to stop at the 50A module - that's the end of the BMR464 line.

 You have 5 boards with 8 VRMs on each.  Let's say you can get them for $30USD each (and I believe the cost is higher).  You're going to spend another 1200 dollars to get what?
I think you'd be better off buying two Bitcoins and mining with the OEM boards using inexpensive cooling modifications. 

 OR we crowd-fund an experiment and get one lucky miner to test your theory on just one of their 5 boxes.  Everyone not interested take a step back!

if you can surface mount then go for it...but that would add alotta cost onto the unit...and also why stop at the 50a model...why not go higher if yer gona do it? hehe more room for oc'in

Did some Google search but cannot find any place where I can buy these 1008s (I found 1002s, but those aren't cheap as well...)

Can't we just desolder the 1002 and remove them to be replaced by soldering some 1008s in their place?  Doesn't that mod seem like a good solution?  What do you think?

Very impressive mods! wish you lived close by..

if you did what elen did then yes warrenty gone..but taking off yeller sticker n putting yer own heatsink on it...something that can be made back to original shouldn't void yer warrenty i don't think upto you

warranty is gone anyway... but I guess my modifications make it more lasting than the original 107 degrees box...