Topic: SRBMiner Cryptonight AMD GPU Miner V1.9.3 - native algo switching - page 309. (Read 237247 times)

core by 5-10Mhz until it's no longer stable..then if you want put a little more volts into it until stable..
once you reach where both kill it.. go back about 10Mhz and maybe even lower volts a bit..if you are going
for extreme anyways, also watch for heat..but huge heat isn't a good thing..means more power wasted..
when sometimes you can get the same with clocking lower.

Nothing really can help that without the kernels that are executed. This also happens on any other gpu miner as well right?

For everyone:

     I'll explain a bit about OpenCL and how the "kernels" work a bit..
The kernel or "payload/work" is sent to the video card and is kind of like a singular pipe, okay that analogy sucked, but imagine
with all the "CU" Compute Units available.. the work from these kernels are sort of like threads on a miniature scale and they all
work according to the OpenCL standard (Hopefully) But it's like a faucet.. turn it on or off.. also you may need to turn down intensity
even more until you have a freeze free desktop which costs all the extra hashing power you would have had.  Remember on SRBMiner you can
use decimals so don't forget that ex: "intensity 20.3" or what have you.

     I've left an earlier comment on a "low intensity" mode that is about splitting
up one big time job to do work into much smaller timed jobs and if done right, it really shouldn't have any negative impact on hashing speed,
or at least any notable impacts on it so that could be maybe be a modulation of 100ms or 1/10th a second per command queue or even smoother
by doing it roughly ~16.67ms for 1/60th a second, but I'm getting a bit to technical trying to explain it.. Anywho, a good way to perhaps implement
this would be to take an argument for say --desktop_mode ex. --desktop_mode 25 and plug that in which would do 40 pulses_per_second.

     I'm not the most intelligible on OpenCL and gpu kernel work, but to the best of my knowledge this is the only way that claymore/phoenix miners are doing this,
please correct me if I'm wrong (Which I probably am on a few things ).  Since I have no source code to these miners, I've been doing a bit of work on XMRig-amd
and testing my findings out on that.. So far I haven't got into splitting up kernel execution times, but i did turn an almost static 32H/s into 35H/s by redoing a bit
of the way the kernel maps memory.. Not sure how big of an improvement it would be on a real card, but it probably would give a good bit more hashes on other cards..
I'm only able to test with an amd radeon HD6530D APU setup and the maximum i've ever got it to do was 42H/s at a slight bit of OC on claymore and around 50H/s if i jacked
timing up to almost 700Mhz from stock 444Mhz for this cards core..(Probably wondering to yourself how that is even achieved with these old APU's?? I call it part of the good
silicon lottery.) If you have an A6-series or A4's or whatnot, I might be able to also help with clocks and suggestions...

     Also if anybody would like to try and help coding for and helping SRBMiner be the serious go to for cryptonight mining with this low intensity thing and kernel work.. I'd gladly
accept the help and can set up a private git branch that we could work on that through xmrig source, if the Doktor gives an ok on doing, he could throw it into SRB with perhaps
minimal changes and nothing else will have it.  My goal as a mining hobbyist is to make these miners work on pretty much all driver versions (Except the ones like the crimson beta
which doesn't include anything to use far as i can tell...)

Drivers tested with using DDU (Without Format):
  14.4 - Confirmed as working and stable for pre-GCN cards.
  15.7 - Can only confirm that it does not work with A6-series Apu's without driver crash.
  15.11 - Also same as 15.7.
  16.2.1 - Forget about it... for pre-GCN computing.

     Perhaps I should just ask Doktor if he in any way, shape, or form would like to collaborate even?  I tend to use bitbucket for it's private git repositories with things such as these.
And by the way, here is something to talk about.. not sure if there is a rant section, but people that totally pack their miners and put in disassembly code that shuts off a miner for
having any debug software going, or show's faked hashrates to ppl who use nofee or cheats as they worded it, or are those numbers not just higher in the code.. SRB i'm 99.9999% sure
there isn't any of that since we're all getting just about the same hashes on all the open sourced ones, but this seems a good bit more optimized already with more share turn in's
compared to hashes, perhaps Dok already found those memory speedups Okay, enough of my novella.. Thanks.

Wow thank you very much. We can expect it in the next version?
Can you give me some advice for fine tuning my r7 370-s and rx 470s for all cryptonight versions, because on all types of algos, my desktop is little freezing, even on default settings, or even a much smaller intensity.

Дa, я тoжe нe зaмeтил пoтepи мoщнocти. Зaтo co вpeмeнeм пoявляeтcя мнoгo oтклoнeнныx шap "above target". Ha дpyгиx мaйнepax тaкoгo нeт.

Дa, я тoжe нe зaмeтил пoтepи мoщнocти. Зaтo co вpeмeнeм пoявляeтcя мнoгo oтклoнeнныx шap "above target". Ha дpyгиx мaйнepax тaкoгo нeт.

Hy вo пepвыx нeт никaкoй пoтepи мoщнocти, пo кpaйнeй мepe зa cyтки y мeня тoчнo нeт ни нa oднoй из фepм. Пo втopoмy мoжнo caмoмy нacтpoить пepeзaгpyзкy фepмы пo pacпиcaнию пpocтo

He ceкpeт, чтo пocлe нeкoтopoй paбoты тepяeтcя пpoизвoдитeльнocть, чacoв зa 6 нa 10%, пo мoeй cтaтиcтикe.
Peaлизyйтe aлгopитм, пpи пoтepe мoщнocти в зaдaнныx пpeдeлax, пepeзaгpyзкa или зaпycк фaйлa или пepeзaпycк пpoгpaммы.
Moнитopинг для кaждoй или выбpaннoй видeo кapты oтдeльнo.

It's no secret that after some work, the performance is lost, the clock is 6 by 10%, according to my statistics.
Implement the algorithm, if the power is lost within the specified limits, restart or run the file or restart the program.
Monitoring for each or the selected video card separately.