CCminer(SP-MOD) Modded GPU kernels.

[chrysophylax]

I think the autotune of the launch config is not working for scrypt-n. You need to add -l 64x2 or something similar..

here is my commandline:

./ccminer -a scrypt-jane:1402845776,4,30 -l t25x2

uses a little more than 3GB

ok - i just read the next messages and you answered the question i asked earlier ...

how are we to know what settings run well with this? ...

sp? ... is there advice you can give on these? ... especially for the 750ti oc lp cards and the 9xx cards ...

#crysx

I took the settings from the nicehash page:

https://www.nicehash.com/index.jsp?p=news&id=41

But they seem to work for tpruvot fork only.
Probably sp's needs different switches.

ok ...

i guess its a matter of fiddling with parameters then ...

btw - ive skyped you and sent you a pm ... ...

#crysx

[chrysophylax]

I took the settings from the nicehash page:
https://www.nicehash.com/index.jsp?p=news&id=41
But they seem to work for tpruvot fork only.
Probably sp's needs different switches.

I think I will remove scrypt-n from the fork. To slow to be profitable.

would you support it whether it profitable or not? ... or are you just focussing on the 'main' algos sp? ...

i guess the less algos in the fork - the less worry you have with maintaining the algo ...

#crysx

[sp_]

yes. If it doesn't work. Bether to remove it.

[djeZo]

I think the autotune of the launch config is not working for scrypt-n. You need to add -l 64x2 or something similar..

here is my commandline:

./ccminer -a scrypt-jane:1402845776,4,30 -l t25x2

uses a little more than 3GB

ok - i just read the next messages and you answered the question i asked earlier ...

how are we to know what settings run well with this? ...

sp? ... is there advice you can give on these? ... especially for the 750ti oc lp cards and the 9xx cards ...

#crysx

I took the settings from the nicehash page:

https://www.nicehash.com/index.jsp?p=news&id=41

But they seem to work for tpruvot fork only.
Probably sp's needs different switches.

Does ~1.3khs on gtx 950; settings: -l t30x1 (uses almost 2 gig of memory) ~ 50% TDP

Cannot go more, runs out of memory.

But I am having issues with GTX 980 Ti, it runs out of memory very quickly, even on settings where it shouldn't (like -l t25x2) 

[myagui]

here is my commandline:

./ccminer -a scrypt-jane:1402845776,4,30 -l t25x2

uses a little more than 3GB

I don't have scrypt-jane:15 settings around, so I'm guesstimating what should work based on the launch parameters I have for scrypt-jane:16. By the way, this: scrypt-jane:1402845776,4,30 doesn't tell me the current nfactor, unless I have a miner at hand (which I don't). I am assuming you need nfactor 15 in this instance. I find it easier to just set a fixed nfactor and be done with it. During nfactor changes, one always needs to switch to different parameters anyhow, so it's not really adding any work.

So here's some settings for you to try out if you want. I haven't tested these at all, I'm just doing simple multiplication based on some nf16 settings I have on my archives.

./ccminer -a scrypt-jane:15 -L 8 -l t48x8
./ccminer -a scrypt-jane:15 -L 4 -l t48x4

One of these should hit a performance sweet spot, if not, perhaps try with -L 6 for both also.
Some settings will not work on Windows, and even those that do, run much slower there.

Given how it was tpruvot that originally added support for scrypt and scrypt variations to ccminer (from the cudaminer base), perhaps his thread is a better place to discuss further -> ccminer by tpruvot

[ldp5500]

I think the autotune of the launch config is not working for scrypt-n. You need to add -l 64x2 or something similar..

here is my commandline:

./ccminer -a scrypt-jane:1402845776,4,30 -l t25x2

uses a little more than 3GB

ok - i just read the next messages and you answered the question i asked earlier ...

how are we to know what settings run well with this? ...

sp? ... is there advice you can give on these? ... especially for the 750ti oc lp cards and the 9xx cards ...

#crysx

I took the settings from the nicehash page:

https://www.nicehash.com/index.jsp?p=news&id=41

But they seem to work for tpruvot fork only.
Probably sp's needs different switches.

Does ~1.3khs on gtx 950; settings: -l t30x1 (uses almost 2 gig of memory) ~ 50% TDP

Cannot go more, runs out of memory.

But I am having issues with GTX 980 Ti, it runs out of memory very quickly, even on settings where it shouldn't (like -l t25x2) 

750ti 1.26k CCMINER 1.6.6

[pallas]

here is my commandline:

./ccminer -a scrypt-jane:1402845776,4,30 -l t25x2

uses a little more than 3GB

I don't have scrypt-jane:15 settings around, so I'm guesstimating what should work based on the launch parameters I have for scrypt-jane:16. By the way, this: scrypt-jane:1402845776,4,30 doesn't tell me the current nfactor, unless I have a miner at hand (which I don't). I am assuming you need nfactor 15 in this instance. I find it easier to just set a fixed nfactor and be done with it. During nfactor changes, one always needs to switch to different parameters anyhow, so it's not really adding any work.

So here's some settings for you to try out if you want. I haven't tested these at all, I'm just doing simple multiplication based on some nf16 settings I have on my archives.

./ccminer -a scrypt:jane:15 -L 8 -l t48x8
./ccminer -a scrypt:jane:15 -L 4 -l t48x4

One of these should hit a performance sweet spot, if not, perhaps try with -L 6 for both also.
Some settings will not work on Windows, and even those that do, run much slower there.

Given how it was tpruvot that originally added support for scrypt and scrypt variations to ccminer (from the cudaminer base), perhaps his thread is a better place to discuss further -> ccminer by tpruvot

thanks!
nfactor is 15 indeed.
I've tried all your settings on both sp and tpruvot, none of them generated valid shares.
The only settings which worked so far is nicehash's on tpruvot.

EDIT: I'm on linux

[ldp5500]

I think the autotune of the launch config is not working for scrypt-n. You need to add -l 64x2 or something similar..

here is my commandline:

./ccminer -a scrypt-jane:1402845776,4,30 -l t25x2

uses a little more than 3GB

ok - i just read the next messages and you answered the question i asked earlier ...

how are we to know what settings run well with this? ...

sp? ... is there advice you can give on these? ... especially for the 750ti oc lp cards and the 9xx cards ...

#crysx

I took the settings from the nicehash page:

https://www.nicehash.com/index.jsp?p=news&id=41

But they seem to work for tpruvot fork only.
Probably sp's needs different switches.

Does ~1.3khs on gtx 950; settings: -l t30x1 (uses almost 2 gig of memory) ~ 50% TDP

Cannot go more, runs out of memory.

But I am having issues with GTX 980 Ti, it runs out of memory very quickly, even on settings where it shouldn't (like -l t25x2) 

750ti 1.26k CCMINER 1.6.6  LEO unprofitable, dead occupy most of the network operator force.

[djeZo]

here is my commandline:

./ccminer -a scrypt-jane:1402845776,4,30 -l t25x2

uses a little more than 3GB

I don't have scrypt-jane:15 settings around, so I'm guesstimating what should work based on the launch parameters I have for scrypt-jane:16. By the way, this: scrypt-jane:1402845776,4,30 doesn't tell me the current nfactor, unless I have a miner at hand (which I don't). I am assuming you need nfactor 15 in this instance. I find it easier to just set a fixed nfactor and be done with it. During nfactor changes, one always needs to switch to different parameters anyhow, so it's not really adding any work.

So here's some settings for you to try out if you want. I haven't tested these at all, I'm just doing simple multiplication based on some nf16 settings I have on my archives.

./ccminer -a scrypt:jane:15 -L 8 -l t48x8
./ccminer -a scrypt:jane:15 -L 4 -l t48x4

One of these should hit a performance sweet spot, if not, perhaps try with -L 6 for both also.
Some settings will not work on Windows, and even those that do, run much slower there.

Given how it was tpruvot that originally added support for scrypt and scrypt variations to ccminer (from the cudaminer base), perhaps his thread is a better place to discuss further -> ccminer by tpruvot

Thanks, this was helpful. Playing with L parameter now and got my GTX 980 Ti to 2.5 kHs now. Will play a bit more...

[myagui]

Typo alert, corrected below and on my original post:

./ccminer -a scrypt-jane:15 -L 8 -l t48x8
./ccminer -a scrypt-jane:15 -L 4 -l t48x4

[...] perhaps try with -L 6 for both also.

[djeZo]

What does -L stand for anyway? Some tech explanation would be great.

[djeZo]

~3.6 khs  with following params: -L 16 -l t128x4

[myagui]

thanks!
nfactor is 15 indeed.
I've tried all your settings on both sp and tpruvot, none of them generated valid shares.
The only settings which worked so far is nicehash's on tpruvot.

EDIT: I'm on linux

Hmm. My settings are based off the 980, though for this algo I'd expect that same settings would work for your 970.
One last shot, just for the 970, you could lower the thread count (replace 48 with 32, if still not working, 24).

[myagui]

What does -L stand for anyway? Some tech explanation would be great.

https://github.com/cbuchner1/CudaMiner/blob/master/README.txt#L131

Ideally, one should run with low'ish -L values, but there's different sweet spots for maximum performance vs best efficiency.
A higher -L gives you the most performance potential for a given compute capability, especially useful in high end cards. A lower -L gives you the most efficiency, and it is especially useful in lower end cards.

[bathrobehero]

Last I checked scrypt-jane in SP's fork didn't work.

LEOcoin was stupidly profitable a few months ago but for a few weeks now it's not that profitable and now with nicehash adding it, it will probably keep diving.
The next n-factor increase will be in 38 hours though (n-factor 16) so don't spend too much time trying to find the best n-factor 15 config.

As a tip though, the fastest kernel launch configs are based on how many SM units a card have:
750 Ti's have 5 so the fastest launch config will look like t5x?, 970's have 13 (so t13x?), 980 have 16 and the 980 Ti have 22. With the correct lookup-gap of course.

I'm using an old version of sp_'s fork (somewhere around commit 800) and added a few basic command line arguments that are helping with scrypt-jane mining like being able to start autotuning from a set amount instead of having to wait from the beginning (so you can start autotuning from like t13x1), added another one that allows setting how much GPU memory ccminer can use (which is 88% by default on windows and 98% on linux), reintroduced the batchsize parameter since it was replaced with api-bind which slightly speeds the hashrate up. Just basic things saving some time and allowing to change stuff without having to recompile each time.
If someone need it just let me know.

For n-factor 15 I get:
750 Ti: 1.26khs (-L 5 -l t5x16)
970: 3.27khs (-L 10 -l t13x24)

N-factor 16:
750 Ti: 0.53khs (-L5 -l t5x15)
970: 1.15 (-L 14 -l t13x28)

These are for windows with oc, linux.

[djeZo]

What does -L stand for anyway? Some tech explanation would be great.

https://github.com/cbuchner1/CudaMiner/blob/master/README.txt#L131

Ideally, one should run with low'ish -L values, but there's different sweet spots for maximum performance vs best efficiency.
A higher -L gives you the most performance potential for a given compute capability, especially useful in high end cards. A lower -L gives you the most efficiency, and it is especially useful in lower end cards.

That explains it why GTX 980 Ti likes L=16

[myagui]

As a tip though, the fastest kernel launch configs are based on how many SM units a card have:
750 Ti's have 5 so the fastest launch config will look like t5x?, 970's have 13 (so t13x?), 980 have 16 and the 980 Ti have 22. With the correct lookup-gap of course.

Right on point bathrobehero. I neglected the SMM/SMX difference between the 980 and 970 upthread.

[djeZo]

So I found best params.

GTX 980 Ti: -L 16 -l t22x23
GTX 950: -L 7 -l t6x32

What are best ones for 970 and 980? Don't have these cards...

[sambiohazard]

Here are my configs with cudaminer for

970: --algo=scrypt-jane:15 -L 8 -i 1 -l t64x4 -C 1 -b 65536

750ti: --algo=scrypt-jane:15 -H 2 -i 0 -m 1 -l t5x24 -L 8 -b 65536

i don't remember the hashrates but they were what most of other people were getting.

[sp_]

Submitted another speedup in quark/x11. +50-100KHASH on the gtx 970.

Not sure about the other cards. Please test..

(More of blake512 - 80 was  precalculated.)