CCminer(SP-MOD) Modded GPU kernels.

[tbearhere]

yep, just fixed that :/ for this version just multiply by 3

will be in 1.4.8 (with a -N param to specify the number of samples for the average, if wanted)

im starting the work on the network json interface (like cgminer, but read-only to report some important stats)

Epsylon3 there is an improvement in hash. 

[tbearhere]

yep, just fixed that :/ for this version just multiply by 3

will be in 1.4.8 (with a -N param to specify the number of samples for the average, if wanted)

im starting the work on the network json interface (like cgminer, but read-only to report some important stats)

display is still showing ~10 mh/s per card on quark algo but really hashing at 5.1 mh/s. on 1.4.8

[Joe9]

Some GTX 970 numbers:

Gigabyte Geforce GTX 970 (non G1, triple fan model, stock clocks, and stock voltages)
Windows 7 64bit
8GB RAM
intel i5 something-k

X11 - 6600
x13 - 5200
x15 - 4500

This thing is whisper quiet compared to the 6950 it replaced. How much power is each algo using? For now I multiplied the 750Ti numbers by 2.2

[sp_]

I have a strange performance bug in my latest source(github). The hashrate on the 750ti sometimes performes 30KHASH higher in average. If you stop ccminer, and then restart a couple of times, you will see that the average hashoutput will sometimes report much higher numbers. I believe it has something to do with alligment in memory.

I have reproduced the problem in the BMW hash. (I comment out the 10 other hashing algos in x11, and run a perfomance check on the BMW alone). Start ccminer, time, stop ccminer, start ccminer stop etc..
you will see that sometimes the hashrate is 10% higher than normal, and will stay this way until the exe is shutdown..
I have tried everything, alligning memory buffers, const buffers. etc etc. The BMW is spilling to memory, so I believe this is related to some failed cachereads.

Any programmers out there that can explain this?

[djm34]

I have a strange performance bug in my latest source(github). The hashrate on the 750ti sometimes performes 30KHASH higher in average. If you stop ccminer, and then restart a couple of times, you will see that the average hashoutput will sometimes report much higher numbers. I believe it has something to do with alligment in memory.

I have reproduced the problem in the BMW hash. (I comment out the 10 other hashing algos in x11, and run a perfomance check on the BMW alone). Start ccminer, time, stop ccminer, start ccminer stop etc..
you will see that sometimes the hashrate is 10% higher than normal, and will stay this way until the exe is shutdown..
I have tried everything, alligning memory buffers, const buffers. etc etc. The BMW is spilling to memory, so I believe this is related to some failed cachereads.

increase the number of register

[djm34]

Secret super performing GPU Kernels like this are probably why GPU mining profitability fell off a cliff very quickly in mid August.

If anyone had to pay for a faster "secret" miner, he would not dump his coin below profitability.
Only those who sell low (with free software) kills the profitability.

So stop blaming others for being a weak hand.

[sp_]

increase the number of register

I already did. I also reduced the threads from 256 to 32. For more speed. But I need to understand why. Is variables allocated on the stack not alligned? The reason is that when the miner first starts with high hashrates it will hold this hashrate for hours. It looks like you have a 25% chance or less to get the best performance. Anyway, will work some more tonight to see if I can solve it.

[Epsylon3]

I think its more related to the gpu boost feature...

[djm34]

for which card is this ?
I have seen some rather strange issue with the 980...
There is a software on my mobie (network genie), yesterday I shut it down the gpu usgae went down by 20%...

[sp_]

Its on a gainward 750ti.

http://www.gainward.com/main/vgapro.php?id=926&lang=en

1085 MHz (base)
1163 MHz (boost)

Could be the boost that is sometimes on and sometimes off. How can I enable the boost clock always?

I need to dump some tables in cuda to file for debug/analysis. How can I do it quickly?

[djm34]

Its on a gainward 750ti.

http://www.gainward.com/main/vgapro.php?id=926&lang=en

1085 MHz (base)
1163 MHz (boost)

Could be the boost that is sometimes on and sometimes off. How can I enable the boost clock always?

I need to dump some tables in cuda to file for debug/analysis. How can I do it quickly?

boost is enable by default.
If you want to check without boost I think it is possible by forcing to run in p2 state (not obvious though if it will work... won't work for 9xx serie).

Alternatively, you could just monitor the core clock. If boost doesn't fire up, the core clock should be rather low in the 900MHz...
(compare with gpu-z should tell you if you are boosting)
Other alternative, Precision X allow to have a card using boost all the time even when idle...

[sp_]

But how can I dump tables to the commandline or to file within device code?
cuprintf doesn't work

With nsight I can set a breakpoint and "add watch" and copy values, but I want a faster way

Edit: found a ok way. I enter tablepointer,16 in the add watch window, and it shows the 16 first elements from the table

[sp_]

My next commit will give much hash.

[Epsylon3]

something interesting : http://epsy.linuxd.org/ccminer/sample.htm

Both are working on the same pool, first is windows with a MSI, second (epsytour) is linux with also the same MSI and a gigabyte black edition (both cards are reporting higher hashrate).

Check the Shares/minutes and accepted count... almost the same :/

[tsiv]

something interesting : http://epsy.linuxd.org/ccminer/sample.htm

Both are working on the same pool, first is windows with a MSI, second (epsytour) is linux with also the same MSI and a gigabyte black edition (both cards are reporting higher hashrate).

Check the Shares/minutes and accepted count... almost the same :/

That would be how vardiff works, higher hash rate leads to higher difficulty to hit the pool's configured share rate target. Or am I missing the point?

[Epsylon3]

indeed, so i need to add that to the stats

[jpouza]

My next commit will give much hash.

Can't wait!

X11 10MH/s@980 on the way 

[sp_]

you have to wait a bit longer, I have some problems with some of the mods. But I checked in +20KHASH @750ti by rewriting ECHO hash. Precalced some more and removed some calculations.


you can compile and check if you like.

The source is available here:

https://github.com/sp-hash/ccminer

I should probobly merge the latest changes in the main branch. (1.4.9) but I'm too lazy. My focus is on the kernals, and not the rest.

The work on the echo is not done. There is more to remove.

[tbearhere]

still very new at compiling, where can i get curl to 7.38.0 windows 8.1 from a safe site please.

[djm34]

still very new at compiling, where can i get curl to 7.38.0 windows 8.1 from a safe site please.

from the site of the author