Topic: CCminer(SP-MOD) Modded NVIDIA Maxwell / Pascal kernels. - page 1227. (Read 2347426 times)
My next commit will give much hash.
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
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
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. 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?
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...
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?
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?
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%...
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%...
I think its more related to the gpu boost feature...
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.
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.
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 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.
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?
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?
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
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
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 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)
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. 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)
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)
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)
There is something wrong, it's not adding the 3 cards hash, it counts as one only.
3x980@9MH/s = 27MH/s
Counts only 9MH/s with the 03 cards...
final 1.4.7 released : https://github.com/tpruvot/ccminer/releases/tag/1.4.7-tpruvot
Looks like the wolf miner(?) (AMD) has been published too some ppl. 10MHASH on 290x. I will checkin some more kernals soon, hopefully the 980 will pass 10MHASH as well.
https://bitcointalksearch.org/topic/x11-x13-x15-with-50-more-hashrate-7mhs-on-280x-or-10-mhs-on-290x-854257
https://bitcointalksearch.org/topic/x11-x13-x15-with-50-more-hashrate-7mhs-on-280x-or-10-mhs-on-290x-854257
Thanks for this. Is there any difference between ccminer-50 and ccminer-50-52 when running on a 750ti?
Good question, test and tell me
Seems to be variable against driver version, the 50 52 is for people with a 750Ti and a 9xx on the same MachineYep, runs great with both, 750 ti and 970's together!
Thanks for this. Is there any difference between ccminer-50 and ccminer-50-52 when running on a 750ti?
Good question, test and tell me
Seems to be variable against driver version, the 50 52 is for people with a 750Ti and a 9xx on the same MachineHash rate jumps around so much it's hard to tell. After a few hours on each I can't see any meaningful difference.
Thanks for this. Is there any difference between ccminer-50 and ccminer-50-52 when running on a 750ti?
Good question, test and tell me
Seems to be variable against driver version, the 50 52 is for people with a 750Ti and a 9xx on the same Machine Jump to: