CCminer(SP-MOD) Modded GPU kernels.

[Angora]

did some testing with Lyra2 n Quark on 3 versions of modded ccminer

DJM34 VTC0.1    Lyra2: 477K-540K                   Quark 2147K-2482K

SP 1.5.1           Lyra2: 500K-580K                    Quark 2230K-2273K     
 
sp release 52:   Lyra2 only 1 accept for 590K      Quark 2116K-2366K

Testing time 5 minutes for each on the donate settings

Card is EVGA 750ti clock n memory  speed 1268M PCIE 2.0   OS Windows 8.1 64bit   

V52 was giving me lots of rejects from extranonce..  Now I recall reading in a sgminer thread that extranonce can cause the miner to show hash as failed but it most likely was accepted on the pool end.   The solution was to use a command line argument to turn off extranonce. something like --noextranonce.

thats strange how in v52 spmod lyra2 was bombing out ...


we had the tests run about an hour ago - and lyra was fine - same with quark and neoscrypt - x11 and x13 ...

x15 was running fine - but the nicehash stratum dropped the link - as it usually does on the us stratum ...

what settings did you use for lyra2 on v52? ... i have only the intensity parameter active -i 16.5 ( -i 22.5 is max but many cpu validation errors ) ...

#crysx

Have not used intensity just the difficulty....
ccminer.exe -a lyra2  -o stratum+tcp://donate-sp.granitecoin.com:7939/ -u GKarB5.nv1 -p x --diff 128"

I will try it with the intensity toggle....

[1714922636]

1714922636

[chrysophylax]

did some testing with Lyra2 n Quark on 3 versions of modded ccminer

DJM34 VTC0.1    Lyra2: 477K-540K                   Quark 2147K-2482K

SP 1.5.1           Lyra2: 500K-580K                    Quark 2230K-2273K      
  
sp release 52:   Lyra2 only 1 accept for 590K      Quark 2116K-2366K

Testing time 5 minutes for each on the donate settings

Card is EVGA 750ti clock n memory  speed 1268M PCIE 2.0   OS Windows 8.1 64bit    

V52 was giving me lots of rejects from extranonce..  Now I recall reading in a sgminer thread that extranonce can cause the miner to show hash as failed but it most likely was accepted on the pool end.   The solution was to use a command line argument to turn off extranonce. something like --noextranonce.

thats strange how in v52 spmod lyra2 was bombing out ...


we had the tests run about an hour ago - and lyra was fine - same with quark and neoscrypt - x11 and x13 ...

x15 was running fine - but the nicehash stratum dropped the link - as it usually does on the us stratum ...

what settings did you use for lyra2 on v52? ... i have only the intensity parameter active -i 16.5 ( -i 22.5 is max but many cpu validation errors ) ...

#crysx

Have not used intensity just the difficulty....
ccminer.exe -a lyra2  -o stratum+tcp://donate-sp.granitecoin.com:7939/ -u GKarB5.nv1 -p x --diff 128"

I will try it with the intensity toggle....

we dont set the diff - as the ccminer seems to adjust accordingly on its own ...

./ccminer -o stratum+tcp://donate-sp.granitecoin.com:7939/ -O chrysophylax.donate:x -a lyra2 -i 16.5 ...

thats all thats used on the commandline ...

seems to run nicely on that setting alone ...

#crysx

[Angora]

did some testing with Lyra2 n Quark on 3 versions of modded ccminer
DJM34 VTC0.1    Lyra2: 477K-540K                   Quark 2147K-2482K
SP 1.5.1           Lyra2: 500K-580K                    Quark 2230K-2273K     
  sp release 52:   Lyra2 only 1 accept for 590K      Quark 2116K-2366K
testing time 5 minutes for each on the donate settings
Card is EVGA 750ti clock n memory  speed 1268M PCIE 2.0   OS Windows 8.1 64bit   
V52 was giving me lots of rejects from extranonce..  Now I recall reading in a sgminer thread that extranonce can cause the miner to show hash as failed but it most likely was accepted on the pool end.   The solution was to use a command line argument to turn off extranonce. something like --noextranonce.

your quark numbers are way off. they should be around 6MHASH on the EVGA 750ti 300% faster than the numbers you get.

I measured by taking the accepted hash rate from the miner...
this is the command line ..... ccminer.exe  -o stratum+tcp://donate-sp.granitecoin.com:7919/ -O GKarB5:x -a quark -i 24

what parameters should I be using?

[Angora]

did some testing with Lyra2 n Quark on 3 versions of modded ccminer

DJM34 VTC0.1    Lyra2: 477K-540K                   Quark 2147K-2482K

SP 1.5.1           Lyra2: 500K-580K                    Quark 2230K-2273K      
  
sp release 52:   Lyra2 only 1 accept for 590K      Quark 2116K-2366K

Testing time 5 minutes for each on the donate settings

Card is EVGA 750ti clock n memory  speed 1268M PCIE 2.0   OS Windows 8.1 64bit    

V52 was giving me lots of rejects from extranonce..  Now I recall reading in a sgminer thread that extranonce can cause the miner to show hash as failed but it most likely was accepted on the pool end.   The solution was to use a command line argument to turn off extranonce. something like --noextranonce.

thats strange how in v52 spmod lyra2 was bombing out ...


we had the tests run about an hour ago - and lyra was fine - same with quark and neoscrypt - x11 and x13 ...

x15 was running fine - but the nicehash stratum dropped the link - as it usually does on the us stratum ...

what settings did you use for lyra2 on v52? ... i have only the intensity parameter active -i 16.5 ( -i 22.5 is max but many cpu validation errors ) ...

#crysx

Have not used intensity just the difficulty....
ccminer.exe -a lyra2  -o stratum+tcp://donate-sp.granitecoin.com:7939/ -u GKarB5.nv1 -p x --diff 128"

I will try it with the intensity toggle....

we dont set the diff - as the ccminer seems to adjust accordingly on its own ...

./ccminer -o stratum+tcp://donate-sp.granitecoin.com:7939/ -O chrysophylax.donate:x -a lyra2 -i 16.5 ...

thats all thats used on the commandline ...

seems to run nicely on that setting alone ...

#crysx

That was it!!   Replaced the diff with intensity 16.5 and it's running steadily @ 700K

[chrysophylax]

did some testing with Lyra2 n Quark on 3 versions of modded ccminer
DJM34 VTC0.1    Lyra2: 477K-540K                   Quark 2147K-2482K
SP 1.5.1           Lyra2: 500K-580K                    Quark 2230K-2273K      
  sp release 52:   Lyra2 only 1 accept for 590K      Quark 2116K-2366K
testing time 5 minutes for each on the donate settings
Card is EVGA 750ti clock n memory  speed 1268M PCIE 2.0   OS Windows 8.1 64bit    
V52 was giving me lots of rejects from extranonce..  Now I recall reading in a sgminer thread that extranonce can cause the miner to show hash as failed but it most likely was accepted on the pool end.   The solution was to use a command line argument to turn off extranonce. something like --noextranonce.

your quark numbers are way off. they should be around 6MHASH on the EVGA 750ti 300% faster than the numbers you get.

I measured by taking the accepted hash rate from the miner...
this is the command line ..... ccminer.exe  -o stratum+tcp://donate-sp.granitecoin.com:7919/ -O GKarB5:x -a quark -i 24

what parameters should I be using?

for quark - they are the exact parameters we use ...

for lyra2 - its obviously the parameters in the previous post ...

i can post the max settings ( and we really dont use much else but intensity -i setting ) if you prefer ...

no overclocks - no firmware tweaking - just straight v52 compile in fedora 20 x64 using cuda 6.5 ...

edit - 'That was it!!   Replaced the diff with intensity 16.5 and it's running steadily @ 700K' ...

glad that worked ... the optimizations that are 'stock' in this fork ( and others like tpruvot's fork ) are quite high intensities normally ...

so a little tweak factor for the card that you may have - could come out with a better result ...

#crysx

[joblo]

sorry if I get offended easily but some comments around (not necessarily yours)  sounds like that if it isn't in sp release it is crap... (or not optimized or whatever...)

If gave you that impression it was not my intent. There are several reasons the focus is on SP. First it's his thread
about his release. It's also the most popular fork at the moment and the center of most of the discussion. It is
also my impression that it is the most complete open source version of ccminer and that optimizations,
new kernels, etc, written by other devs will be merged by SP. If there are other open source forks that have advantages
over SP's I am not aware of them. Even if there are I don't think it would be fair to SP to hijack his thread to
discuss another fork of ccminer at length.

However, I do take notice when SP posts that he has imported code from another dev. And I do visit the gits of
these devs to see what else they are up to.

[chrysophylax]

sorry if I get offended easily but some comments around (not necessarily yours)  sounds like that if it isn't in sp release it is crap... (or not optimized or whatever...)

If gave you that impression it was not my intent. There are several reasons the focus is on SP. First it's his thread
about his release. It's also the most popular fork at the moment and the center of most of the discussion. It is
also my impression that it is the most complete open source version of ccminer and that optimizations,
new kernels, etc, written by other devs will be merged by SP. If there are other open source forks that have advantages
over SP's I am not aware of them. Even if there are I don't think it would be fair to SP to hijack his thread to
discuss another fork of ccminer at length.

However, I do take notice when SP posts that he has imported code from another dev. And I do visit the gits of
these devs to see what else they are up to.

same jo ...

delving into the other gits and compile the other fork versions for testing is a valuable tool for comparison ...

it is time consuming - but also gives a good indication as to whether the optimizations are at any one time ...

respect for devs is generally for MOST devs ... but use of any certain app - like ccminer - is based on its flexibility and optimization as well as its ease of use ( and ease of compile ) ...

we use a few different ccminer compiles - and always continue to compare ...

but ccminer-spmod does seems to be the most complete of the open source miners ...

we will still support the other devs in as much as we can do also ... and there is never any real intentional malice towards devs in general ...

not from us anyway ...

the latest of spmod commit is running on our farm at the moment - and running quite stable ...

#crysx

[joblo]

970 (GV-N970WF3OC-4GD - 250w OC edition instead of 145w):
stock - 2.75 mh/s at 187W
oc - 3.0 mh/s at 208W (+185/0 - 1501mhz)

I had assumed the 970 TDP would be within the 145W Nvidia spec except if OCed. It seems that
assumption was way off. I didn't find an actual TDP spec for this card, just PSU and connector reqs.
Is it really 250W? This changes the balance of power (bad pun) and makes me wonder about the 980
rated at 165W.

[djm34]

970 (GV-N970WF3OC-4GD - 250w OC edition instead of 145w):
stock - 2.75 mh/s at 187W
oc - 3.0 mh/s at 208W (+185/0 - 1501mhz)

I had assumed the 970 TDP would be within the 145W Nvidia spec except if OCed. It seems that
assumption was way off. I didn't find an actual TDP spec for this card, just PSU and connector reqs.
Is it really 250W? This changes the balance of power (bad pun) and makes me wonder about the 980
rated at 165W.

more or less same tdp as the 980 150w most likely (2x 6pin + pice alim)

[Angora]

For my card on lyra2... sweet spot is intensity 17.  within 1 minute it reached 705K and slowly climbing... 16.5  took close to an hr to reach 705k.   
One thing I did notice is that @ 16.5 memory use was 1.2M. Above that memory is 2.7M. 
Edit:   spoke too soon... seems 16.5 is it.  17 has become unstable. lots of stales and reconnects.


Name of the game is fastest rate and how quickly it reaches it.   

got neoscrypt working on spmod52.   Got rid of the difficulty.  Now to tweak the intensity.


 

[chrysophylax]

For my card on lyra2... sweet spot is intensity 17.  within 1 minute it reached 705K and slowly climbing... 16.5  took close to an hr to reach 705k.   
One thing I did notice is that @ 16.5 memory use was 1.2M. Above that memory is 2.7M. 
Edit:   spoke too soon... seems 16.5 is it.  17 has become unstable. lots of stales and reconnects.


Name of the game is fastest rate and how quickly it reaches it.   

got neoscrypt working on spmod52.   Got rid of the difficulty.  Now to tweak the intensity.


 

thats what we have done ...

all the intensities maxed out for our cards - gigabyte 750ti oc lp ...

i cant remember neoscrypt at the moment - but i will publish them all via the donation thread tomorrow ...

sleep soon for me for now ... ...

#crysx

[Angora]

With SP's mod 52, do not use --diff parameter.   Prior versions you could on some algos.    
 
My thanks to all for the help troubleshooting.

[5w00p]

970 (GV-N970WF3OC-4GD - 250w OC edition instead of 145w):
stock - 2.75 mh/s at 187W
oc - 3.0 mh/s at 208W (+185/0 - 1501mhz)

I had assumed the 970 TDP would be within the 145W Nvidia spec except if OCed. It seems that
assumption was way off. I didn't find an actual TDP spec for this card, just PSU and connector reqs.
Is it really 250W? This changes the balance of power (bad pun) and makes me wonder about the 980
rated at 165W.

I am by no means going to sit here and write as if I am an authority on this stuff.
However, I have taken note of something that I think deserves mention and discussion.

TDP is the term we all seem to use (myself included) to refer to and to determine the power consumption of a GPU.
But, TDP is  short for Thermal Design Power. Further definition is:
"TDP is the average power a device can dissipate when running real applications." (aka "normal" apps that an average user would run)

TDP, then, is NOT exactly equal to the device's maximum power consumption, nor is it necessarily measured at 100% load.

TDP is self-reported by manufacturers. It seems that several years ago, AMD & Intel used different percentages of processor loading to measure and report their CPU's TDP.
AMD used ~100% load and intel something like 80-85%. Well, intel's method seems to have become the norm.
This is what led to the definition of: average power dissipated when running "real applications."

So, under intense loading situations, a device can definitely consume more power and dissipate more heat than its TDP would indicate.
Of course, overclocking raises the amount of power consumed. And mining intensive algorithms certainly is not what would be called "real applications."

From http://www.cpu-world.com/Glossary/M/Minimum_Maximum_power_dissipation.html:

Maximum power dissipation is the maximum power dissipated by the CPU under the worst conditions - at the maximum core voltage, maximum temperature and maximum signal loading conditions.
Maximum Power dissipation is always higher than Thermal Design Power.

I believe that when mining the more intensive algos, these GTX 970 and 980 video cards are in fact consuming somewhat more power than their nvidia-stated TDP might indicate.

EDIT: Using an AC/DC clamp meter, I measured the actual current through the two 6-pin PCIe connectors on my stock BIOS Zotac GTX 970 w/ moderate overclock, while mining Quark algo with sp_mod release version 52.

One connector showed a relatively consistent 4.5A.
The other connector fluctuated between 2.5A and ~7A. I think it is safe to average it to 4.5A.

9A @ 12V = 108W + 75W (assumed) through the PCIe bus = 183 Watts

[bensam1231]

Bugreport for Neoscrypt. Running 5 970s on the same system causes it to spit out "Cuda error in func 'scanhash-neoscrypt' at line 67 : out of memory."

Limiting it to any 4 devices with -d eliminates the problem. Tried running separate instances with different devices, doesn't work. System memory is fine. DJM's miner has the same problem.

you need more memory... or pagefile... (which is required to allocate that memory to the gpu...)
alternatively you can lower the intensity, that should free up some memory
other alternative, try to start 1 or 2 ccminer session

As mentioned running more then one instance does not fix this, system memory is fine. It doesn't get maxed out or anywhere close to it while monitoring it. I'll throw some more memory into the system and see what happens though. Currently it's at 4GB.

Haven't tried a lower intensity, this seems more like a bug then anything if that's the case.

Memory is not necessary, pagefile will most likely work. I have 14GB (!) of pagefile on my rigs because I think there was a cudaminer algo that only ran if I had tons of pagefile but it never used more than a few megabytes.
Same goes for sgminer to a certain degree, although I was never able to run it with more than 4 cards enabled.

I will try a bigger pagefile. The pagefile sits at the smallest possible size when the miner is in use and starting up, just like memory. I don't know why this will fix it. My current pagefile already maxes out at 8192, but sits at 1024.

In order to run 5 AMD cards you need to be on W8.1. W7 only allows 4. 6 cards requires a lot of hacking about to make it work no matter what.

[antonio8]

Bugreport for Neoscrypt. Running 5 970s on the same system causes it to spit out "Cuda error in func 'scanhash-neoscrypt' at line 67 : out of memory."

Limiting it to any 4 devices with -d eliminates the problem. Tried running separate instances with different devices, doesn't work. System memory is fine. DJM's miner has the same problem.

you need more memory... or pagefile... (which is required to allocate that memory to the gpu...)
alternatively you can lower the intensity, that should free up some memory
other alternative, try to start 1 or 2 ccminer session

As mentioned running more then one instance does not fix this, system memory is fine. It doesn't get maxed out or anywhere close to it while monitoring it. I'll throw some more memory into the system and see what happens though. Currently it's at 4GB.

Haven't tried a lower intensity, this seems more like a bug then anything if that's the case.

Memory is not necessary, pagefile will most likely work. I have 14GB (!) of pagefile on my rigs because I think there was a cudaminer algo that only ran if I had tons of pagefile but it never used more than a few megabytes.
Same goes for sgminer to a certain degree, although I was never able to run it with more than 4 cards enabled.

I will try a bigger pagefile. The pagefile sits at the smallest possible size when the miner is in use and starting up, just like memory. I don't know why this will fix it. My current pagefile already maxes out at 8192, but sits at 1024.

In order to run 5 AMD cards you need to be on W8.1. W7 only allows 4. 6 cards requires a lot of hacking about to make it work no matter what.

I am no pro by any means but I had one rig with 5 AMD cards and had no issues with it at all on Windows 7.

I just didn't like having to change core or memory clocks for different algo's. Became to much of a job. Now that one is just down to a 280x and just waiting to see what the new 300 series does.

[bathrobehero]

TDP

TDP

Yes, it's not how much a card can pull but TDP does correlate fairly well to that amount (at least it did until the 900-series) and we don't have better figures from the specs. (Linus' explaination: https://www.youtube.com/watch?v=yDWO177BjZY )

I downloaded some random stock BIOSes from here and opened them in Maxwell II BIOS Tweaker to check their maximum TD.. I mean power consumption using this image to translate the figures and here are the results (in watts):

Code:

Bios				Flavor		Maximum	Target
Asus.GTX970.4096.141028		Strix OC	250	163.46-193.152
EVGA.GTX970.4096.141020		FTW		250	170-187
Galaxy.GTX970.4096.140912	EXOC		200	200-250
Gigabyte.GTX970.4096.141105	Windforce OC	250	250-280
Gigabyte.GTX970.4096.141910_1	G1 Gaming	250	250-280
MSI.GTX970.4096.141029		Gaming		250	200-220
NVIDIA.GTX970.4096.140826	Reference?	250	151.2-160.3
Palit.GTX970.4096.140903	Standard	250	151.2-160.3
Palit.GTX970.4096.140910	JetStream OC	250	180-200
PNY.GTX90.4096.140912		VCGGTX9704XPB	250	151.2-160.3
Zotac.GTX970.4096.141024	Standard	196	151.2-160-3
Zotac.GTX970.4096.141910	AMP Omega	350	325-345
...
Gigabyte GTX 980		G1 Gaming	250	300-366
Gigabyte GTX 980  		Windforce OC	250	270-300

I added two 980 (CBA to add more) to the list to show the figures does seem to be correct - at least compared to a stress test from a Tom's hardware review:



Ps: we should really have more threads instead of flooding this one with everything

[5w00p]

Very good info and great idea, bathrobehero.

GeForce GTX 900-series power consumption discussion thread: https://bitcointalk.org/index.php?topic=1091755.0

[Angora]

Hmmm..... sp mod 52 crashed when I tried algo x15 intensity 22.  

error msg :   cuda erron in func 'x11_simd512_cpu_init' line 614: unknown error.

testing to see if intensity setting changes hash rate. w/o it, hash is 3100 ksh

this is on EVGA 750ti running @ stock

edit:  error poped up again on intensity 21. this time it said out of memory.  mystery solved.

intensity 17 card keeps getting work but shows no hashing and then stratum fails and disconnects.

Intensity 19 is the best @ 3175 khs.  20 gave a net loss of 150k per hash vs no setting.

Does the PCI-e slot type have any bearing on hash rate?  It is PCI-ex16 2.0

[chrysophylax]

Hmmm..... sp mod 52 crashed when I tried algo x15 intensity 22.  

error msg :   cuda erron in func 'x11_simd512_cpu_init' line 614: unknown error.

testing to see if intensity setting changes hash rate. w/o it, hash is 3100 ksh

this is on EVGA 750ti running @ stock

edit:  error poped up again on intensity 21. this time it said out of memory.  mystery solved.

intensity 17 card keeps getting work but shows no hashing and then stratum fails and disconnects.

Intensity 19 is the best @ 3175 khs.  20 gave a net loss of 150k per hash vs no setting.

Does the PCI-e slot type have any bearing on hash rate?  It is PCI-ex16 2.0

no - the data that goes through teh pci slot is small for mining ...

ALL of our farm cards ( nvidia and amd ) run off x1 risers ...

so you should have no issues at all with data in and out of any of the pcie ports ...

the x15 algo seems to work with our ones at -i 25 - BUT - that does not mean it is the best setting for it ...

dont forget - nicethash have categorically stated ( on the donation thread ) that if there is no work orders on the site - the link gets dropped automatically ...

i am not sure if that has any bearing to ccminer crashing or not - but the lack of work to the miner itself 'may' have bearing on it IF there is an issue with ccminer / x15 not receiving any work or dropping links to the site ...

of course - these are just contemplations 'out there' - and not based on factual tests ...

#crysx

[hashbrown9000]

any linux guys having trouble with commit # 829?

If I try to run the newest miner, the command prompt drops to a "#" (pound sign) and no mining happens.  Reverting to an earlier version of the miner shows the same results as well.