CCminer(SP-MOD) Modded GPU kernels.

[scryptr]

I have now fixed the groestl speedup 2 on github. Klaus_t has been optimizing quark, so I am waiting for a pull request, and then build release 31.

SP_ BUILD v30a/b --

Version 30b is faster yet.  Both my rigs see an increase of about 1200-1300kh/s mining quark.  This difference is from v29 to v30b.

However, in the long-haul, 24hrs or so, build 30a/b is not stable on my slower rig.  Again, it is crashing.

I hope v31 comes soon, and with smooth running code.  My fast rig is now at 37.5mh/s, my slow rig at 36.2mh/s, and my fastest 750ti is at 6340kh/s (quark).       --scryptr

EDIT:  I've reverted my slower rig to v29 after a third crash.       --scryptr

[sp_]

Just removed 1296 assembly instructions from fugue hash, but something wrong with the byteorder of the registers. Boost in x13,x14,x15 comming soon

[sp_]

I have now fixed the groestl speedup 2 on github. Klaus_t has been optimizing quark, so I am waiting for a pull request, and then build release 31.

SP_ BUILD v30a/b --
Version 30b is faster yet.  Both my rigs see an increase of about 1200-1300kh/s mining quark.  This difference is from v29 to v30b.
However, in the long-haul, 24hrs or so, build 30a/b is not stable on my slower rig.  Again, it is crashing.
I hope v31 comes soon, and with smooth running code.  My fast rig is now at 37.5mh/s, my slow rig at 36.2mh/s, and my fastest 750ti is at 3275kh/s.       --scryptr
EDIT:  I've reverted my slower rig to v29 after a third crash.       --scryptr

Release 30 is running stable here. Could be that your cards get too hot? old driver? one sick card?

[sp_]

NVIDIA domination:



The people who are running ccminer/2014.06.15 could get a 40% boost if they switch to 1.5.30-git(SP-MOD). I guess someone has rented some tesla hashpower, and they are not smart enough to compile my fork for compute 3.0, 3.5.

[sp_]

NVIDIA supports the demoscene:

http://nv.scene.org/2015/

The demoscene support NVIDIA

www.pouet.net

[exemplaar]

Sp, did you tried new drivers 347.09?

Any improvements in hashrate towards 344.75?

Appreciate your work and dedication.

[scryptr]

I have now fixed the groestl speedup 2 on github. Klaus_t has been optimizing quark, so I am waiting for a pull request, and then build release 31.

SP_ BUILD v30a/b --
Version 30b is faster yet.  Both my rigs see an increase of about 1200-1300kh/s mining quark.  This difference is from v29 to v30b.
However, in the long-haul, 24hrs or so, build 30a/b is not stable on my slower rig.  Again, it is crashing.
I hope v31 comes soon, and with smooth running code.  My fast rig is now at 37.5mh/s, my slow rig at 36.2mh/s, and my fastest 750ti is at 3275kh/s.       --scryptr
EDIT:  I've reverted my slower rig to v29 after a third crash.       --scryptr

Release 30 is running stable here. Could be that your cards get too hot? old driver? one sick card?

Thanks--

I am running nVidia drivers 346.35, the latest for Linux.  My cards are in the mid 50 degree range.  This rig has the Sempron 145 unlocked to 2 cores.  v29 just crashed on me, also.  Since v28, stability has again been an off-and-on issue with this rig.  I suspect the CPU-workload, but there are too many loose variables.

I hope release 31 runs better.       --scryptr

[chrysophylax]

I have now fixed the groestl speedup 2 on github. Klaus_t has been optimizing quark, so I am waiting for a pull request, and then build release 31.

SP_ BUILD v30a/b --

Version 30b is faster yet.  Both my rigs see an increase of about 1200-1300kh/s mining quark.  This difference is from v29 to v30b.

However, in the long-haul, 24hrs or so, build 30a/b is not stable on my slower rig.  Again, it is crashing.

I hope v31 comes soon, and with smooth running code.  My fast rig is now at 37.5mh/s, my slow rig at 36.2mh/s, and my fastest 750ti is at 3275kh/s.       --scryptr

EDIT:  I've reverted my slower rig to v29 after a third crash.       --scryptr

scryptr - have you thought about a different ( possibly underlying ) issue with the slower rig?

i mean - some of my rigs would crash if i used the compiled versions of ccminer on other rigs ... when i compiled ccminer of the rig itself - it worked flawlessly ... no crashes - no issues - same version - just compiled locally ...

hence the reason why im 'standardizing' the farm with all the same components ... getting rid of the older motherboards ( which are nice giagbyte ones ) and rebuilding them into computers and rebuilding the farm with asrock h81-pro-btc motherboards with the same cpu and ram ...

i use the same version build of ccminer on ALL the machines - but some of them have locally compiled ones - as i get very similar issues to your slower machies if i dont compile locally - and some of them refuse to work at all ...

#crysx

[tbearhere]

I have now fixed the groestl speedup 2 on github. Klaus_t has been optimizing quark, so I am waiting for a pull request, and then build release 31.

SP_ BUILD v30a/b --

Version 30b is faster yet.  Both my rigs see an increase of about 1200-1300kh/s mining quark.  This difference is from v29 to v30b.

However, in the long-haul, 24hrs or so, build 30a/b is not stable on my slower rig.  Again, it is crashing.

I hope v31 comes soon, and with smooth running code.  My fast rig is now at 37.5mh/s, my slow rig at 36.2mh/s, and my fastest 750ti is at 3275kh/s.       --scryptr

EDIT:  I've reverted my slower rig to v29 after a third crash.       --scryptr

scryptr - have you thought about a different ( possibly underlying ) issue with the slower rig?

i mean - some of my rigs would crash if i used the compiled versions of ccminer on other rigs ... when i compiled ccminer of the rig itself - it worked flawlessly ... no crashes - no issues - same version - just compiled locally ...

hence the reason why im 'standardizing' the farm with all the same components ... getting rid of the older motherboards ( which are nice giagbyte ones ) and rebuilding them into computers and rebuilding the farm with asrock h81-pro-btc motherboards with the same cpu and ram ...

i use the same version build of ccminer on ALL the machines - but some of them have locally compiled ones - as i get very similar issues to your slower machies if i dont compile locally - and some of them refuse to work at all ...

#crysx

If you have the cards all synced together ....try unsyncing them. I cant run my cards synced.

[AdamWhite]

Hi, is there a possibility to improve the 750ti performance on CryptoNight algorithm?

Sorry if this has been discussed already.. thanks for sharing your work!

[djm34]

NVIDIA domination:



The people who are running ccminer/2014.06.15 could get a 40% boost if they switch to 1.5.30-git(SP-MOD). I guess someone has rented some tesla hashpower, and they are not smart enough to compile my fork for compute 3.0, 3.5.

lol, if you still expecting crypto people to be smart...

[AdamWhite]

Hi, is there a possibility to improve the 750ti performance on CryptoNight algorithm?

Sorry if this has been discussed already.. thanks for sharing your work!

Yes.

Thanks. I'm using tsiv's September 2014 release.. is there another (public) fork being developed?

[Dotcommie]

I have now fixed the groestl speedup 2 on github. Klaus_t has been optimizing quark, so I am waiting for a pull request, and then build release 31.

SP_ BUILD v30a/b --

Version 30b is faster yet.  Both my rigs see an increase of about 1200-1300kh/s mining quark.  This difference is from v29 to v30b.

However, in the long-haul, 24hrs or so, build 30a/b is not stable on my slower rig.  Again, it is crashing.

I hope v31 comes soon, and with smooth running code.  My fast rig is now at 37.5mh/s, my slow rig at 36.2mh/s, and my fastest 750ti is at 6340kh/s (quark).       --scryptr

EDIT:  I've reverted my slower rig to v29 after a third crash.       --scryptr

If you're getting that on even your slow rig I'm guessing you're overclocking them too much.  My 6xEVGA FTW cards can't be OC'd more than 10mhz over stock OC without crashing eventually and it's been that way since the beginning.  They get 36.55mh so I think you're pushing them too much.  I clocked the memory to -150 since it really has no effect on a lot of algos besides wasting power. Try dropping mem clocks to stock maybe?

My slow rig of Gigabyte 750ti's only get 33mh and it's a clone of the other rig with an even better i3 cpu, so I'm not sure why your numbers are quite a bit more than mine besides maybe OC.  Haven't tried much OC on the Gigabyte on quark yet.  I left it at +25mhz core and 100mhz memory I think since too high and other algos have a problem when it autoswitches to them. Rather have them work consistently 24hrs than have one algo crash them and lose half a day or more.

[scryptr]

I have now fixed the groestl speedup 2 on github. Klaus_t has been optimizing quark, so I am waiting for a pull request, and then build release 31.

SP_ BUILD v30a/b --

Version 30b is faster yet.  Both my rigs see an increase of about 1200-1300kh/s mining quark.  This difference is from v29 to v30b.

However, in the long-haul, 24hrs or so, build 30a/b is not stable on my slower rig.  Again, it is crashing.

I hope v31 comes soon, and with smooth running code.  My fast rig is now at 37.5mh/s, my slow rig at 36.2mh/s, and my fastest 750ti is at 6340kh/s (quark).       --scryptr

EDIT:  I've reverted my slower rig to v29 after a third crash.       --scryptr

If you're getting that on even your slow rig I'm guessing you're overclocking them too much.  My 6xEVGA FTW cards can't be OC'd more than 10mhz over stock OC without crashing eventually and it's been that way since the beginning.  They get 36.55mh so I think you're pushing them too much.  I clocked the memory to -150 since it really has no effect on a lot of algos besides wasting power. Try dropping mem clocks to stock maybe?

My slow rig of Gigabyte 750ti's only get 33mh and it's a clone of the other rig with an even better i3 cpu, so I'm not sure why your numbers are quite a bit more than mine besides maybe OC.  Haven't tried much OC on the Gigabyte on quark yet.  I left it at +25mhz core and 100mhz memory I think since too high and other algos have a problem when it autoswitches to them. Rather have them work consistently 24hrs than have one algo crash them and lose half a day or more.

SP_ v30a/b/c/d (four separate build changes to my count) --

Hello all!  Glad that several people replied to my posts yesterday and today.  I'll try to answer all...

First, I run two nearly identical Linux rigs, same MB,OS, different CPUS, slightly different cards.  Six cards per rig, the fast, reliable rig has 6x EVGA FTW 750ti cards, the slower rig has 6x EVGA SC 750ti cards.  The 750ti SC cards have no 6-pin PCI connector, and these cards run on a MB with a 2-core Sempron 145 (unlocked).

There is no Linux overclocking control that I know of, all my cards run at stock clocks.  I don't know how to synch my cards in Linux, either.  My faster, reliable rig has an AMD FX chip, 2 cores, and is capable of 3.2 GHz.  The unlocked Sempron 145 has 2 cores, and is capable of 2.8GHz.  They have equal cache (64/64/1024).

Right now the slowest of 12 cards is running at 6000kh/s, the fastest at 6340kh/s.  This is stock speed, and I am running on SP_ v30d, a fresh build within the last hour on both rigs.  My fast rig is at 37.60mh/s, my slow rig is at 36.35mh/s.  This is about 100kh/s better per rig than v30b. Both rigs are running smoothly.

SP_ build v30c was a disaster, my cards ran at 4+ mh/s and the rig crashed in short order.  I shut my slow rig off (early am) and went to bed.  My fast rig was running v30b smoothly, I just let it run.

As I was typing this, my slow rig crashed.  It had 232/235 accepts. I am going to reboot it and keep trying.       --scryptr

[sp_]

I merged with the klaus_t branch and the configuration was way off. I have now fixed the problem and restored performance.

I have tweaked the intensity parameters. Now the miner will work on the new cards as well. (gtx 960).

building release 31.

[sp_]

First, I run two nearly identical Linux rigs, same MB,OS, different CPUS, slightly different cards.  Six cards per rig, the fast, reliable rig has 6x EVGA FTW 750ti cards, the slower rig has 6x EVGA SC 750ti cards.  The 750ti SC cards have no 6-pin PCI connector, and these cards run on a MB with a 2-core Sempron 145 (unlocked).

swap 3 of the cards with a 6 pin power connector to the slow rig. Some motherboards doesn't work if you have 6 cards without powered raisers. How many watt is the PSU. and is it gold rated?

[scryptr]

First, I run two nearly identical Linux rigs, same MB,OS, different CPUS, slightly different cards.  Six cards per rig, the fast, reliable rig has 6x EVGA FTW 750ti cards, the slower rig has 6x EVGA SC 750ti cards.  The 750ti SC cards have no 6-pin PCI connector, and these cards run on a MB with a 2-core Sempron 145 (unlocked).

swap 3 of the cards with a 6 pin power connector to the slow rig. Some motherboards doesn't work if you have 6 cards without powered raisers. How many watt is the PSU. and is it gold rated?

All my cards have powered USB 3 risers, connected to the PSU.  None draw power from the MB directly on either rig.  The slow rig has a 600w bronze PSU, the fast one has a 750W bronze PSU.  Neither draws 500w at the wall.

I am thinking that I might lower the stock intensity.  It is a simple setting.  What are your current stock intensities for x11/x13/x14/x15/quark?       --scryptr

[sp_]

release 31 is here:

- improved groestl512 part2. faster x11,x13,x14,x15,nist5,quark etc.
- merged branch with klaus_t.  Faster quark + other fixes.
- fixed crash on 960 cards. (default intensity)

1.5.31(sp-MOD) is available here: (18-jan-2015)

https://github.com/sp-hash/ccminer/releases/tag/1.5.31

The sourcecode is available here:

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

[sp_]

All my cards have powered USB 3 risers, connected to the PSU.  None draw power from the MB directly on either rig.  The slow rig has a 600w bronze PSU, the fast one has a 750W bronze PSU.  Neither draws 500w at the wall.

I am thinking that I might lower the stock intensity.  It is a simple setting.  What are your current stock intensities for x11/x13/x14/x15/quark?       --scryptr

Try release 31. I have reduced the intensity for quark to save some memory. Performance was increased as well.

new default is:

(240* 4096)

old was:

// 256*4096  (-i 20)

[scryptr]

SP_MOD v31--

Initial impressions on v31 are good, however I haven't passed 1000 accepts on either rig.  I am getting 37.7mh/s on the fast rig, and 36.5mh/s on the slow rig.  This is an early readout, but indicates an improvement of at least 200kh/s per rig.

I am using KBomba's CCmonitor utility; it has a nice, easy interface.  It also does a standard deviation calculation on the hash rates, indicating how they vary as they are read.  What I am noticing here is that the standard deviation for my slow rig has decreased from roughly 2X the standard deviation for my faster rig.  My slow rig is now 36.5mh/s +/- 33.9kh/s, my fast rig is now 37.7mh/s +/- 27.9kh/s.  Earlier versions of SP_ ccminer running on my slow rig had a standard deviation of ~ +/-70kh/s.  My hope is that this is an indication of stable performance.       --scryptr