CCminer(SP-MOD) Modded GPU kernels.

[tbearhere]

@sp could you change the retrys on pool disconnects to 15 seconds default please. thx

[1714932666]

1714932666

[1714932666]

1714932666

[1714932666]

1714932666

[1714932666]

1714932666

[Epsylon3]

yep, after some tests i can say these application clocks are not used at all on linux... could be some day...

the setting is indeed changed in nvidia-smi -q but doesnt really change the device clocks (i also tried cudaResetDevice to be sure after changes)

http://docs.nvidia.com/deploy/nvml-api/group__nvmlDeviceCommands.html#group__nvmlDeviceCommands_1gc2a9a8db6fffb2604d27fd67e8d5d87f

[djm34]

new numbers from lyra (work still in progress):

gtx980: 1900kh/s
gtx780ti: 2400kh/s
gtx750ti: 990kh/s

nice, my gtx750 with latest sp_ mod can do only 905khs. (1510/1600) Performance is memory constrained ...

actually mine with latest sp_mod mine (1365/2900) does 770kh/s... now at 1031kH/s  (same setting)
gtx780ti 2588kh/s
gtx980 1900kh/s (no change)

[flipclip]

yep, after some tests i can say these application clocks are not used at all on linux... could be some day...

the setting is indeed changed in nvidia-smi -q but doesnt really change the device clocks (i also tried cudaResetDevice to be sure after changes)

http://docs.nvidia.com/deploy/nvml-api/group__nvmlDeviceCommands.html#group__nvmlDeviceCommands_1gc2a9a8db6fffb2604d27fd67e8d5d87f

From that nvmlDeviceCommands page, it looks like it is only used by: "For Tesla products from the Kepler family."

On the page: http://docs.nvidia.com/deploy/nvml-api/nvml-api-reference.html#nvml-api-reference

1. NVML API Reference
Supported products:
Full Support
NVIDIA GeForce Line: None

Limited Support
NVIDIA GeForce Line: All current and previous generation GeForce-branded parts

So only limited support, but it doesn't seem to list what is/isn't supported.

[Epsylon3]

yep, but look like its under work from nvidia, could appear soon. I will try to find beta versions and follow them.

but no, same behavior in 352.09 (released the 18 May)

[joblo]

what would you suggest for the higher level card? ... even if all we do is test with it ...

Titan X )))

I don't think so.

http://cryptomining-blog.com/tag/gtx-titan-x-hashrate/

Might be worth waiting for the 980ti.

[sp_]

new numbers from lyra (work still in progress):
gtx980: 1900kh/s
gtx780ti: 2400kh/s
gtx750ti: 990kh/s

nice, my gtx750 with latest sp_ mod can do only 905khs. (1510/1600) Performance is memory constrained ...

actually mine with latest sp_mod mine (1365/2900) does 770kh/s... now at 1031kH/s  (same setting)
gtx780ti 2588kh/s
gtx980 1900kh/s (no change)

good job. More interested in your quark optimalizations so I submitted another one @github.

[rednoW]

Titan X )))

I don't think so.

http://cryptomining-blog.com/tag/gtx-titan-x-hashrate/

Might be worth waiting for the 980ti.

I said in context that Titan X should be able to change frequencies with nvml api. Someone should test this guess )))

[sp_]

-Faster quark
-Faster groestlcoin
-Faster aes (tiny speedup in most algos)
-added --gpu-memspeed and --gpu-engine but it only works on windows and for some gpus (nvidia-smi limitation)


1.5.51(sp-MOD) is available here: (29-may-2015)

https://github.com/sp-hash/ccminer/releases/tag/1.5.51

The sourcecode is available here:

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

Post your stats here. Card name/gpu clock/memclock

[djm34]

-Faster quark
-Faster groestlcoin
-Faster aes (tiny speedup in most algos)
-added --gpu-memspeed and --gpu-engine but it only works on windows and for some gpus (nvidia-smi limitation)


1.5.51(sp-MOD) is available here: (29-may-2015)

https://github.com/sp-hash/ccminer/releases/tag/1.5.51

The sourcecode is available here:

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

Post your stats here. Card name/gpu clock/memclock

how faster is the new groeslt ?
By the way there is a bug in ccminer.cpp which is at the origin of the duplicate share:
the line reseting the nonce value when checking if the target has changed in miner_thread

newer number for lyra:
gtx980:   1990kh/s
gtx750ti: 1140kh/s
gtx780ti: 2787kh/s (lol doesn't seem to want to stop increasing...)

[sp_]

speeds build 51 (factory standard clocks)

groestl:
gtx 970(gigabyte oc): 23.300 MHASH
gtx 750ti(gigabyte windforce): 7.560 MHASH

quark:
gtx 970(gigabyte oc): 16.200 MHASH (up from 15.700 in build 50)
gtx 750ti(gigabyte windforce): 5,7 MHASH

[sp_]

newer number for lyra:
gtx980:   1990kh/s
gtx750ti: 1140kh/s
gtx780ti: 2787kh/s (lol doesn't seem to want to stop increasing...)

I suspect you have done something like:

1. Reduced the register count
2. changed launchconfig like in cudaminer/the cryptonight miner -l 8x60 etc.

am I right?

[joblo]

-Faster quark
-Faster groestlcoin
-Faster aes (tiny speedup in most algos)
-added --gpu-memspeed and --gpu-engine but it only works on windows and for some gpus (nvidia-smi limitation)


1.5.51(sp-MOD) is available here: (29-may-2015)

https://github.com/sp-hash/ccminer/releases/tag/1.5.51

The sourcecode is available here:

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

Post your stats here. Card name/gpu clock/memclock

Quark seeing 6120 KH/s, up from 6050, on EVGA GTX750ti SC LP no aux pwr, default clocks and ccminer parms, Centos 6.

I noticed you don't include sm52 in the makefile. Any particular reason?

[sp_]

Quark seeing 6120 KH/s, up from 6050, on EVGA GTX750ti SC LP no aux pwr, default clocks and ccminer parms, Centos 6.

I noticed you don't include sm52 in the makefile. Any particular reason?

Add this to the makefile:

nvcc_ARCH += -gencode=arch=compute_52,code=\"sm_52,compute_52\"

you could try to only compile for 52. On old drivers it was faster to compile for 52 only than for 50 and 52 mixed

[joblo]

Quark seeing 6120 KH/s, up from 6050, on EVGA GTX750ti SC LP no aux pwr, default clocks and ccminer parms, Centos 6.

I noticed you don't include sm52 in the makefile. Any particular reason?

Add this to the makefile:

nvcc_ARCH += -gencode=arch=compute_52,code=\"sm_52,compute_52\"

you could try to only compile for 52. On old drivers it was faster to compile for 52 only than for 50 and 52 mixed

Thanks, I already compile for 35/50/52. Compile time is not a concern of mine. I was just wondering if you
had a specific reason for leaving it out, Seems not.

[sp_]

It will use less power, but run slower. (if you run compute 5.0 builds on compute 5.2 cards)

I added 5.2 to the makefile now..

My mod is optimized for 32bit and wondows, but I try to keep the linux compability.

[Epsylon3]

Quark seeing 6120 KH/s, up from 6050, on EVGA GTX750ti SC LP no aux pwr, default clocks and ccminer parms, Centos 6.

I noticed you don't include sm52 in the makefile. Any particular reason?

He didnt wrote this part. sm52 is not supported in official Cuda 6.5 (They made a special 6.5.19 at the 970/980 release date)

Due to bugs in Cuda 7.0 related to our code.... we stay on that and sm52 is not enabled by default

[joblo]

Quark seeing 6120 KH/s, up from 6050, on EVGA GTX750ti SC LP no aux pwr, default clocks and ccminer parms, Centos 6.

I noticed you don't include sm52 in the makefile. Any particular reason?

He didnt wrote this part. sm52 is not supported in official Cuda 6.5 (They made a special 6.5.19 at the 970/980 release date)

Due to bugs in Cuda 7.0 related to our code.... we stay on that and sm52 is not enabled by default

In case anyone gets tripped up with sm52 here's a link to cuda 6.5 with 9xx support.

https://developer.nvidia.com/cuda-downloads-geforce-gtx9xx

[djm34]

newer number for lyra:
gtx980:   1990kh/s
gtx750ti: 1140kh/s
gtx780ti: 2787kh/s (lol doesn't seem to want to stop increasing...)

I suspect you have done something like:

1. Reduced the register count
2. changed launchconfig like in cudaminer/the cryptonight miner -l 8x60 etc.

am I right?

no, I never really looked at the launch config as it is done in cudaminer or cryptonight.
Actually, it is a full rewrite of lyra2

[Grout]

speeds build 51 (factory standard clocks)

quark:
gtx 970(gigabyte oc): 16.200 MHASH (up from 15.700 in build 50)
gtx 750ti(gigabyte windforce): 5,7 MHASH

Do you have power consumption metrics for these? Specifically the 970. I'm trying to decide if they are worth it or if I should stick to 750 Tis...