CCminer(SP-MOD) Modded GPU kernels.

[SS2006]

Very nice! I'm now seeing 6600-6700 KH/S on my GTX 970 OC'd by 200 on clock on X11. (Original was 6100-6200), so an extra 500 KH/S
Def the best ccminer I've seen for X11 (i only mine X11)
Thanks! I hope there is more improvement, donations coming soon

I should add the hash rate is not too stable at times, it does bounce around a lot from 6300 to 6700

[jpouza]

I have emailed some of you a version of the miner with compute 5.2

Please test and report in the thread.

If I forgot anyone please resend your email on pm.

JESUS!!!!WTF!!!
Almost 8MH/s@X11 on my 980 cards. WOW!!!
Keep up the awesome job man!

[jpouza]

Monster!


share image

[jjjordan]

EVGA GTX 970 SC ACX1 (stock)
X11 5900-6000
X13 4800-4900

[sp_]

Very nice! I'm now seeing 6600-6700 KH/S on my GTX 970 OC'd by 200 on clock on X11. (Original was 6100-6200), so an extra 500 KH/S
Def the best ccminer I've seen for X11 (i only mine X11)
Thanks! I hope there is more improvement, donations coming soon

I should add the hash rate is not too stable at times, it does bounce around a lot from 6300 to 6700

Try to mine with a higher diff. Each time the miner is finding a nounce the hashrate is dropping. You can alse see this in GPU-Z. I think the author of the fork tried to fix this. Thread issue... I am focusing on the kernals right now. I beleive I can push the 980 above 10MHASH. NVIDIA will sell a shitload of cards with my improved miner, ask them to donate a card for me.

[SS2006]

heck if you can push the 970 that far relatively might get you card myself

[SS2006]

Very nice! I'm now seeing 6600-6700 KH/S on my GTX 970 OC'd by 200 on clock on X11. (Original was 6100-6200), so an extra 500 KH/S
Def the best ccminer I've seen for X11 (i only mine X11)
Thanks! I hope there is more improvement, donations coming soon

I should add the hash rate is not too stable at times, it does bounce around a lot from 6300 to 6700

Try to mine with a higher diff. Each time the miner is finding a nounce the hashrate is dropping. You can alse see this in GPU-Z. I think the author of the fork tried to fix this. Thread issue... I am focusing on the kernals right now. I beleive I can push the 980 above 10MHASH. NVIDIA will sell a shitload of cards with my improved miner, ask them to donate a card for me.

the difficulty fixed it. I always thought the faster i saw yay the better. Even tho now I see consistent high rates (6700), the accepts are much slower, is that OK, am I still getting paid the same/more on the pool side?
Thanks man

[sp_]

the difficulty fixed it. I always thought the faster i saw yay the better. Even tho now I see consistent high rates (6700), the accepts are much slower, is that OK, am I still getting paid the same/more on the pool side?
Thanks man

Yes, you will be payed the same. Probobly abit more because higher stable hashrates.

[sp_]

I don't know if the 980 will ever beat a 290X in raw hashrate; power consumption, certainly, but not plain performance. I've got 8.2MH/s on low clocks now - getting a better card in a couple of days.

Still alot than can be done. Half of the Hashfunctions are still not modified in my mod. My compute 5.2 build is close to 8 MHASH with heavy overclock. 10MHASH is only 25% faster.

25% faster but 100% more profit for the miners.

happy hashing

[go6ooo1212]

@sp_
I've tested the new reliese all day on x13.
970 went up to 5050 kH/s ; 980 reaches 5960 kH/s
The losses are about 2.4-2.5%

[sp_]

Today I have integrated the changes from Schleicher that he posted on the ccminer thread into my mod:

Quote from: Schleicher on October 19, 2014, 05:07:14 PM
I managed to increase the quark and nist5 speed a little bit.
Source code:
https://github.com/KlausT/ccminer

simd512 seems to be 10% faster. 
My stock 750TI is approaching 2700 without overclock. and 3000 with oclock.

This is without the faster Keccak (created by nvidia) in cudaminer.

These are the optimized numbers so far.

Blake         xxx (not done)
skein    1.5%
BMW       60%
jh512    4.5%
keccac    1%
cubehash: 7.5%
shavite: 3.6%
simd512: 9,2%
fuge:   4,70%      
hamsi:  6.97% 
shabal: 22%   
wirlpool:1.87%   
echo: 5.5% 
luffa: 0.4%

3 coders have contributed to the new speedup.

The sourcecode will be checked into the blakecoin fork by Epsylon3

https://bitcointalk.org/index.php?topic=770064.0

[Amph]

Very nice! I'm now seeing 6600-6700 KH/S on my GTX 970 OC'd by 200 on clock on X11. (Original was 6100-6200), so an extra 500 KH/S
Def the best ccminer I've seen for X11 (i only mine X11)
Thanks! I hope there is more improvement, donations coming soon

I should add the hash rate is not too stable at times, it does bounce around a lot from 6300 to 6700

hey that's good, can you tell me the consumption?

[djm34]

Today I have integrated the changes from Schleicher that he posted on the ccminer thread into my mod:

Quote from: Schleicher on October 19, 2014, 05:07:14 PM
I managed to increase the quark and nist5 speed a little bit.
Source code:
https://github.com/KlausT/ccminer

simd512 seems to be 10% faster.  
My stock 750TI is approaching 2700 without overclock. and 3000 with oclock.

This is without the faster Keccak (created by nvidia) in cudaminer.

These are the optimized numbers so far.

Blake         xxx (not done)
skein    1.5%
BMW       60%
jh512    4.5%
keccac    1%
cubehash: 7.5%
shavite: 3.6%
simd512: 9,2%
fuge:   4,70%      
hamsi:  6.97%  
shabal: 22%   
wirlpool:1.87%   
echo: 5.5%  
luffa: 0.4%

3 coders have contributed to the new speedup.

The sourcecode will be checked into the blakecoin fork by Epsylon3

https://bitcointalk.org/index.php?topic=770064.0

That uint2 keccak was done by mtrlt, if I'm not mistaken.

actually using it makes no difference on x11 it is already the fastest kernel of the bunch by far, also for some strange reason it does not work with compute 5.2 (registers gets confused   had some weird issues when testing where some variables weren't updated at all...)

[sp_]

Have you investigated whatever is going on with jackpotcoin?  I emailed you what was going on when I tried mining it.  It seemed like a nice improvement, but the pools only saw about 1/8 of the reported hash in the cmd prompt.

In the first version the blake implementation was commented out for jackpoint coin. 1/8 of the reported hash. What are the odds for that? The chained cryptohash is missing one of the algorithms and still reports found nounces. Is this the killer Blake?

[sp_]

Today I have integrated the changes from Schleicher that he posted on the ccminer thread into my mod:

Quote from: Schleicher on October 19, 2014, 05:07:14 PM
I managed to increase the quark and nist5 speed a little bit.
Source code:
https://github.com/KlausT/ccminer

simd512 seems to be 10% faster. 
My stock 750TI is approaching 2700 without overclock. and 3000 with oclock.

This is without the faster Keccak (created by nvidia) in cudaminer.

These are the optimized numbers so far.

Blake         xxx (not done)
skein    1.5%
BMW       60%
jh512    4.5%
keccac    1%
cubehash: 7.5%
shavite: 3.6%
simd512: 9,2%
fuge:   4,70%      
hamsi:  6.97% 
shabal: 22%   
wirlpool:1.87%   
echo: 5.5% 
luffa: 0.4%

3 coders have contributed to the new speedup.

The sourcecode will be checked into the blakecoin fork by Epsylon3

https://bitcointalk.org/index.php?topic=770064.0

That uint2 keccak was done by mtrlt, if I'm not mistaken.

// Experimental Kernel for Kepler (Compute 3.5) devices
// code submitted by nVidia performance engineer Alexey Panteleev

https://github.com/cbuchner1/CudaMiner/blob/35984c723eb786d614158daca9b07ac20de8645d/nv_kernel2.cu

[sp_]

Today I managed a few percent in groestl!

All the kernals in x11 have now been optimized, but there is more potential.

[djm34]

Today I have integrated the changes from Schleicher that he posted on the ccminer thread into my mod:

Quote from: Schleicher on October 19, 2014, 05:07:14 PM
I managed to increase the quark and nist5 speed a little bit.
Source code:
https://github.com/KlausT/ccminer

simd512 seems to be 10% faster.  
My stock 750TI is approaching 2700 without overclock. and 3000 with oclock.

This is without the faster Keccak (created by nvidia) in cudaminer.

These are the optimized numbers so far.

Blake         xxx (not done)
skein    1.5%
BMW       60%
jh512    4.5%
keccac    1%
cubehash: 7.5%
shavite: 3.6%
simd512: 9,2%
fuge:   4,70%      
hamsi:  6.97%  
shabal: 22%   
wirlpool:1.87%   
echo: 5.5%  
luffa: 0.4%

3 coders have contributed to the new speedup.

The sourcecode will be checked into the blakecoin fork by Epsylon3

https://bitcointalk.org/index.php?topic=770064.0

That uint2 keccak was done by mtrlt, if I'm not mistaken.

// Experimental Kernel for Kepler (Compute 3.5) devices
// code submitted by nVidia performance engineer Alexey Panteleev

https://github.com/cbuchner1/CudaMiner/blob/35984c723eb786d614158daca9b07ac20de8645d/nv_kernel2.cu

It's been done before, in OpenCL. See Maxcoin's CL file.

hmm, except that nvidia miner was there first...  
(but, I think Reorder used it a lot)
And actually it makes no difference on compute 3.0 and 3.5 (might explain why it wasn't used anymore)

[sp_]

// Experimental Kernel for Kepler (Compute 3.5) devices
// code submitted by nVidia performance engineer Alexey Panteleev
https://github.com/cbuchner1/CudaMiner/blob/35984c723eb786d614158daca9b07ac20de8645d/nv_kernel2.cu

It's been done before, in OpenCL. See Maxcoin's CL file.

(Compute 3.5) is outdated. All kernals needs to be rewritten for maxwell for optimal performance. preferably in Assembly language. This work takes months.

[sp_]

My groestl fixes seems to increase x11 with around 20 khash on the 750ti. around 60khash on the 980. The groestl is now only the 3rd slowest of the x11 hashes.

[sp_]

Managed to fix quark and Nist5 now. I had to rollback my assembly blake and the optimalisations I merged from Schleicher.

750ti standard clock
quark/nist5
4950/ 8000

750ti oc.
quark/nist5
5600/9000

But all algorithms seems to work now.