CCminer(SP-MOD) Modded GPU kernels.

[sp_]

new build 39
GTX 750 1480core/1567mem testing with --cpu-priority 5 on yaamp:
rates in khs
qubit: ~4350
quark: ~5650
fresh: ~3450
x11: 2859
x13: 2248
x14: 2214
x15: 1937
lyra2: 862
sha256 : 193950 (didn't wait for yay's )
keccak on NiceHash:
150mhs

Good rates for a non TI card.
I think the kernals needs to be finetuned for each algo. When I set optimal launch configs in the quark kernals it slows down the the x15 since they share kernals.
I might need to duplicate kernal calls to be able to tweak the optimal hashrate for every algo.

[duboisi]

For both R38 and R39, 750 Ti get error "Cuda error in func 'x11_simd512_cpu_init' at line 634 : out of memory." when running x11 in HamsterPool. No such error with x11 in YAAMP.

Edit: Upon further testing, running up to 3x 750 Ti has no error (i.e.: only 4x and beyond get the above error). To run 4x and beyond, I have to use -i 18 to run fine. Error could be unique to HamsterPool only.

Edit2: 'x11_simd512_cpu_init' <-- corrected

[rednoW]

For both R38 and R39, 750 Ti get error "Cuda error in func 'x11_xim512_cpu_init' at line 634 : out of memory." when running x11 in HamsterPool. No such error with x11 in YAAMP.

hmmm, 750 without Ti works fine on hamster. Maybe you should try to play with --diff parameter.

[Schleicher]

For both R38 and R39, 750 Ti get error "Cuda error in func 'x11_xim512_cpu_init' at line 634 : out of memory." when running x11 in HamsterPool. No such error with x11 in YAAMP.

xim512 ? You mean simd512?

[bathrobehero]

Good job!

ccminer r39 on 750 Ti at +150 core and stock memory solomining to avoid vardiff fluctuation:

fresh 3770
groestl 8200
jackpot 5900
keccak 170000
lyra2 713
nist5 9900
quark 6100
qubit 4780
s3 9400
x11 3150
x13 2500
x14 2450
x15 2120
bitcoin 216000

spreadminer v5: 1730

On a sidenote, starting ccminer with --cpu-priority 5 sometimes freezes the PC without ever starting hashing so I'm not using it and I manually set the process priority to realtime after it started.

[bronan]

My GTX 780 limited to 80% power
core 1097 mhz / memory 3100
at hamsterpool :

x11: 4613 kh
x11 bluecoin performance 5312 kh

after new case and better cooling i might test full speed its limited to run maximum temp of 70c now
pretty darn good speed already with so much limits
oops lol different ccminer
lol
the new 39 crashes


[2015-02-15 20:46:46] Starting Stratum on stratum+tcp://eu.hamsterpool.com:7773
[2015-02-15 20:46:46] NVAPI GPU monitoring enabled.
[2015-02-15 20:46:46] 1 miner thread started, using 'x11' algorithm.
[2015-02-15 20:46:46] Stratum difficulty set to 0.00904417
Cuda error in func 'x11_simd512_cpu_init' at line 641 : invalid texture reference.

[sp_]

For both R38 and R39, 750 Ti get error "Cuda error in func 'x11_sim512_cpu_init' at line 634 : out of memory." when running x11 in HamsterPool. No such error with x11 in YAAMP.
Edit: Upon further testing, running up to 3x 750 Ti has no error (i.e.: only 4x and beyond get the above error). To run 4x and beyond, I have to use -i 18 to run fine. Error could be unique to HamsterPool only.

I get this error sometimes too on rigs with 6-7 cards, or with a 960 card. The default intensity is a bit high. Ho ever if you restart the miner it usually works.  It seems like the driver is freeing up som memory when the out of memory exception is triggered.

Workaround: Restart the miner a couple pof times until it  starts, or lower the intensity. -i 18

[bronan]

lol

i just wanted to ask whats going on
have tried starting 25 times with -i between 2 and 18 not starting


[2015-02-15 20:55:31] Intensity set to 8, 256 cuda threads
[2015-02-15 20:55:31] Starting Stratum on stratum+tcp://eu.hamsterpool.com:7773
[2015-02-15 20:55:31] NVAPI GPU monitoring enabled.
[2015-02-15 20:55:31] 1 miner thread started, using 'x11' algorithm.
[2015-02-15 20:55:31] Stratum difficulty set to 0.00904417
Cuda error in func 'x11_simd512_cpu_init' at line 641 : invalid texture reference.

another 20 times tried not starting

[sp_]

lol
i just wanted to ask whats going on

Your card is compute 3.5 and this miner is compiled for compute 5.0 and higher. You will have to compile a version yourself. (change compute version in the project file to 3.5)
But not sure how fast it will run though...

[bronan]

ahh only maxwell this version
wait are you sure the other ccminer 5.0 runs perfect

[sp_]

ahh only maxwell this version
wait are you sure the other ccminer 5.0 runs perfect

My mod will work, but I didn't compile it for you card in the exefile. This is a maxwell mod, and all the improvements are only tested on maxwell cards. I don't know how the performance is on older cards. It's probobly slower than the other ccminer forks.

[bronan]

I am not sure but this card runs proper on all ccminer50.exe files, but its not indeed not a new maxwell its the CK110 gpu

[hefty]

New build with more hashing power in most algos.

from release 38 we have

-faster keccak
-faster fugue
-faster hamsi
-faster aes
-faster echo
-some bugfixes

1.5.39(sp-MOD) is available here: (15-feb-2015)

https://github.com/sp-hash/ccminer/releases/tag/1.5.39

The sourcecode is available here:

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

thanks man keccak is faster by 15000 to 17000 khash on 970 from realase 38, thanks for advice you give me a week ago about crashing driver but nothing worked i think that mjollnir algo is broken in all ccminer realases for 970 not just yours. I have only mjollnir crashing and few other people from this forum pmd me that have same problem all of them have 970gtx.

[sp_]

Tonight I will try to use the BFI instruction to modify pointers, I will allign sharemem to 1024 bytes boundaries. and insert the 8 bit offset in bits 2-10 directly into the pointer.

Today a byteperm+shift+add is needed to calculate the sharedmem adress
(echo/shavite)
Hopefully AES will be faster.

3 instructions down to 1 instruction.

In echo alone there are 5120 random sharemem accesses. If this works, 1 round of Echo(hashalgo #11 in x11) will use around 10240 less instructions to do the same work. IMHO massive improvement.

In my bitslice groestl improvement i removed around 1000 instructions and the x11  hash got 50-100KHASH faster on the 750ti.
With 10 000 instructions removed, perhaps 500KHASH faster is possible.

But I assume 32bit pointers, so it might not work on 64bit builds. And it might not work in cuda code alone. Perhaps I need to rewrite AESround function to asm.
And the BFI might be slower than other instructions, so less instructions but more expensive instructions.

anyway. will try it out.

[scotte]

qubit has slowed from 3609 kH/s on the previous version to 3571 kH/s on the current one - GTX 850M Linux, using "git pull" to update, build.sh has -O3 passed to configure

[bronan]

New build with more hashing power in most algos.

from release 38 we have

-faster keccak
-faster fugue
-faster hamsi
-faster aes
-faster echo
-some bugfixes

1.5.39(sp-MOD) is available here: (15-feb-2015)

https://github.com/sp-hash/ccminer/releases/tag/1.5.39

The sourcecode is available here:

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

did you check how much memory is being needed for this, could it be the 970 mem limit at 3.5 Gb?

The 980 has 4 gb besides that there is not difference between 970 and 980 i think besides the 980 has more cores

still waiting for my watercooled 980 to be delivered

[djm34]

New build with more hashing power in most algos.

from release 38 we have

-faster keccak
-faster fugue
-faster hamsi
-faster aes
-faster echo
-some bugfixes

1.5.39(sp-MOD) is available here: (15-feb-2015)

https://github.com/sp-hash/ccminer/releases/tag/1.5.39

The sourcecode is available here:

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

did you check how much memory is being needed for this, could it be the 970 mem limit at 3.5 Gb?

The 980 has 4 gb besides that there is not difference between 970 and 980 i think besides the 980 has more cores

still waiting for my watercooled 980 to be delivered

those algo don't use much vram, it shouldn't be a problem (actually there isn't much algo which use more than 2Gb)

[antantti]

those algo don't use much vram, it shouldn't be a problem (actually there isn't much algo which use more than 2Gb)

Sorry sp_ for hijacking this thread but need to ask you and djm34, what are those algos that benefit from vram --> 2GB. And also, why some algos get a boost from memory oc and most
don't? Usually it is all about core clock, you can heavily downclock memory to save power and there is no difference in mining performance.

And if that 2GB doesn't matter what is holding back gtx 960 vs 750ti? On djm's neoscrypt miner 960 does +100% against 750ti, on other algos most of the time the difference is
somewhere in +60-90% area.

Sorry guys, computer enthusiast without coding skills just want's to learn...

[tbearhere]

#39 less hashrate on most algos. I'm still using the best one #33. But keccak was 162,000kh/s now with #39  170,000kh/s. 750ti.

[RtRaway]

Can you tell me a good video card for $370?
Tq