NVIDIA Kepler (K20) from 134MHash/s to 330MHash/s with CUDA

[charliemaggot]

That's the Microsoft runtime, but it was linked to the debug version so I've changed it to a release version.

You can download again and I've added the missing file in case you don't have the runtime installed.

edit: CUDA module isn't compiled...working on it.

edit2: I've added the CUDA ptx file, so you can download the latest binaries.

For for expediency, I've thrown in a "buildcuda.bat" in the root that will compile the ptx and put it in bin.

I'll tidy it all up tomorrow (or later today as they say).

[charliemaggot]

Windows binaries: I've added a "kepler" branch and rebuilt it. I don't have any suitable h/w, so if anyone could try it, that'd be grand.

https://github.com/cdmackie/rpcminer-mod/tree/kepler

[gateway]

I have a GTX 680.. just tried to run it but it spit this out..

Client will start 1 miner threads
Work will be refreshed every 4000 ms
1 CUDA GPU devices found
Setting CUDA device to first device found
Loading module bitcoinminercuda.ptx
Unable to load CUDA module: 209

then it exits..

[charliemaggot]

I have a GTX 680.. just tried to run it but it spit this out..

Client will start 1 miner threads
Work will be refreshed every 4000 ms
1 CUDA GPU devices found
Setting CUDA device to first device found
Loading module bitcoinminercuda.ptx
Unable to load CUDA module: 209

then it exits..

Did you use the master or kepler branch?
Does the file bitcoinminercuda.ptx exist in the same directory as the exe?
Can you try compiling the ptx yourself (run buildcuda)?

[psychocoder]

For GTX690 you must compile with sm_20. I work on a version which shuld be run better on Fermi GPUs and I have get the register count down from 63 to 34 but without any speedup :-(.

@charliemaggot: Thanks for building Win version.

[charliemaggot]

@charliemaggot: Thanks for building Win version.

Anytime. I just compiled it. Kudos for your coding it.

So for 209 errors, change buildcuda.bat

  -gencode=arch=compute_35,code=\"sm_35,compute_35\"

to

  -gencode=arch=compute_20,code=\"sm_20,compute_20\"

and run it. If you don't have VS2010, I can build it and throw it up.

[Barrel]

I'd like a compiled binary for the 209 error, I tried doing it on VS2012 and got a bunch of errors.

[charliemaggot]

I'd like a compiled binary for the 209 error, I tried doing it on VS2012 and got a bunch of errors.

Grab https://github.com/cdmackie/rpcminer-mod/blob/kepler/bin/bitcoinminercuda.20.ptx and rename it removing the ".20", replacing the other ptx file. Let me know how you get on.

[Barrel]

About the same hashrate on aggression=10 as guiminer on a GTX 650 Ti, 50-ish MH/s.

I'm not too familiar with the other settings (gpugrid and gputhreads) so I'm not sure if those would raise the hash rate a significant amount,

[charliemaggot]

About the same hashrate on aggression=10 as guiminer on a GTX 650 Ti, 50-ish MH/s.

I'm not too familiar with the other settings (gpugrid and gputhreads) so I'm not sure if those would raise the hash rate a significant amount,

I can't say, but psychocoder may well know better settings for you. Try info from his first post in this thread.

[gateway]

Ok got this working, any recommended settings ?

default settings:

Done allocating CUDA resources for (32,128)
Target = 00000000ffffffffffffffffffffffffffffffffffffffffffffffffffffff
68767 khash/s
73804 khash/s
74725 khash/s
74484 khash/s
73926 khash/s
73477 khash/s
73711 khash/s

cgminer does about 118-120~

some settings I found just testing but no idea if they are right for the 680 card
Client will start 1 miner threads
Work will be refreshed every 4000 ms
1 CUDA GPU devices found
Setting CUDA device to first device found
Loading module bitcoinminercuda.ptx
CUDA initialized
Done allocating CUDA resources for (256,512)
Finding best configuration step end (256,512) 1ms  prev best=9223372036854775807ms
Done allocating CUDA resources for (256,512)
Target = 00000000ffffffffffffffffffffffffffffffffffffffffffffffffffffff
113145 khash/s
119789 khash/s
126394 khash/s
119730 khash/s
120627 khash/s
122473 khash/s
122970 khash/s

-aggression=8 -gpugrid=256 -gputhreads=512

[psychocoder]

I think I update the code tomorrow (german time) and add multi sm support to the make system (linux) and code.

@Barrel: Please test with parameters -aggression=8 -gpugrid=64 -gputhreads=384

[Evan]

I think I update the code tomorrow (german time) and add multi sm support to the make system (linux) and code.

@Barrel: Please test with parameters -aggression=8 -gpugrid=64 -gputhreads=384

Hey so I am still having issues

Create Kernel rval=-46
error creating kernel

[charliemaggot]

I think I update the code tomorrow (german time) and add multi sm support to the make system (linux) and code.

@Barrel: Please test with parameters -aggression=8 -gpugrid=64 -gputhreads=384

Hey so I am still having issues

Create Kernel rval=-46
error creating kernel

That's OpenCl, not cuda, in case you are running the wrong one. Otherwise can you check you have copied the bitcoinmineropencl.cl from the src\opencl directory into where you are trying to run the exe.

[Barrel]

68 MH/s with those settings. A bit better, but not as drastic as I had hoped.

[Evan]

I think I update the code tomorrow (german time) and add multi sm support to the make system (linux) and code.

@Barrel: Please test with parameters -aggression=8 -gpugrid=64 -gputhreads=384

Hey so I am still having issues

Create Kernel rval=-46
error creating kernel

That's OpenCl, not cuda, in case you are running the wrong one. Otherwise can you check you have copied the bitcoinmineropencl.cl from the src\opencl directory into where you are trying to run the exe.

C:\Users\Glutony>C:\Users\Glutony\Desktop\rpcminer-mod-master\bin\rpcminer-mod-c
uda.exe -u Master-Galaxy7 -p *****-o http://localhost:8332
Client will start 1 miner threads
Work will be refreshed every 4000 ms
Could not retrieve work from RPC server.
CURL return value = 22
4 CUDA GPU devices found
Setting CUDA device to first device found
Loading module bitcoinminercuda.ptx
Unable to load CUDA module: 301


For BuildCuda.bat
With Cuda 5.0.5 installed:



C:\Users\Lab>"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v5.0\bin\nvcc.exe" -ptx        --use-local-env --cl-version 2010       -gencode=arch=compute_35
code=\"sm_35,compute_35\" -Xptxas=-v -Xopencc=-LIST:source=on  -m32    -D_BITCOIN_MINER_CUDA_  -DNVCC  -I"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v5.0\include"     -I"src\cuda"    -o "bin\bitcoinminercuda.ptx"   "bin\bitcoinminercuda.cu"
Nvcc : fatal error : Cannot find compiler 'cl.exe' in PATH

[InqBit]

I am getting a 200 error. I didn't see it covered above. I downloaded all the files in the bin folder into the same folder, is that correct?

Windows 7 x64


C:\miners\rpcminer nvopt>rpcminer-mod-cuda.exe -url=http://stratum.bitcoin.cz:3333 -user=user -password=pass
Client will start 1 miner threads
Work will be refreshed every 4000 ms
2 CUDA GPU devices found
Setting CUDA device to first device found
Loading module bitcoinminercuda.ptx
Unable to load CUDA module: 200

[charliemaggot]

Nvcc : fatal error : Cannot find compiler 'cl.exe' in PATH

Means that it can't find the C compiler. You need to have the Microsoft C compiler installed (e.g. Visual Studio 10 or Visual Studio 10 Express) to be able to build the cuda files. If you do have it already, you just need to run "C:\Program Files (x86)\Microsoft Visual Studio 10.0\VC\bin\vcvars32.bat" first in the command window to set up the paths.

If you don't have it, either download and install the Express version or just copy down the the prebuilt .ptx files from the bin directory in https://github.com/cdmackie/rpcminer-mod/tree/kepler/bin. There are now different ones, so try the bitcoinminercuda.ptx first. If that works, rename bitcoinminercuda.20.ptx to bitcoinminercuda.ptx and try again. Likewise with bitcoinminercuda.30.ptx. USe the last one that worked.

[charliemaggot]

I am getting a 200 error. I didn't see it covered above. I downloaded all the files in the bin folder into the same folder, is that correct?

@InqBit What cards do you have?

Save your bitcoinminercuda.ptx and rename bitcoinminercuda.30.ptx to bitcoinminercuda.ptx and try again. Also, try bitcoinminercuda.20.ptx.

[InqBit]

I am getting a 200 error. I didn't see it covered above. I downloaded all the files in the bin folder into the same folder, is that correct?

@InqBit What cards do you have?

I have a GTX480 and GTS450

edit: Fermi cards, but figured it would run. Maybe not?