CUDA Optimalized BTC miner for NVIDIA cards

[icedev576]

I`ve started to optimalize the rpc-miner CUDA version. Currently it runs on Linux (Ubuntu 12.04), and tested on 2 different cards:
Asus GTX-590: 260 MH/s (no OC)
Asus GTX-690: 230 MH/s (no OC)
It is about 15-30% faster then the original.
I started this topic to discuss is it worth it to continue this project or no one interested in...

I am intersted in your current hash rates on newer nvidia cards also.

I accept donations for this project in the following BTC address: 1GgnBSzLq6DioYS4haPSx6gZCeL9KMUYFW.

It will be opensource.

[1713939909]

1713939909

[1713939909]

1713939909

[1713939909]

1713939909

[1713939909]

1713939909

[1713939909]

1713939909

[gaborjakab]

Hi!

I have an overclocked 680 and have only 135 Mhash/s. I would love to see if your modification helps a bit. By the way I am using on older Ubuntu. When will it be available?

[roy7]

You might find these helpful

https://bitcointalk.org/index.php?topic=163750.0
https://bitcointalk.org/index.php?topic=167229.0

[icedev576]

This is interesting. I am reading the post, maybe I can mix the solutions together.

[klintay]

yeah i would be interest too in this!

[smoothie]

Hopefully this hasn't already been optimized. But if so, good job.

[icedev576]

My code is a mess now, so it isn`t ready for a public release, but because gaborjakab uses almost the same hardware, and OS like me, I can send a binary from the current stage just for the test. So if you like please send a private message.

[LighningChun]

i have a gtx card, looking forward for this! 

[icedev576]

Well, I read the whole "NVIDIA Kepler (K20) from 134MHash/s to 330MHash/s with CUDA" topic. As far I know psychocoder has optimalized the cuda code for the GK110 architecture (sm_35 in cuda). There was only 2 card supporting this: Tesla K20, Geforce Titan, but from yesterday you can buy the Geforce 780! It is the same arch, a little cheaper and more consumer. If I had enough donations, I will buy one.

My solution is speeding up the older generations: Fermi and Kepler(sm_20, sm_30), that means geforce 4xx to 6xx. So I hope I can merge them together.

I sent a binary to gaborjakab, others please wait and keep me posted.

[gaborjakab]

Thanks man, I have tried it, now it goes with about 150 MHash/s, about 10% faster then before. It is not a miracle, but it is something.

[thomassie12]

I used a gtx 660 (non Ti) i couldnt get it higher then 73 mhash/s

[icedev576]

Thomassie12 are you still mining or you give up? What kind of miner do you use or used?

[kenshin23]

I'm currently using a EVGA GTX 670 with cgminer and I get 91-95 MH/s. Seems low, but I don't have any sort of optimizations for it just yet. So I'm very interested to see what you can achieve with this.

[icedev576]

HI Kenshin23!

Can you give me some information: what OS you use is it 64bit win or linux(distro?)? And the most important are you using rpcminer or something else?

[Cyberclones]

I started by running a EVGA GTX 550 Ti 1 GB was getting about 45Mhash/s without a overclock on it but I no longer run it

[icedev576]

I`ve managed to merge the source. Tomorrow I will upload the source code to github. I am also thinking to create an cuda mining survival guide to add some help for nvidia card owners.

[kenshin23]

HI Kenshin23!

Can you give me some information: what OS you use is it 64bit win or linux(distro?)? And the most important are you using rpcminer or something else?

Hi, sure! Win 7 64 bit and currently using cgminer 3.1.1 as-is, but willing to switch to another miner if it means it'll be more efficient.

The CUDA mining survival guide sounds great; looking forward to it.

[OnLashoc]

I'm in same boat you are ^^

[xh4]

Looking forward to using all my old graphics cards for something useful...

[gaborjakab]

Any news from here? I am still waiting the sourcecode.

[Devnant]

Hey man! Very interested in testing your code on a GTX TITAN!

[gourmet]

I`ve managed to merge the source. Tomorrow I will upload the source code to github. I am also thinking to create an cuda mining survival guide to add some help for nvidia card owners.

Looking forward to improve my GTX 580 performance at least a bit. It's not my main worker, but there's no reason to leave its power unused.

[icedev576]

Hey man! Very interested in testing your code on a GTX TITAN!

Sorry currently I am only optimalized for 4xx to 6xx, the titan is a different issue. maybe later

[icedev576]

Any news from here? I am still waiting the sourcecode.

Sorry I am little busy at my job, I don`t have time for this today. I`ve created the github repo, but it is relatively empty now.

[LuckyXIII]

Has anyone actually gotten an Nvidia card to mine efficiently enough to pay for it's own power consumption? Or are these solely for people mining with "free" electricity?

[R3dcell]

very interesting, any info as to whether this would increase the hash rate on my GTX 660?

[icedev576]

Source has been released. You can found here:
https://github.com/icedev576/rpcminer-optcuda

Please follow the instructions, to build and run in: README, README.md, HOW-TO files

Anyone who was able to build any variation of rpcminer, can this also.

Compiled Binaries for different OS, may come later.

[djmagicmetz]

Asics are going to fry these things. Game over.

[kenshin23]

Thanks for this! I'm downloading the CUDA toolkit to see if I can compile it for my system. Otherwise, I'll wait for the binaries.

[icedev576]

Has anyone actually gotten an Nvidia card to mine efficiently enough to pay for it's own power consumption? Or are these solely for people mining with "free" electricity?

I think the short answer is no. I don`t think anyone bought an nvidia card just for mining. I think (just like me) we got nv card because other reasons (for me, I use it for work). In the computer spare time, mining can run, but it`s not worth unless you have "free" electricity.

[kenshin23]

Has anyone actually gotten an Nvidia card to mine efficiently enough to pay for it's own power consumption? Or are these solely for people mining with "free" electricity?

I think the short answer is no. I don`t think anyone bought an nvidia card just for mining. I think (just like me) we got nv card because other reasons (for me, I use it for work). In the computer spare time, mining can run, but it`s not worth unless you have "free" electricity.

In my particular case, I had an ATI HD 5770 but it sucked at flightsimming (which is what I built my computer for) so I sold that (stupid, I know, but I really didn't know much about bitcoins then) and bought the 670 I'm now using. It runs FSX great, but now I got interested in bitcoins and found out that the card I sold can hash at least twice as much than what the 670 can do normally. I can't really undo the sale, and buying an ATI card that can mine efficiently is more or less out of the question right now.

Hence, my interest in optimizing my card's performance. I also know that GPU mining is almost dead, but it's worth a shot, I guess.

[gourmet]

Source has been released. You can found here:
https://github.com/icedev576/rpcminer-optcuda

Please follow the instructions, to build and run in: README, README.md, HOW-TO files

Anyone who was able to build any variation of rpcminer, can this also.

Compiled Binaries for different OS, may come later.

I've compiled cgminer on Ubuntu recently.
But there is Radeon card on that machine.
Never have been using rpc-miner. (Is there
any big difference in compilation/usage?)
My Nvidia is in a machine running Windows.
So I'll wait for Windows binaries, too.

[edit]
I'm running guiminer there now. I'm really
curious what the difference may be. :-)

[dentldir]

Testing a 660ti with this right now, seeing about 104MH/s on stock clocks.

cgminer comes in at 100MH/s.  Both are steady, so it is a solid 4% improvement without any tweaking.

The same card, on scrypt using cudaminer, now gets around 160KH/s after several kernel iterations.  I don't know if there is anything instructive inside the spinlock kernels in cudaminer 4-30, but it might be worth a look.

Cheers and thanks for the good work.

[minehe]

260 MH/s is too slow.

[dentldir]

On that 660ti, the auto config found:

Best configuration step found (-gpugrid=448,-gputhreads=1024) 104206 KHs  prev best=105422 KHs
Done allocating CUDA resources for (448,1024)

on a second run, and settled in at 105MH/s.

[kenshin23]

On that 660ti, the auto config found:

Best configuration step found (-gpugrid=448,-gputhreads=1024) 104206 KHs  prev best=105422 KHs
Done allocating CUDA resources for (448,1024)

on a second run, and settled in at 105MH/s.

Are you using that card exclusively to mine, or do you have it configured for normal desktop usage?

[dentldir]

It has an X server running on it, but I'm not using it as a desktop.  It was a default Ubuntu 12.04 install running the driver in the CUDA 5.0 package.  (304.86?).

[Cyberburner]

Problem with all these programs is you have to be a technician to be able to work it. The first one to release some more user friendly things, that actually work, will be my bitcoin jezus...

[cyberkiller]

interesting, definitely a good thing. Everyone is so focused on AMD cards, yes rightfully so as they are faster but people have tons of nvidia cards laying around.

[icedev576]

I have written instructions in the readme files, but I write here some important information:
 - Use 4.2 if you compile the cu files. CU files should be compiled as a second pass and compiled next to the executable
 - I've added precompiled cubin files to the project, it is OS independent so you can copy it next to the executable file, if you don't want to compile the cu files
 - The program has an auto-tuning feature, if you don't define the gpuGrid, and gputhreads, it will find the best configuration, it is a wider search then the original (it can take several minutes!)
 - Once it is done you can use it next time
 - It does't have native stratum support, use a proxy instead
 - For measuring the time please run it for at least a few minutes

[nightyj]

With this miner I manage to get from my GTX480 from 110 till 140 Mh/s at default, thanks. 

[icedev576]

With this miner I manage to get from my GTX480 from 110 till 140 Mh/s at default, thanks. 

You welcome!

[Eli0t]

would be great to get some more MH out of my 560ti, currently 83 MH

waiting for win7 64bit binaries

have you fixed rpc-miner-cuda so it doesnt crash the display driver when closing it, or was that a windows issue

[zhuravno]

what performance increase is expected compared to guiminer or cgminer?

[llieco]

260 MH/s to slow.

[icedev576]

what performance increase is expected compared to guiminer or cgminer?

Well actually CG miner is quite good, I don't expect more then 10%. For Titan and GTX780, if somebody will do some optimalization CUDA can be better. If I have enough money I'll buy a 780, and try it.

[icedev576]

260 MH/s to slow.

Unfortunately NV cards not so good for mining. Buy AMD or some asic if you want more or you can design a new asic.

[icedev576]

would be great to get some more MH out of my 560ti, currently 83 MH

waiting for win7 64bit binaries

have you fixed rpc-miner-cuda so it doesnt crash the display driver when closing it, or was that a windows issue

Well, I don't know, I have dedicated cards for CUDA (no display or X attached to it)

[kenshin23]

Any update on binaries for Windows? I've tried compiling with Cygwin, but cmake complains it can't find CUDA_CUDART_LIBRARY.

Code:

$ cmake ../
-- The C compiler identification is GNU 4.5.3
-- The CXX compiler identification is GNU 4.5.3
CMake Warning at /usr/share/cmake-2.8.9/Modules/Platform/CYGWIN.cmake:15 (message):
  CMake no longer defines WIN32 on Cygwin!

  (1) If you are just trying to build this project, ignore this warning or
  quiet it by setting CMAKE_LEGACY_CYGWIN_WIN32=0 in your environment or in
  the CMake cache.  If later configuration or build errors occur then this
  project may have been written under the assumption that Cygwin is WIN32.
  In that case, set CMAKE_LEGACY_CYGWIN_WIN32=1 instead.

  (2) If you are developing this project, add the line

    set(CMAKE_LEGACY_CYGWIN_WIN32 0) # Remove when CMake >= 2.8.4 is required

  at the top of your top-level CMakeLists.txt file or set the minimum
  required version of CMake to 2.8.4 or higher.  Then teach your project to
  build on Cygwin without WIN32.
Call Stack (most recent call first):
  /usr/share/cmake-2.8.9/Modules/CMakeSystemSpecificInformation.cmake:36 (INCLUDE)
  CMakeLists.txt:3 (PROJECT)


-- Check for working C compiler: /usr/bin/gcc.exe
-- Check for working C compiler: /usr/bin/gcc.exe -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++.exe
-- Check for working CXX compiler: /usr/bin/c++.exe -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Boost version: 1.53.0
-- Found the following Boost libraries:
--   date_time
--   filesystem
--   program_options
--   regex
--   system
--   thread
-- Found OpenSSL: /usr/lib/libssl.dll.a;/usr/lib/libcrypto.dll.a (found version "1.0.1e")
-- Looking for include file pthread.h
-- Looking for include file pthread.h - found
-- Looking for pthread_create
-- Looking for pthread_create - found
-- Found Threads: TRUE
CMake Error at /usr/share/cmake-2.8.9/Modules/FindPackageHandleStandardArgs.cmake:97 (MESSAGE):
  Could NOT find CUDA (missing: CUDA_CUDART_LIBRARY)
Call Stack (most recent call first):
  /usr/share/cmake-2.8.9/Modules/FindPackageHandleStandardArgs.cmake:288 (_FPHSA_FAILURE_MESSAGE)
  cmake-modules/FindCUDA.cmake:682 (find_package_handle_standard_args)
  CMakeLists.txt:32 (FIND_PACKAGE)


-- Configuring incomplete, errors occurred!

If anyone has any how-to's or some kind of guidance on how to get this working, please share.

[mikal]

Can you explain the increase of hash rate? I thought the original miner got all could be achieved on sm30...
I'll start digging into the code, but I'd like to read high-level explanation.

[icedev576]

Any update on binaries for Windows? I've tried compiling with Cygwin, but cmake complains it can't find CUDA_CUDART_LIBRARY.

Code:

$ cmake ../
-- The C compiler identification is GNU 4.5.3
-- The CXX compiler identification is GNU 4.5.3
CMake Warning at /usr/share/cmake-2.8.9/Modules/Platform/CYGWIN.cmake:15 (message):
  CMake no longer defines WIN32 on Cygwin!

  (1) If you are just trying to build this project, ignore this warning or
  quiet it by setting CMAKE_LEGACY_CYGWIN_WIN32=0 in your environment or in
  the CMake cache.  If later configuration or build errors occur then this
  project may have been written under the assumption that Cygwin is WIN32.
  In that case, set CMAKE_LEGACY_CYGWIN_WIN32=1 instead.

  (2) If you are developing this project, add the line

    set(CMAKE_LEGACY_CYGWIN_WIN32 0) # Remove when CMake >= 2.8.4 is required

  at the top of your top-level CMakeLists.txt file or set the minimum
  required version of CMake to 2.8.4 or higher.  Then teach your project to
  build on Cygwin without WIN32.
Call Stack (most recent call first):
  /usr/share/cmake-2.8.9/Modules/CMakeSystemSpecificInformation.cmake:36 (INCLUDE)
  CMakeLists.txt:3 (PROJECT)


-- Check for working C compiler: /usr/bin/gcc.exe
-- Check for working C compiler: /usr/bin/gcc.exe -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++.exe
-- Check for working CXX compiler: /usr/bin/c++.exe -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Boost version: 1.53.0
-- Found the following Boost libraries:
--   date_time
--   filesystem
--   program_options
--   regex
--   system
--   thread
-- Found OpenSSL: /usr/lib/libssl.dll.a;/usr/lib/libcrypto.dll.a (found version "1.0.1e")
-- Looking for include file pthread.h
-- Looking for include file pthread.h - found
-- Looking for pthread_create
-- Looking for pthread_create - found
-- Found Threads: TRUE
CMake Error at /usr/share/cmake-2.8.9/Modules/FindPackageHandleStandardArgs.cmake:97 (MESSAGE):
  Could NOT find CUDA (missing: CUDA_CUDART_LIBRARY)
Call Stack (most recent call first):
  /usr/share/cmake-2.8.9/Modules/FindPackageHandleStandardArgs.cmake:288 (_FPHSA_FAILURE_MESSAGE)
  cmake-modules/FindCUDA.cmake:682 (find_package_handle_standard_args)
  CMakeLists.txt:32 (FIND_PACKAGE)


-- Configuring incomplete, errors occurred!

If anyone has any how-to's or some kind of guidance on how to get this working, please share.

Is your CUDA_TOOLKIT_ROOT_DIR environment variable set?

[icedev576]

Can you explain the increase of hash rate? I thought the original miner got all could be achieved on sm30...
I'll start digging into the code, but I'd like to read high-level explanation.

Everything about the used register-count. If a CUDA code uses too many registers 2 things can happen: you can run the code on less threads, or you can tell the nvcc compiler to force use less register. There is no magic if your code needs 72 register, but you force the compiler use only 63 will mean 9 register will be in  local memory (available via spill loads), and this is slower then the registers.

So optimizing a CUDA code usually means write a code that can run with less register. On the other hand you can use a lot of special functions and types (I moved the input storing to __constant__ memory ). Memory reading and writing is slow, avoid unnecessary memory ops. And there are a lot of other tricks...

[icedev576]

Code now works on cards with compute mode 3.5 (Titan, GTX780).
Tested on GTX-780 :-) It is about 275 MH/s.

[kenshin23]

Any update on binaries for Windows? I've tried compiling with Cygwin, but cmake complains it can't find CUDA_CUDART_LIBRARY.

Code:

$ cmake ../
-- The C compiler identification is GNU 4.5.3
-- The CXX compiler identification is GNU 4.5.3
CMake Warning at /usr/share/cmake-2.8.9/Modules/Platform/CYGWIN.cmake:15 (message):
  CMake no longer defines WIN32 on Cygwin!

  (1) If you are just trying to build this project, ignore this warning or
  quiet it by setting CMAKE_LEGACY_CYGWIN_WIN32=0 in your environment or in
  the CMake cache.  If later configuration or build errors occur then this
  project may have been written under the assumption that Cygwin is WIN32.
  In that case, set CMAKE_LEGACY_CYGWIN_WIN32=1 instead.

  (2) If you are developing this project, add the line

    set(CMAKE_LEGACY_CYGWIN_WIN32 0) # Remove when CMake >= 2.8.4 is required

  at the top of your top-level CMakeLists.txt file or set the minimum
  required version of CMake to 2.8.4 or higher.  Then teach your project to
  build on Cygwin without WIN32.
Call Stack (most recent call first):
  /usr/share/cmake-2.8.9/Modules/CMakeSystemSpecificInformation.cmake:36 (INCLUDE)
  CMakeLists.txt:3 (PROJECT)


-- Check for working C compiler: /usr/bin/gcc.exe
-- Check for working C compiler: /usr/bin/gcc.exe -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++.exe
-- Check for working CXX compiler: /usr/bin/c++.exe -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Boost version: 1.53.0
-- Found the following Boost libraries:
--   date_time
--   filesystem
--   program_options
--   regex
--   system
--   thread
-- Found OpenSSL: /usr/lib/libssl.dll.a;/usr/lib/libcrypto.dll.a (found version "1.0.1e")
-- Looking for include file pthread.h
-- Looking for include file pthread.h - found
-- Looking for pthread_create
-- Looking for pthread_create - found
-- Found Threads: TRUE
CMake Error at /usr/share/cmake-2.8.9/Modules/FindPackageHandleStandardArgs.cmake:97 (MESSAGE):
  Could NOT find CUDA (missing: CUDA_CUDART_LIBRARY)
Call Stack (most recent call first):
  /usr/share/cmake-2.8.9/Modules/FindPackageHandleStandardArgs.cmake:288 (_FPHSA_FAILURE_MESSAGE)
  cmake-modules/FindCUDA.cmake:682 (find_package_handle_standard_args)
  CMakeLists.txt:32 (FIND_PACKAGE)


-- Configuring incomplete, errors occurred!

If anyone has any how-to's or some kind of guidance on how to get this working, please share.

Is your CUDA_TOOLKIT_ROOT_DIR environment variable set?

It wasn't, but I just set it prior to compilation and the problem persists. Any other clues? Or is cygwin not appropriate for compiling this? (might have to make a liveUSB with linux just for mining...)

[Eli0t]

Code now works on cards with compute mode 3.5 (Titan, GTX780).
Tested on GTX-780 :-) It is about 275 MH/s.

any hope for compute 2.1? also how does that rate vs your cgminer speed

[icedev576]

Code now works on cards with compute mode 3.5 (Titan, GTX780).
Tested on GTX-780 :-) It is about 275 MH/s.

any hope for compute 2.1? also how does that rate vs your cgminer speed

UPDATE:
Automatic GPU architecture chooser was added it works on any architecture from 1.0 to 3.5. The relevant cubin file should be in cmake-rpcminer. I added these files on github, so use them.

[Eli0t]

Code now works on cards with compute mode 3.5 (Titan, GTX780).
Tested on GTX-780 :-) It is about 275 MH/s.

any hope for compute 2.1? also how does that rate vs your cgminer speed

UPDATE:
Automatic GPU architecture chooser was added it works on any architecture from 1.0 to 3.5. The relevant cubin file should be in cmake-rpcminer. I added these files on github, so use them.

i guess the difficulty is too high now, but if you could include all the relevant information from the whole thread in the first post and possibly some windows binaries im sure it would help someone out.

think i missed some words from my last post. what is your rpcminer-optcuda MH vs your cgminer MH ?

[gourmet]

Also looking forward to Windows binaries and the comparison with cgminer. ;-)