CUDA miner

18911131469

Comments

  • crsminercrsminer Member Posts: 75
    Also when doing cmake -DBUNDLE=miner -DETHASHCU=1 I get this:

    -- The CXX compiler identification is unknown
    CMake Error: your CXX compiler: "CMAKE_CXX_COMPILER-NOTFOUND" was not found. Please set CMAKE_CXX_COMPILER to a valid compiler path or name.
    -- ctest path: /usr/bin/ctest
    CMake Error at cmake/FindPackageHandleStandardArgs.cmake:138 (message):
    Could NOT find CryptoPP (missing: CRYPTOPP_INCLUDE_DIR CRYPTOPP_LIBRARY)
    (Required is exact version "5.6.2")
    Call Stack (most recent call first):
    cmake/FindPackageHandleStandardArgs.cmake:374 (_FPHSA_FAILURE_MESSAGE)
    cmake/FindCryptoPP.cmake:107 (FIND_PACKAGE_HANDLE_STANDARD_ARGS)
    cmake/EthDependencies.cmake:44 (find_package)
    CMakeLists.txt:298 (include)


    -- Configuring incomplete, errors occurred!
    See also "/home/crsminer10/cpp-ethereum/CMakeFiles/CMakeOutput.log".
    See also "/home/crsminer10/cpp-ethereum/CMakeFiles/CMakeError.log".


    Sorry for my noobishness, but it's my first time on linux!
  • skunkskunk Member Posts: 13
    crsminer said:


    Anybody knows why, when I do dpkg -i cuda-repo-ubuntu1404-7-5-rc_7.5-7_amd64.deb, I get the following error?

    corrupted archive, download it again...
    btw, don't know with ethminer, but cuda 6.5 is generally faster for mining.
  • skunkskunk Member Posts: 13
    install http://tahoe-lafs.org/trac/pycryptopp
    don't know exact debian/ubuntu package name
  • Bensam123Bensam123 Member Posts: 62
    Yup, 6.5 is the almost universally accepted solution for cuda mining. They've tried a lot of other variants in the CCminer thread and they never match up.
  • crsminercrsminer Member Posts: 75
    skunk said:

    crsminer said:


    Anybody knows why, when I do dpkg -i cuda-repo-ubuntu1404-7-5-rc_7.5-7_amd64.deb, I get the following error?

    corrupted archive, download it again...
    btw, don't know with ethminer, but cuda 6.5 is generally faster for mining.
    Thanks! downloaded directly, installed and now nvidia-smi showing all cards!

  • crsminercrsminer Member Posts: 75
    skunk said:

    install http://tahoe-lafs.org/trac/pycryptopp
    don't know exact debian/ubuntu package name

    I git cloned it but can't figure it out how to install... sorry :smile:

  • skunkskunk Member Posts: 13
    apt-get install pycryptopp ?
  • crsminercrsminer Member Posts: 75
    Managed to install python-pycryptopp_0.5.29-1build1_amd64.deb , but still getting that error.

    ethminer folder is emty...

    maybe some other help please?

  • crsminercrsminer Member Posts: 75
    Do not know what I done, but it works now! It's generating the DAG file.
  • FlatulentfoxFlatulentfox Member Posts: 1
    Thanks for the hard work on the cuda miner. I'm running an evga gtx 750 ti sc but i'm only getting 7.4 mh/s. Any ideas on how to reach the 9.8mh/s that i'm seeing on forums?

    I switched from windows 8 to windows 7 using the same hardware and got a big jump in hashrates. Windows 8 was giving me 0.4mh/s for some reason. Would switching to ubuntu give me close to 9.8mh/s? Thanks in advance from a newbie.

    OS - Windows 7
    CPU - Celeron G3220
    Motherboard - Asrock B85
    Ram - 4gb
    GPU - 1x EVGA 750ti superclocked (1300 mhz)

  • PhantomPhantom Member Posts: 46
    I also gave up with win 7 and win 8.1. With 6 x GTX 750 ti, I run perfectly on Ubuntu with 52 Mh/s
    However, I got no luck with ATI Radeon HD 9750 and R9 290 Tri-X on Ubuntu, could not go higher than 30 Mh/s. So, I switch back to Win 8.1 and get about 48 Mh/s with HD 9750 overclock and R9 290 standard clock.

    Now, I am trying to run GTX 750 ti on Windows again, because I can easy overclock those cards with Windows.
  • Bensam123Bensam123 Member Posts: 62
    Something worth noting Genoil, the memory controller is pegged at 100% when looking at GPUZ.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    edited August 2015
    @Bensam123 it better be at 100%, or high at least. The kernel is memory-bound, meaning that its speed it limited by memory access times. What still puzzles me however, is that AMD cards get much closer to the theoretical memory bandwidth than NVidia.

    @Flatulentfox the normal speed for a GTX750Ti is about 8MH/s in becnhamrk mode, bit lower for actual mining. The 9.8 probably came from a highly tuned card.

    The low hashrate on Win8.1/Win10 is a known issue.
  • Bensam123Bensam123 Member Posts: 62
    edited August 2015
    This problem is also present in Cryptonote, that was covered a bit in the CCMiner thread. Although I don't know if this can be cleaned up at all.

    https://bitcointalk.org/index.php?topic=826901

    I don't know if this is applicable, but certain data formats move faster through memory then others. Not sure if you get to pick or choose.

    http://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    edited August 2015
    It already loads vec4 (4 * 32 bit). The problem (or the solution, depends how you look at it) is that the ethash algo has to do 64 of those vec4 loads sequentially per GPU thread at unpredictable locations in the DAG file. It's highly inefficient from a GPU standpoint, but exactly what the designers of the algo were after.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    edited August 2015
    I've completed merging a reasonably recent develop branch with cudaminer. No performance gains, but it should be more compatible with latest ethminer now.
    • new branch at https://github.com/Genoil/cpp-ethereum/tree/cudaminer-frontier
    • build instructions: install cuda sdk; add cmake flag -DBUNDLE=cudaminer
    • win64 binaries attached
    • many cmd line params changed. See --help section
    • bug with specific multiple cuda device selection likely fixed (please test, I have only 1 GPU)
    • Linux compilation untested. Help/test is highly appreciated
    • donations welcome @ 0xeb9310b185455f863f526dab3d245809f6854b4d
    Post edited by Genoil on
  • cryptletcryptlet Member Posts: 29
    @Genoil This cudaminer-fontier release doesn't show hashrate?

    Also I tried compilling on ubuntu. I get an error. I have no problems compilling your previous existing builds and geth on my linux machine. So far just this.

    [ 53%] Building CXX object libdevcore/CMakeFiles/devcore.dir/TrieHash.cpp.o [ 55%] Building CXX object libdevcore/CMakeFiles/devcore.dir/TransientDirectory.cpp.o [ 57%] Building CXX object libdevcore/CMakeFiles/devcore.dir/Hash.cpp.o Linking CXX shared library libdevcore.so [ 57%] Built target devcore make: *** [all] Error 2
  • Bensam123Bensam123 Member Posts: 62
    Not sure if that's what you were trying to fix. But --opencl-device doesn't do anything nor does ---opencl-platform.
    Genoil said:

    It already loads vec4 (4 * 32 bit). The problem (or the solution, depends how you look at it) is that the ethash algo has to do 64 of those vec4 loads sequentially per GPU thread at unpredictable locations in the DAG file. It's highly inefficient from a GPU standpoint, but exactly what the designers of the algo were after.

    int8/int16/fp32 doesn't have anything to do with mining? I have no idea, just looking at benchmarks.
  • skunkskunk Member Posts: 13
    [ 64%] Building CXX object libethash-cu/CMakeFiles/ethash-cu.dir/ethash_cu_miner.cpp.o /home/skunk/bitcoin/cpp-ethereum/libethash-cu/ethash_cu_miner.cpp: In static member function ‘static bool ethash_cu_miner::configureGPU(unsigned int, unsigned int, unsigned int, unsigned int, bool, uint64_t)’: /home/skunk/bitcoin/cpp-ethereum/libethash-cu/ethash_cu_miner.cpp:155:45: error: invalid operands of types ‘const char [29]’ and ‘char [256]’ to binary ‘operator
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    Thanks guys. Got work to do :)
  • scryptrscryptr Member Posts: 20
    750ti PERFORMANCE-- I am sorry to report that after a fresh install of Win 7 x64, and the latest nVidia drivers (355.60), there was no improvement on the 750ti bug. But, because the Ethminer only reports the total hashing rate for the rig, and not the individual cards and the hash rate accepted, it is difficult to see what is really happening.

    The hash rate starts high, then tapers to a point where it equals the average hash rate for the 1 960 card, plus about 0.5Mh/s to 1Mh/s for each 750ti card. --scryptr
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @scryptr not sure if --gpu-devices or --opencl-device works fully in either cudaminer or cudaminer-frontier. that way you could isolate the 750ti. but i've got a bad feeling about this...
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    cryptlet said:

    @Genoil This cudaminer-fontier release doesn't show hashrate?
    ..

    Yes I noticed and found the cause. I'll take it up with ethdevs, probably a (Windows) build issue.
  • GruberGruber Member Posts: 6
    Genoil said:

    @Gruber I measure CPU load caused by ethminer in NSight Profiler. In that tool, the sp-mod CPU offloading variant as suggested by him in ihis bitcointalk thread actually reports 85% CPU, vs. 4% on the tpruvot ("ed...") version. So I went with tpruvot's solution.

    Attached is a binary with sp's CPU fix reenabled. With --high-cpu-load enabled, it should give similar performance as the sp-mod binary. If that one is still better, then it is good news because there's a few other optimizations in there that don't do anything on my GTX780 but may work on your 670

    @Genoil I measured CPU load caused ethminer using Process Explorer. Tested platform: GeForce GTX 560M, CPU i5-3470, Windows 8.1. Using CUDA miner version posted August 25.

    ethminer_CUDA_20150826.exe -M -U
    inner mean: 10621516 H/s
    CPU load: 0.26% - 0.28%

    ethminer_CUDA_20150826.exe -M -U --high-cpu-load
    inner mean: 12223879 H/s
    CPU load: 0.13% - 0.15%

    Result hashrate using --high-cpu-load higher and close to ethminer sp_mod. But the load on the processor is less than when using --high-cpu-load.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @Gruber, strange that NSight then reports different values. Would be useful if somebody with a wall power meter would point out the actual difference, because that's where it's all about. Next version will have --high-cpu-load renamed to --cuda-turbo, not making any claims on load any longer ;)
  • GruberGruber Member Posts: 6
    @Genoil OK Play around evening =)
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @skunk @cryptlet : it should build on Ubuntu now. I tried on an EC2 instance, needed 8GB of RAM. Then after that I upgraded the instance to have GPUs, but failed at installing NVidia drivers. I'm so not going to spend time fixing that :). Can you please build and test if (multi-)GPU mining works?

    Thanks.
  • scryptrscryptr Member Posts: 20
    I have decided to skip the 750ti struggle. :) I mined solo for a week and earned 4 blocks, and I mined at NanoPool and earned a couple more ether coins. The pool reported me at up to 56Mh/s, mining on 2 GTX 960 cards and 4 750ti cards, but the average was more like 24Mh/s at best, poolside.

    Instead, I will rebuild my 280x rig as I can afford new hardware. If ether coins improve on the market, I can always point my 960s at it, because your Cuda code does work well for them. The 280x cards can be set up for Ethereum without conflicting with my CCminer builds.

    Thank you for your work. I am going to follow your thread! --scryptr
  • skunkskunk Member Posts: 13
    @Genoil it builds fine after removing -Werror
    results for single 960 follows:

    ./ethminer -M -U
    min/mean/max: 10310997/10363425/10398378 H/s
    inner mean: 3466126 H/s

    ./ethminer -M -U --cuda-turbo
    min/mean/max: 10310997/10380901/10398378 H/s
    inner mean: 3466126 H/s

    ./ethminer -M -G
    min/mean/max: 10223616/10276044/10310997 H/s
    inner mean: 3436999 H/s

    it seems --cuda-turbo doesn't make any difference anymore, cpu utilization is low with or without the flag...
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @skunk thanks. Apparently NVidia has done some great work on their OpenCL implementation for Maxwell 5.2 cards. Differences are soo small. I'm going to try adding some NVidia specific tweaks to the opencl kernel, see if I can surpass native CUDA :)
Sign In or Register to comment.