CUDA miner

1303133353669

Comments

  • Marvell9Marvell9 Member Posts: 593 ✭✭✭
    Yeah im getting 20.9 mhs on single gtx 970s on windows 10
  • Marvell9Marvell9 Member Posts: 593 ✭✭✭


    - One problem I have is the SMI power limit command does not work with the older driver on windows 10 and

    - also I cant seem to really get any boost on hash rate when rasing the memory clock. (I already did all the SMI clocks stuff)

    I tried 3800 memory and + 80 Core no real boost just power temp and instability.

    I'll post screenies this evening.
  • vexdvexd Member Posts: 2
    Genoil said:

    @vexd @promojo the black screen and timeouts are caused by the kernel running longer than the timeout defined by TDR: https://msdn.microsoft.com/en-us/library/windows/hardware/ff569918(v=vs.85).aspx

    This is not supposed to happen, but in case of Windows 10, I can imagine it happens. 20Mh/s for 5 x 970 is way too low of course. If you want to continue with this setup anyway, I suggest you lower --cuda-grid-size (default 8192) by factor 2 until it runs stable.

    I get normal hashrates but then mining just stops and gives this error (at line 288): https://imgur.com/yThs5rv
    OS is win 8.1. How can I get rid of it, because this happens quite a lot?
  • restlessrestless Member Posts: 80
    I am trying to use 2 different workers on 1 machine. Its Athlon x2 4600, am2

    So I start 1 ethminer.exe thread on device 0 and another on device 1, using different dag location
    thread 0 device 0 is ethminer from official branch, set affinity to core 0
    thread 1 device 1 is genoil 1.0.4b3, set affinity to core 1
    thread1 is consuming 30-50% of core 1, while thread 0 consumes 3-4% on core 0


    Any idea/swicth how to decrease cpu consumption?
    To make things worse, the cards are radeon 7970, so I use open-cl ...
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    edited March 2016
    @restless I'll have to figure out what was changed in the official miner that my fork apparently doesn't have fixed yet. I've only been looking at CPU consumption on NVidia before. Lucky for you is that i'm going to shift focus to the opencl kernel soon.

    But before that I'm trying to get CUDA Compute 3.0 (GK104, GK106) support back in. There's a "recompute30" branch now that compiles for 3.0 and upwards and has all the bells and whistles of 1.0.4, but I haven't tested it on an actual Compute 3.0 device yet. If you have such a card and know how to compile, feel free to test.

    I've also upgraded the master branch to 1.0.4, but haven't changed anything to the stratum implementation yet. So consider it experimental :)
  • 4gun4gun Member Posts: 16
    edited March 2016
    I am sorry for the off-topic - can somebody help me find the latest AMD miner version?
    (Fastest that is ;) Any private ones?)
  • wirelessnet2wirelessnet2 Member Posts: 385 ✭✭✭
    @Genoil How would I be able to fix the "CUDA driver version is insufficient for CUDA runtime version" error? Im using 347.52 drivers
  • Marvell9Marvell9 Member Posts: 593 ✭✭✭
    use a newer version 350.12 i think
  • BiteTheCoinBiteTheCoin Member Posts: 6
    edited March 2016
    I tried compiling on Ubuntu 14.04 and get this:
    Edit never mind, forgot to compile just the miner.
    Genoil said:

    @restless I'll have to figure out what was changed in the official miner that my fork apparently doesn't have fixed yet. I've only been looking at CPU consumption on NVidia before. Lucky for you is that i'm going to shift focus to the opencl kernel soon.

    But before that I'm trying to get CUDA Compute 3.0 (GK104, GK106) support back in. There's a "recompute30" branch now that compiles for 3.0 and upwards and has all the bells and whistles of 1.0.4, but I haven't tested it on an actual Compute 3.0 device yet. If you have such a card and know how to compile, feel free to test.

    I've also upgraded the master branch to 1.0.4, but haven't changed anything to the stratum implementation yet. So consider it experimental :)

  • BiteTheCoinBiteTheCoin Member Posts: 6
    Tried new miner recompute30 on the g2.2 and got 9.08mh vs old miner 10.3mh
    Using:
    ethminer --farm-recheck 200 -G
    then my proxy information on dwarfpool.
  • MegaquakeMegaquake Member Posts: 11
    I have tried to mine with compoute 3.0 on my MSI laptop specs here i have 32gb ram upgrade
    https://www.msi.com/Notebook/GT70-Dominator-Dragon-Edition.html#hero-specification

    Windows 7 64 Benchmark

    Genoil's ethminer 0.9.41-genoil-1.0.3
    =====================================================================
    Forked from github.com/ethereum/cpp-ethereum
    CUDA kernel ported from Tim Hughes' OpenCL kernel
    With contributions from RoBiK, tpruvot and sp_

    Please consider a donation to:
    ETH: 0xeb9310b185455f863f526dab3d245809f6854b4d

    Using grid size 8192, block size 128

    Found suitable CUDA device [GeForce GTX 870M] with 3221225472 bytes of GPU memory
    Benchmarking on platform: CUDA
    Preparing DAG for block #0
    i 21:27:14|cudaminer0 workLoop 0 #00000000ΓǪ #00W0a0r0m0i0n0gΓ Çuªp...

    i 21:27:14|cudaminer0 Initialising miner...
    Using device: GeForce GTX 870M (Compute 3.0)
    Trial 1... 9754195
    Trial 2... 9754195
    Trial 3... 9754195
    Trial 4... 9754195
    Trial 5... 9747718
    min/mean/max: 9747718/9752899/9754195 H/s
    inner mean: 3251398 H/s

    When I mine I never get any shares.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @Megaquake the output shows you compiled 1.0.3, which doesn't support Compute 3.0. @BiteTheCoin thanks a lot for testing. Does it submit shares?
  • MegaquakeMegaquake Member Posts: 11
  • MegaquakeMegaquake Member Posts: 11
    i can run cuda-grid-size 16384 for faster hashing still no submitted shares
  • MegaquakeMegaquake Member Posts: 11
    where can i get the 1.0,4? I cant compile the source but rus up too 16-18mh i think thats impressive for a moble GPU
  • MegaquakeMegaquake Member Posts: 11
    Cant get 1.0.4 to work on my laptop gtx 870

    Cuda error in func 'set_constants' at line 45 : invalid device synbol
  • happytreefriendshappytreefriends Member Posts: 537 ✭✭✭
    Take troubleshooting to ANOTHER THREAD!
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @Megaquake GTX870m is a Compute 3.0 device. It's not supported by 1.0.4. Tonight I can test on Grid K520 and may release 1.0.5 with Compute 3.0 support, provided it works ok.
  • MegaquakeMegaquake Member Posts: 11
  • blueboxbluebox Member Posts: 181 ✭✭
    edited March 2016
    Had an old Tesla M2090 (compute 2.0) laying around that we're not using in the data center. Just for kicks & grins fired up the 1.0.3 version using opencl and it ran 11MH (CUDA ran 3MH, not surprising...). This card is about 5+ years old, quite surprised it worked at all with that ethminer version. Wish I could try out one of our K40's... B)
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @bluebox K40 is a GK110B, so performance should be similar to GTX780. Around 17.5MH/s..nothing spectacular :neutral:
  • blueboxbluebox Member Posts: 181 ✭✭
    @Genoil Yep, but I'll bet the M40 would work pretty well (if mining is single-precision?)... Sorry, HPC is my day job but the geek in me likes to dream/play sometimes. :#
  • BiteTheCoinBiteTheCoin Member Posts: 6
    Genoil said:

    @Megaquake the output shows you compiled 1.0.3, which doesn't support Compute 3.0. @BiteTheCoin thanks a lot for testing. Does it submit shares?

    Yeap! Seems to work nice and stable from what I have seen, but the older version from the main branch seems to go around 1MH faster.
    Maybe I'm missing a tweak or instructions for Ubuntu 14.04, but ran the two installs and could reproduce the 1MH variance.
    Now if the actual share amount is higher on the compute 3.0 one, that I'm not sure.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @bluebox M40 is GM200, so it will perform like GTX980.
  • blueboxbluebox Member Posts: 181 ✭✭
    50% more cores, 30% higher bandwidth (albeit 6GHz vs 7), 384bit width vs 256... I would hope it outperforms the 980, though we still don't know pricing yet.

    We won't be getting any; all of the compchem and cfd/fea apps we host are DP-intensive (on which maxwell thoroughly stinks).
  • phazzphazz Member Posts: 4
    quick question, is there a way to limit gpu use just enough to render the screens the cards support?
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @BiteTheCoin I just quickly set up an AWS g2 spot instance to check if it really works. I guess you have been using the opencl kernel, as the CUDA kernel does not submit any shares on Compute 3.0. It did hash at 11.5MH though. While I was trying to figure out what was wrong, the spot price went above the threshold and my instance was terminated so I have to start over again...
  • Bensam123Bensam123 Member Posts: 62
    edited March 2016
    Looking a bit on the quality of life side of things, Genoil, could you work on robustness of the miner? Say if one of the GPUs in a rig resets or locks up, it'll reset the miner and continue on with the other GPUs? That'd be incredibly helpful.

    Right now if a GPU locks up, the whole system locks or the miner itself just errors out. Usually the only way to tell what went wrong is by looking at the GPU speeds and see which has changed (if any). If you're unlucky it's really hard to diagnose as nothing changes.

    A error log that tells you what GPU locked up right before things went bad would be insanely helpful too.
  • BiteTheCoinBiteTheCoin Member Posts: 6
    Genoil said:

    @BiteTheCoin I just quickly set up an AWS g2 spot instance to check if it really works. I guess you have been using the opencl kernel, as the CUDA kernel does not submit any shares on Compute 3.0. It did hash at 11.5MH though. While I was trying to figure out what was wrong, the spot price went above the threshold and my instance was terminated so I have to start over again...

    Correct, using opencl, CUDA didn't send any shares on the old or new miner.
    Happy to hear you got 11.5MH though! What method of compiling did you use to obtain that?

  • Marvell9Marvell9 Member Posts: 593 ✭✭✭
    @Genoil windows 10 was bust even with my hack it still later installs other driver that mess up the ethminer forked version.

    trying Ubuntu using these instructions I ran into this error :

    [email protected]:~/cpp-ethereum/ethminer# ./ethminer -U
    modprobe: ERROR: ../libkmod/libkmod.c:556 kmod_search_moddep() could not open moddep file '/lib/modules/4.2.0-27-generic/modules.dep.bin'
    modprobe: ERROR: ../libkmod/libkmod.c:556 kmod_search_moddep() could not open moddep file '/lib/modules/4.2.0-27-generic/modules.dep.bin'
    modprobe: ERROR: ../libkmod/libkmod-module.c:809 kmod_module_insert_module() could not find module by name='nvidia_352'
    modprobe: ERROR: could not insert 'nvidia_352': Function not implemented
    Genoil's ethminer 0.9.41-genoil-1.0.4
    =====================================================================
    Forked from github.com/ethereum/cpp-ethereum
    CUDA kernel ported from Tim Hughes' OpenCL kernel
    With contributions from RoBiK, tpruvot and sp_

    Please consider a donation to:
    ETH: 0xeb9310b185455f863f526dab3d245809f6854b4d

    Selected CUDA mining without having compiled with -DETHASHCUDA=1 or -DBUNDLE=cudaminer


    These are the instructions I followed from page 8


    -------------------
    Install Ubuntu Linux 14.03 x64

    sudo -s

    wget -O - http://llvm.org/apt/llvm-snapshot.gpg.key | sudo apt-key add -
    add-apt-repository "deb http://llvm.org/apt/trusty/ llvm-toolchain-trusty main"
    add-apt-repository -y ppa:ethereum/ethereum-qt
    add-apt-repository -y ppa:ethereum/ethereum
    add-apt-repository -y ppa:ethereum/ethereum-dev
    wget http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1404/x86_64/cuda-repo-ubuntu1404_7.0-28_amd64.deb
    dpkg -i cuda-repo-ubuntu1404_7.0-28_amd64.deb
    apt-get -y update
    apt-get -y upgrade

    apt-get -y install cpp-ethereum screen cuda build-essential git cmake libboost-all-dev libgmp-dev libleveldb-dev libminiupnpc-dev libreadline-dev libncurses5-dev libcurl4-openssl-dev libcryptopp-dev libjson-rpc-cpp-dev libmicrohttpd-dev libjsoncpp-dev libargtable2-dev llvm-3.8-dev libedit-dev mesa-common-dev ocl-icd-libopencl1 opencl-headers libgoogle-perftools-dev qtbase5-dev qt5-default qtdeclarative5-dev libqt5webkit5-dev libqt5webengine5-dev ocl-icd-dev libv8-dev

    git clone https://github.com/Genoil/cpp-ethereum
    cd cpp-ethereum
    nano cmake/EthCompilerSettings.cmake
    # remove "-Werror" around line 5
    nano libethash-cu/ethash_cu_miner.cpp
    # remove cudaDeviceSetSharedMemConfig(cudaSharedMemBankSizeEightByte); around line 140
    # then add cudaSetDeviceFlags(cudaDeviceScheduleBlockingSync);
    # then add cudaDeviceSetCacheConfig(cudaFuncCachePreferL1);
    cmake -DBUNDLE=miner -DETHASHCU=1
    make -j4
    reboot
    sudo -s
    cd cpp-ethereum/ethminer/
    screen
    ./ethminer -U -F http://(yourpool) --gpu-devices 0 1 2

    -----------------------

    Only difference I can see is a have Ubuntu 14.04 64 bit followed them to the letter except for this part which it seems is not relevant anymore :

    nano cmake/EthCompilerSettings.cmake
    # remove "-Werror" around line 5
    nano libethash-cu/ethash_cu_miner.cpp
    # remove cudaDeviceSetSharedMemConfig(cudaSharedMemBankSizeEightByte); around line 140
    # then add cudaSetDeviceFlags(cudaDeviceScheduleBlockingSync);
    # then add cudaDeviceSetCacheConfig(cudaFuncCachePreferL1

    any ideas ?



Sign In or Register to comment.