CUDA miner

1596062646569

Comments

  • Mo35Mo35 Member Posts: 43
    finally got hands on my 1080 , pulling out 950 and 970 from main box to build dedicated nvidia mining rig , question is witch drivers is best for win7 setup , latest will perform ok , or i have to tinker like with win10 setup ?
  • Bensam123Bensam123 Member Posts: 62
    Tested 368.69 in W8.1 x64, it's still broken in OCL and in Cuda mode. I get a queuebuffer error in OCL and in Cuda I get a 'unspecified launch failure' after it finishes building the dag.

    Perhaps a chunk style loading might be needed (even if devices have enough memory) on W7/8. Since in W10, the newer WDDM version increases the chunk size.
  • Wolf0Wolf0 Member Posts: 329 ✭✭✭
    bitcanuck said:

    Genoil said:

    Trying to fix an issue with the OpenCL kernel for GCN1.1 and up that I grossly overlooked for months...

    I wonder if this is not a windows-only thing. I do most of my mining on Linux (Ubuntu 14.04/fglrx), but now and then will test things on my Win7/64 desktop (Catalyst 15.7.1). Today I just tested out a R9 290x at stock 1000/1250 clocks, and noticed I was only getting 25.6-26Mh instead of ~28 that I get on Linux. I tried Claymore (-ethi 8), and get 28.2-28.3.

    My R9 380's (GCN 1.2) the same hashrate on Linux as they do on the Win7 machine, so maybe the issue is just GCN1.1 (i.e. Hawaii) on Windoze. I suspect Claymore uses a precompiled kernel (like @Wolf0 does), so it wouldn't be dependent on the OpenCL compiler version included in the driver.

    So maybe you don't have to re-write the kernel to get 3 concurrent waves, but instead you could build the kernel .bin and have ethminer use a kernel binary instead of opencl source.
    I told him this and offered to help.
  • SmokyishSmokyish Member Posts: 203 ✭✭
    I can get a stable 30-31MH/s per card with my 2x MSI 1070 Gaming X cards at 150w with OC, stock was 25MH/s.



    (Note: This is my gaming rig, so cards are in a case and connected straight to mobo, no risers and im running X, thus card 0 uses some overhead and has a lower clock&rate.)
  • AmphAmph Member Posts: 106
    Smokyish said:

    I can get a stable 30-31MH/s per card with my 2x MSI 1070 Gaming X cards at 150w with OC, stock was 25MH/s.



    (Note: This is my gaming rig, so cards are in a case and connected straight to mobo, no risers and im running X, thus card 0 uses some overhead and has a lower clock&rate.)
    is that on etheruem and windows 10?
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    It was a driver problem indeed. I'm still rewriting the kernel to optimize the first/last rounds of keccak, but still 1 VGPR short.
  • SmokyishSmokyish Member Posts: 203 ✭✭
    Amph said:

    Smokyish said:

    I can get a stable 30-31MH/s per card with my 2x MSI 1070 Gaming X cards at 150w with OC, stock was 25MH/s.



    (Note: This is my gaming rig, so cards are in a case and connected straight to mobo, no risers and im running X, thus card 0 uses some overhead and has a lower clock&rate.)
    is that on etheruem and windows 10?
    Ethereum and Linux Mint 17.3 (Ubuntu 14.04 "derivative").

    Also, @Genoil for some reason with these 1070's i can't set cuda-block-size to 256 or 1024, any idea why? I've tried with different -L modes too.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @Smokyish try halving grid size while doubling block size. This keeps intensity at the same level. I know, it's not very intuitive...

    1024 won't work, 896 is the highest the kernel supports.

    128 has always been optimal for Maxwell, so most likely for Pascal, too.
  • SmokyishSmokyish Member Posts: 203 ✭✭
    edited July 2016
    Genoil said:

    @Smokyish try halving grid size while doubling block size. This keeps intensity at the same level. I know, it's not very intuitive...

    1024 won't work, 896 is the highest the kernel supports.

    128 has always been optimal for Maxwell, so most likely for Pascal, too.

    Thanks for this, i will test to see if there are differences with different settings.

    Attempting to push the card to a max, but even though powerwise there would be room for more (the cards have 8+6pci connectors) it seems the driver doesn't allow to push much further, keep getting illegal memory access and end of file errors after a while with higher settings, 62MH/s is the best i've been able to get to mine for a longer period of time.

    I have no idea how @kruisdraad got 37MH/s with a FE card...
  • thesmokingmanthesmokingman Member Posts: 152 ✭✭
    Smokyish said:

    Genoil said:

    @Smokyish try halving grid size while doubling block size. This keeps intensity at the same level. I know, it's not very intuitive...

    1024 won't work, 896 is the highest the kernel supports.

    128 has always been optimal for Maxwell, so most likely for Pascal, too.

    Thanks for this, i will test to see if there are differences with different settings.

    Attempting to push the card to a max, but even though powerwise there would be room for more (the cards have 8+6pci connectors) it seems the driver doesn't allow to push much further, keep getting illegal memory access and end of file errors after a while with higher settings, 62MH/s is the best i've been able to get to mine for a longer period of time.

    I have no idea how @kruisdraad got 37MH/s with a FE card...
    @Smokyish - Ditto

    I have a Gigabyte G1 Gaming and Gigabyte FE GTX 1070s, and the highest I ever saw was 34 MHs at 18~1900 Mem clock, but the driver would crash shortly thereafter. 1600 Mem clock puts out a consistent 62MHs for both cards. @kruisdraad was using Genoil v1.1.5, so maybe there's a difference between the two? (I'm running 1.1.7) I've been trying to figure out how to build previous versions from source to test, but no luck so far.

    As an aside, does increasing the GPU core clock affect your hashrate? Only increasing the mem clock affected my hash rate, and just wondering if I have my coolbits set correctly, or Nvidia cards overclock using the mem clock vs AMD using the core clock.
  • workwork Member Posts: 2,084 ✭✭✭✭
    @thesmokingman that all depends on if the GPU is memory limited or compute limited.
  • SmokyishSmokyish Member Posts: 203 ✭✭

    Smokyish said:

    Genoil said:

    @Smokyish try halving grid size while doubling block size. This keeps intensity at the same level. I know, it's not very intuitive...

    1024 won't work, 896 is the highest the kernel supports.

    128 has always been optimal for Maxwell, so most likely for Pascal, too.

    Thanks for this, i will test to see if there are differences with different settings.

    Attempting to push the card to a max, but even though powerwise there would be room for more (the cards have 8+6pci connectors) it seems the driver doesn't allow to push much further, keep getting illegal memory access and end of file errors after a while with higher settings, 62MH/s is the best i've been able to get to mine for a longer period of time.

    I have no idea how @kruisdraad got 37MH/s with a FE card...
    @Smokyish - Ditto

    I have a Gigabyte G1 Gaming and Gigabyte FE GTX 1070s, and the highest I ever saw was 34 MHs at 18~1900 Mem clock, but the driver would crash shortly thereafter. 1600 Mem clock puts out a consistent 62MHs for both cards. @kruisdraad was using Genoil v1.1.5, so maybe there's a difference between the two? (I'm running 1.1.7) I've been trying to figure out how to build previous versions from source to test, but no luck so far.

    As an aside, does increasing the GPU core clock affect your hashrate? Only increasing the mem clock affected my hash rate, and just wondering if I have my coolbits set correctly, or Nvidia cards overclock using the mem clock vs AMD using the core clock.
    Yes, raising the coreclock affects the hashrate a bit, but i'm actually now underclocking it a bit for a more stable hashrate.
  • MadpacketMadpacket Member Posts: 25
    So what's the verdict with the 1070? Worth it or better off with 480's?
  • daemonixdaemonix Member Posts: 15
    hmm this doesnt look good..

    ℹ 17:16:25|cudaminer0 Initialising miner...
    ℹ 17:16:25|cudaminer1 Initialising miner...
    Using device: Tesla K80 (Compute 3.7)
    Using device: Tesla K80 (Compute 3.7)
    Trial 1... 27962026
    Trial 2... 27962026
    Trial 3... 28661077
    Trial 4... 27962026
    Trial 5... 27962026
    min/mean/max: 27962026/28101836/28661077 H/s

    Am I missing something?
    Using: nvidia-docker run -it anthonytatowicz/eth-cuda-miner -M -U -t 2

    Any special setting for a K80? (or 2xK40 :P )
  • SmokyishSmokyish Member Posts: 203 ✭✭
    Madpacket said:

    So what's the verdict with the 1070? Worth it or better off with 480's?

    Just read one or two pages back on this thread and you can make your own conclusions.
    There has been discussion on the matter and people have different opinions on the matter, so is it worth it is really up to many factors which you need to weigh yourself.
  • ChkoupiChkoupi Member Posts: 3
    edited July 2016
    Hello,
    After a lot of research about bitcoins, altcoins and mining (i'm totally new to all of this) i figured out i could try to mine with my computer and my GTX760 i bought for gaming so i've downloaded a lot of drivers, wallets and miners and someone on reddit (here https://www.reddit.com/r/EtherMining/comments/4sx3hv/newb_trying_to_get_into_ethermining/ ) told me that i should use genoil (since claymore doesn't work for Nvidia cards) so i first tried before changing anything, with open cl i had a hash of 1.5MH/s so i did more research and i installed CUDA 7.5, and the 347.52 driver but then cuda didn't work because the driver is too old so i had to get back on the latest driver and then i tried and i got around 3MH/s on one single block then again 1 to 2MH/s, i would really like to know how i could get a better hash i saw that a lot of people had between 10 and 15MH/s in this list http://62.212.74.86/~mining/list/index.php and fix whatever problem i have now, for the record i have a very very slow internet speed so i'm using poolmining and myetherwallet.com .

    This is the .bat file i use :

    setx GPU_FORCE_64BIT_PTR 0 setx GPU_MAX_HEAP_SIZE 100 setx GPU_USE_SYNC_OBJECTS 1 setx GPU_MAX_ALLOC_PERCENT 100 ethminer -U -S pool.alpereum.ch:3001 -O 0xA178dFABa746aEcE081Cf4e6825Db07aAf84e206.ChkouPCGPU --cuda-grid-size 8192 --cuda-block-size 128 pause
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @daemonix 28Mh/s is about right for K80. Very dissappointing indeed.
    @Chkoupi Use Win7 instead of Win10. I know..that doesn't sounds like a real solution but there is no real other.
  • daemonixdaemonix Member Posts: 15
    @Genoil Indeed! At least it does well of neural nets.
  • ChkoupiChkoupi Member Posts: 3
    @Genoil uhm i already have ubuntu 14.04 on dual boot will it work well ? i could wipe win 10 but it will be a lot of trouble for me :/
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @Chkoupi yes Ubuntu is also fine.
  • ChkoupiChkoupi Member Posts: 3
    @Genoil so on an attempt to try to avoid having to setup everything again i installed CUDA 6.5 and the 347.52 driver but i have a window that says "ethminer.exe" has stopped working (Windows is checking for solutions and all the jazz) and after a short moment i can close the program, so is it really impossible to run it on windows 10 ? i have some stuff on steam and on some other launchers i'd like to download while mining i would have to choose between mining or downloading (if i mine on ubuntu) :/
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @Chkoupi you need a exe built against CUDA 6.5. If I recall well, the 1.1.7 CUDA6.5 exe doesn't support the GTX760, so you'd either have to use the 1.0.5 that sits there or build your own.

    On a different note: the current Win10 Insider build (14390) has WDDM 2.1 activated again, so using this build with the WU driver that comes with it, should possibly solve the issues with Pascal. Don't havea Pascal card yet, so can't verify.
  • nisanisa Member Posts: 8
    Iam tested win10 14390 +newest vga driver.(gtx1070)
    No crash , smooth dag file creation , but i have 3-4Mh/s

    Anyone minning at win10+ pascal?
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @nisa try removing the driver and wait for Windows Update to install one. That may fix the issue (not sure though).
  • nisanisa Member Posts: 8
    4Mh/s -gtx1070

    driver-368.39
    genoil 1.1.7 cuda8
    win10 sr1 (14390)
    gpuz ---> Memory controller load 18%
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    If I remember well, the system of @vaulter that I saw this working on had 368.60.
  • AmphAmph Member Posts: 106
    try last one, which are .81
  • FingerFinger Member Posts: 11
    edited July 2016
    ~4 MH/s 1070

    driver-368.81
    also tested 368.95
    Win 10 insider (14393)
    genoil 1.1.7


    than i uninstalled the Driver and Let Windows 10 Insider Install a Driver (368.39) but dxdiag says WDDM 2.0 and i only have ~3 MH/s

    Somone has a link to an .exe file for this 2.1 Driver? (368.60 / 368.20 or something else? )
    Post edited by Finger on
  • SmokyishSmokyish Member Posts: 203 ✭✭
    @Finger I doubt there is a .exe to be shared, as the version with WDDM 2.1 can only be obtained via Windows Update, not a distributed exe file.

    If you can get the latest WIP, you should use Wagnards DDU to completely remove all previous driver installations, then let only Windows install your Nvidia drivers, do not install anything yourself, and then you can see if dxdiag still shows WDDM 2.0, not 2.1 to know if you can use your 1070 for Ethereum mining on Win10.
  • AmphAmph Member Posts: 106
    wait 2 august and you will be able to mine ethereum, they will release the drivers officially
Sign In or Register to comment.