CUDA miner

1192022242569

Comments

  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @syaoran99 yes you will get the full 17MH/s while benchmarking. The problem is in the DAG size in combination with Windows 8/10. Currently at 1280MB (as opposed to 1024MB during benchmarking) the GPU's TLB runs full and requires old memory page mappings to be removed from the TLB to make place for new ones. This is known as TLB trashing. The threshold for this to happen is 1024MB on Windows 10 with Maxwell 2 cards. But on Windows 7 and Linux it is 2048MB. See the below graph of approximate hashrate on GTX780 on Windows 7:




  • syaoran99syaoran99 Member Posts: 204
    Hi @Genoil I don't quite know much about these but if I'm understanding it right, it's only happening because I'm using windows 8/10?

    If I were on a windows 7 it would correctly show? What can I do to tweak the system?

    Or am I understanding it wrongly where I don't need to do anything and that I'm hashing at the right speed just that it's showing differently?
  • syaoran99syaoran99 Member Posts: 204
    Alternatively, is there any way to reduce the DAG size or requirement for running Windows 8/10? That way the memory can stay within the maximum output threshold of the GPU?

    Also with that you're saying if I can make space for 1024MB, the performance of my GPU will be significantly boosted? Sorry if this is a simple issue. I'm pretty new and have no programming knowledge whatsoever but I'm very interested in picking it up as I go :)

    Thanks again for the patience and responses!
  • syaoran99syaoran99 Member Posts: 204
    edited January 2016
    @Genoil

    I just googled about it and I found your thread here https://bitcointalk.org/index.php?topic=1268355.0

    So there's no simple workaround for it at the moment for non-programmers? D:

    Anyways, thanks once again for doing so much for the community. When I do get my ETH off mining, i'll definitely return the favor :) atm I don't even have 0.5 ETH no thanks to the TLB issue :(


    Post Edit: Sorry I should have read the pages before this prior to posting. It seems I'm not the only one suffering this but it's more so because I'm running on Windows 10. I guess if they do not fix this It's a matter of time before every GPU in the world will not be able to mine ETH.

    I hope that they either find a way to reduce the DAG or somehow split it up, as even if new drivers come out to optimize space, eventually there's a hardware limit from my understanding.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    Most cards will be fine on Windows 7 / Linux untile the switch to POS. It's mainly GTX750/ti that suffers now, as the threshold for TLB trashing is 1GB on Win 7 / Linux.

    Some older AMD GPU's also have TLB issues at the current DAG size. And then there is a potential (haven't completely wrapped my head around it, yet) threat for some AMD GPUs with DAG allocations above 1280MB, which may happen very soon. I may be mistaken about that, though.
  • ImAMiner?ImAMiner? Member Posts: 208 ✭✭
    @Genoil Which AMD cards are you thinking this could potentially happen to?
  • EastwindEastwind Member Posts: 107
    I think the 280x might have problem as when I run the test program, it cannot allocate 1408MB data.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @ImAMiner? See this post: https://bitcointalk.org/index.php?topic=1268355.msg13531406#msg13531406

    As you can see in the graph I posted earlier today, there also is a slight downslope in the curve, before the big drop @ 2048MB. It could very well be that this slight downslope is what AMD users are also experiencing.

    From what I've seen in other reports though, the AMD TLB drop on GCN 1.0 cards already starts after 1024MB, but has a less steep drop than what we see on Nvidia. The result however is that after ~1280MB (which is around this time), many AMD cards could become useless for ETH mining.

    I haven't got a lot of data on GCN 1.1 and 1.2 yet.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    Eastwind said:

    I think the 280x might have problem as when I run the test program, it cannot allocate 1408MB data.

    Yes. This is something else than TLB trashing, I dubbed it DAGpocalypse ;). This is why I built the chunky miner, but as you pointed out in the other thread, it doesn't find any solutions. The cause is probably in memorty fragmentation or some kind of misalignment, but it's impossible to fix without AMD hardware. That said, I don't have any use for AMD hardware other than fixing this potential issue :)
  • ImAMiner?ImAMiner? Member Posts: 208 ✭✭
    Thanks, I read that thread earlier too fast and missed that post. Subscribed to it now. I may have time to get some data for GCN 1.1 this weekend after I get my new geth and proxy working together.
  • ImAMiner?ImAMiner? Member Posts: 208 ✭✭
    DAGpocalypse haha. Somebody get @Genoil an AMD card quick. I'd like to say this would reduce the security of the network etc but really nvidia can't rake in all the ether...can they :neutral:

  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    edited January 2016
    @Eastwind try this on an AMD 280X. Unzip dagpocalypse next to ethminer. It's just ethminer with some hardcoded tweaks in it. You may need cudart as well becuase I'm to lazy to disable CUDA ;)
    dagpocalypse.exe -D 1000000
    ..wait it out...
    dagpocalypse.exe -G -S
    The first call precreates the DAG of DOOM ;). And then the second call will try to run a mining simulation on this 1350MB DAG.

    If it does indeed fail to allocate the GPU RAM for this DAG, there is indeed a need for a fix. Make sure to add all that GPU_MAX_ALLOC stuff that's required on some or all AMD cards.
  • ImAMiner?ImAMiner? Member Posts: 208 ✭✭
    @Genoil I'll try this on a 290x this weekend.

    I run your old ethminer, the one with --gpu-workgroup-size 92 --gpu-batch-size 20

    Will that make a difference.
  • grimlergrimler Member Posts: 12
    @Genoil Do you have a Link for me to your latest cuda ethminer? (windows binaries)
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @grimler the dagpocalypse.exe in my last post is just ethminer. If you run it with -U you'll have CUDA. You'll need some DLLs though but these are included in the cryptomining blog Zips.

    Btw it isn't any faster than earlier releases that can be found in this thread.
  • syaoran99syaoran99 Member Posts: 204
    @Genoil
    I've been doing a lot of reading lately and though I'm not a programmer nor have any knowledge in coding at all, I've tried to understand the issue.

    That said, only way I can think of helping out in a situation like this for non-programmers is to get you that AMD hardware that you need to hopefully run tests on? :D That is if you still haven't managed to get one.

    I'm thinking everyone would need to try and help to fix this otherwise we will be all unable to mine anything in a few month's time

    (well setting aside the possibility that Ethereum will be changing to PoS by June this year that is) which I doubt because at the moment I don't see how they have enough $$ and people to secure their "flaws" in the PoS system if I didn't understand it wrongly. Afterall, the combined value of all existing of Ethereum is but little and nothing significant that a band of businessmen can't buy let alone a centralised government if I'm not mistaken.

    That said, I was wondering if you require funding to get your hands on an AMD hardware and what hardware would you require? I think that it's high time the mining community assist with your efforts in helping them rack in $$ lol. An EthWallet for this purpose would be a good place to start hehe! I might just be a naive and young guy with a stupid idea tho lol.
  • grimlergrimler Member Posts: 12
    @Genoil Your dagpocalypse won't submit any shares neither does your chunky miner. (both in cuda mode with 970)

    So I am still looking for a ethminer cuda download?!

    (this thread only has very old ones)
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    The old ones are fine.
  • ImAMiner?ImAMiner? Member Posts: 208 ✭✭
    Genoil said:

    @Eastwind try this on an AMD 280X. Unzip dagpocalypse next to ethminer. It's just ethminer with some hardcoded tweaks in it. You may need cudart as well becuase I'm to lazy to disable CUDA ;)

    dagpocalypse.exe -D 1000000
    ..wait it out...
    dagpocalypse.exe -G -S
    The first call precreates the DAG of DOOM ;). And then the second call will try to run a mining simulation on this 1350MB DAG.

    If it does indeed fail to allocate the GPU RAM for this DAG, there is indeed a need for a fix. Make sure to add all that GPU_MAX_ALLOC stuff that's required on some or all AMD cards.
    How will dagpocalypse fail? The dag precreated successfully and after running dagpocalypse -G -S I have Preparing DAG.... and a blinking cursor. Did if fail to allocate the GPU RAM?
  • ImAMiner?ImAMiner? Member Posts: 208 ✭✭
    Nevermind. It was successful. It works with a 290x
  • ImAMiner?ImAMiner? Member Posts: 208 ✭✭
    Actually I'm not sure if it worked. I mean dagpocalypse was hashing but another DAG was created in my Ethash directory.

    So -D 1000000 created a DAG with a size of 1,318,908 KB but when I ran -G -S it looks like a second DAG was created with a size of 1,048,575 KB.

    Also the second DAG has a name of full-R23-000000000000...

  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @ImAMiner? it could be that the 0000 dag is created, but the bigger one used. Check GPU RAM usage while running it.
  • ImAMiner?ImAMiner? Member Posts: 208 ✭✭
    @Genoil You're right, Afterburner shows it using 1358MB

    Did @Eastwind ever get back to you whether his 280x allocate 1358MB?

    Do you want me to try any other DAG sizes?
  • EastwindEastwind Member Posts: 107
    See my result in the Chunky thread.
  • kenshirothefistkenshirothefist Member Posts: 73
    I have three NVIDIA GPUs in my rig (GTX 980, GTX 970, GTX 960) and I have a few questions about CUDA miner
    c:\Mining\ethminer-cuda-0.9.41>ethminer.exe -U --list-devices
    
    Listing CUDA devices.
    FORMAT: [deviceID] deviceName
    [0] GeForce GTX 980
            Compute version: 5.2
            cudaDeviceProp::totalGlobalMem: 4294967296
    [1] GeForce GTX 970
            Compute version: 5.2
            cudaDeviceProp::totalGlobalMem: 4294967296
    [2] GeForce GTX 960
            Compute version: 5.2
            cudaDeviceProp::totalGlobalMem: 2147483648

    1. How can I benchmark each individual GPU?

    ethminer.exe -U -M --phone-home off --cuda-devices 0
    or
    ethminer.exe -U -M --phone-home off --cuda-devices 1
    or
    ethminer.exe -U -M --phone-home off --cuda-devices 2
    or
    ethminer.exe -U -M --phone-home off --cuda-devices 0 1 2

    always runs benchmark only on the first GPU (device 0, GTX 980)!

    2. How can I run actual mining on a particular GPU only?

    This: c:\Mining\ethminer-cuda-0.9.41>ethminer.exe -U -F http://... --cuda-devices 0 1 2

    Using device: GeForce GTX 980(5.2)
    Using device: GeForce GTX 970(5.2)
    Using device: GeForce GTX 960(5.2)

    is OK. However, if I specify --cuda-devices 1 or --cuda-devices 2 it always starts mining on first device (gtx 980) only!

    c:\Mining\ethminer-cuda-0.9.41>ethminer.exe -U -F http://... --cuda-devices 1
    Using device: GeForce GTX 980(5.2)

    c:\Mining\ethminer-cuda-0.9.41>ethminer.exe -U -F http://... --cuda-devices 2
    Using device: GeForce GTX 980(5.2)

    Thanks for help!
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    edited January 2016
    @grimler, as interest in CUDA ethminer seems to raise these days (very understandable with current ETH value), I'll see if I can find some time today to fully test the code again and add the latest binaries to the repo. @kenshirothefist I believe I have addressed this issue earlier, so you may be using an older version of my fork. There are so many 0.9.41 mods I made that has become a complete mess :). That aid, I currently have only one GPU available, so I won't be able to test if it (still) works.

    @Eastwind your test seems to pass. I hope to get access to a Tahiti system today to further investigate why dagSimCL does seem to have these allocation issues, while ethminer doesn't. For now, no DAGpocalypse :)
  • kenshirothefistkenshirothefist Member Posts: 73
    Genoil said:

    @kenshirothefist I believe I have addressed this issue earlier, so you may be using an older version of my fork.

    @Genoil OK, so where can we find your latest version release binary so that I can check this out? Thanks.

  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    @kenshirothefist when it's done. I'm building it right now, generating DAG, then I'll be mining for a short while to see if it submits any shares. Then I'll posted the link here.
  • syaoran99syaoran99 Member Posts: 204
    @Genoil
    Genoil said:



    @Eastwind your test seems to pass. I hope to get access to a Tahiti system today to further investigate why dagSimCL does seem to have these allocation issues, while ethminer doesn't. For now, no DAGpocalypse :)

    Does this mean the DAG file being too large issue can be solved using the current ethminer?
  • kenshirothefistkenshirothefist Member Posts: 73
    edited January 2016
    @Genoil OK, great, thanks. Another short question: is your fork of ethminer just as good for OpenCL mining as well, or did you remove any code related to OpenCL mining? If not, who is the primary maintainer of OpenCL fork of ethminer (and where are latest bins)?
Sign In or Register to comment.