Testnet-mining performance benchmark

acrosacros Member Posts: 63 ✭✭
Hi all!

I was just curious to know what kind of performances people were getting on the testnet with their setup.

Personally I got ~88khs with CPU mining on Intel Xeon (4 cores)/16GB RAM and Ubuntu 14.04

It looks like @larz got ~130khs with his setup (Ubuntu 14.04 on a Core i7 with 16GB RAM)

I'd be curious to know what others are getting! Maybe even some GPU mining performances if someone managed to create such miners already (expected to be ~x100 better I believe).



  • larzlarz Sussex, EnglandMember Posts: 17
    hi @acros,

    curiously my hash rates vary between 60-80kHs with the C++ implementation (eth -m on -f and neth -m on -f) and the 130kHs I've mentioned in the other thread I get out of the go implementation (ethereum --mine) on the exact same hardware.
    Whilst the go implementation seems to be happily mining away, updating the hash rate (which varies between 125 and 130KHs), I do get a peer not found error at the very beginning (as discussed in the other thread), so I'm wondering is it actually mining at all or is this just local?
    So my stats might have to be taken with a grain of salt...
  • acrosacros Member Posts: 63 ✭✭
    Interesting, I'm also using the go implementation for the hash rate I mentioned. Maybe someone from the go team could comment on that message ("[P2P NAT] mapping error: no devices discovered") and clarify whether mining really does take place or not? I'm not sure who to ask though, @StephanTual maybe?
  • larzlarz Sussex, EnglandMember Posts: 17

    These screenshots were taken within a minute from one another. neth runs at 40kHz and ethereum seems to run at close to 130kHz.
  • acrosacros Member Posts: 63 ✭✭
    I also tried on another machine I have (AMD A6 7400K dual-core 3.5GHz/8GB RAM/Ubuntu 14.04) and got ~84khs (go-ethereum only)
  • carloscarlsoncarloscarlson Member Posts: 10
    edited March 2015
    Hey guys, I'm somewhat new to all of this, but I figured that I would give you my data.

    I am running an Intel i7-4770K @3.5ghz, with 16gb ram. It's running Ubuntu 14.10, with no video card currently installed.

    I'm getting a pretty steady 136 khash

    There is no GPU miner publicly available is there? If so, I would love to see some specs on GPUs tested.
  • acrosacros Member Posts: 63 ✭✭
    Hey @carloscarlson, thanks a lot for the input!

    AFAIK it doesn't make use of the GPU at this point anyway, though it has been said on reddit that Frontier wouldn't launch without a GPU miner (for fairness).

    Which code base did you use by the way?
  • carloscarlsoncarloscarlson Member Posts: 10
    edited March 2015
    It was the go implementation, using the command line.
  • MarioFortierMarioFortier Boston, MAMember Posts: 30
    edited March 2015
    Comment taken from GPU/OpenCL code:

    // author Tim Hughes
    // Tested on Radeon HD 7850
    // Hashrate: 15940347 hashes/s
    // Bandwidth: 124533 MB/s

    (Reference: GitHub Link )

    I have not verified such performance, just reporting here what I read.

    Keep in mind this seems to be a raw measurement done with the ethash benchmark, not a measurement while integrated to run against testnet... though it gives us a glimpse of what to expect.
  • acrosacros Member Posts: 63 ✭✭
    Thanks Mario, the 15,940,347 (~16mhs) number would seem a bit high considering that 130khs seem to be the trend with CPU (except for me!) and that the GPU mining is expected to be ~100x time faster, namely ~1,3mhs. Though as you say, we don't have much context on the code comment!

    What does bandwdith impact exactly?
  • MarioFortierMarioFortier Boston, MAMember Posts: 30
    edited March 2015
    (~130K x ~100 = ~13M)
    acros said:

    What does bandwdith impact exactly?

    The bandwidth displayed by Ethash benchmark is the number of MBytes/s read from main memory.

    In latest code, I see that it is not an independent measurements. It is simply a function of the hashrate:

    bandwidth = hashrate * hashimoto access * bytes read per access = hashrate * 64 * 128 bytes

    This is likely calculated just to check if near the hardware limit.
    Post edited by MarioFortier on
  • MarioFortierMarioFortier Boston, MAMember Posts: 30
    edited March 2015
    I get ~90 khashes/s on a Xeon E5-2609 v2 @ 2.5 GHz (Ethash benchmark, Ubuntu, single threaded).

    From these numbers I assume the Go implementation is single threaded (I don't know the language well enough to be sure).

    Can someone familiar with the GO implementation confirm about the CPU mining being currently single threaded or not?
  • acrosacros Member Posts: 63 ✭✭

    (~130K x ~100 = ~13M)

    Typical me... o:)

    You seem to be getting a similar hashrate as I do. Thanks for clarifying the bandwith bit btw! I'm not overly familiar with Go either, hopefully someone can jump in on that.

  • MarioFortierMarioFortier Boston, MAMember Posts: 30
    edited March 2015
    GO Ethereum was single threaded, just spotted changes done ~4 hours ago to make it multithreaded.

    Using my modified Ethash benchmark for multithread tests shows a hashrate increase to be a bit less than a multiple of the number of hardware cores (as expected).

    I get ~700 khashes/s with 16 threads running over 8 hardware cores. The setup is a dual socket Xeon E5-2609 v2 @ 2.5 GHz.

    To me, this is further confirming that CPU mining will not be economically competitive compare to GPU mining rigs (as intended by the design).
  • ConradJohnsonConradJohnson Member Posts: 130 ✭✭
    I'm still not seeing any arguments in there to access GPUs installed on the machine.
  • joshuad31joshuad31 Member Posts: 30
    This may be too much to ask but is there someone who can take an educated guess at the setup to purchase to mine Ethereum today and what might benefits there might be for a GPU mining solution in the future? I am willing to pay a .05 btc bounty for this information if someone can explain the following:
    1. How much better is it to mine with Ubuntu vs Windows
    2. If I have 2MB up and 20MB download connection speed does it make sense to pay an extra 10 dollars a month to get more bandwith to be able to mine more?
    3. Can you give me three side by side comparisons of two medium costs systems and one expensive system?
  • ImAMiner?ImAMiner? Member Posts: 208 ✭✭
    1. There is no difference
    2. Your connection is just fine
    3. You basically just need a normal computer, you don't need a top of the line cpu or ram or hdd. If you plan on running more than one gpu, you will need a motherboard that can handle that and usb riser cards. Your expense is going to come down to which gpus you want to run and how many and then that will determine which power supplies you need, which are going to be expensive as well. If you are going to run multiple cards there are going to be other expenses as well, structure to hold all the parts ( I just use adjustable shelves from Home Depot ), how to get all that power to the cards, cooling etc. So basically a one card setup can cost you just the card if you can build a computer from spare parts, but a 'rig' with 6 top of the line cards running off one motherboard can easily cost more than $4000. That's probably conservative too. In general nvidia cards are more efficient than AMD's but AMDs are faster. They could be on equal playing ground depending on how dagger hashimoto turns out.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    As long as there is no CUDA implementation I would stay away from NVidia. Their OpenCL drivers are way behind on AMD's.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    edited April 2015
    That said, I just ran the OpenCL benchmark on a GTX780 (OpenCL 1.1) and got this output:
    hashrate: 15094211, bw: 117923 MB/s

    But that kind of points out the weak NVidia OpenCL implementation, as in gaming circumstances, the GTX780 would blow the HD7850 mentioned earlier out of the water.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    And I just clocked 16Mh/s on an AMD R9 270 card.

    It may be useful for people that are planning to purchase hardware that having plenty of VRAM is critical to GPU mining of ether. This is because the full Dagger dataset is loaded onto the GPU. For now that's 1GB, but it will eventually grow to more GB than current cards have.
  • acrosacros Member Posts: 63 ✭✭
    Thanks for the info @Genoil! Which client did you use for mining with that AMD card? I'd like to see what my hashrate is too!
  • ImAMiner?ImAMiner? Member Posts: 208 ✭✭
    @acros I think he just ran the OpenCL benchmark, no mining software yet.
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    Yep the cpp OpenCL benchmark (on Win32).
  • GenoilGenoil 0xeb9310b185455f863f526dab3d245809f6854b4dMember Posts: 769 ✭✭✭
    I only get about 22Kh/s on the ethash cpp CPU benchmark, single-threaded. CPU is a XEON E5-1620, 3.6Ghz. That's really low compared to what @MarioFortier is reporting. Then again, with 8 threads on this CPU it would be about 170Kh/s and then seems to match with what the rest is reporting.

    Also tried to build against Intel OpenCL, but it wouldn't let me allocate 1GB of RAM for the dag buffer.
  • MarioFortierMarioFortier Boston, MAMember Posts: 30
    @Genoil, thanks for sharing benchmarking information.

    Assuming your setup has fast memory, then may be the lower CPU single threaded rate you got is related to the compiler not doing auto-vectorization. Just guessing.

    I use the Intel compiler on Linux. Using their vector report option I did verify that vectorization was done to match the hardware (AVX).

    I will start working on GPU optimization in about one week, and will give a shot to Intel OpenCL also. I will post my findings here.

  • MarioFortierMarioFortier Boston, MAMember Posts: 30
    If anyone has benchmark info about Intel Xeon PHI cards, then please let me know.

    I did re-write the mining code (for larger vector size and prefetching optimization) and I can hardly break above 5 MH/s on the 57 cores 31S1P model. Anyone found ways to do better with Xeon PHIs?
Sign In or Register to comment.