Skip to content

NoncerPro/noncerpro-nimiq-cuda

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

32 Commits
 
 
 
 
 
 

Repository files navigation

Nimiq GPU miner for Nvidia graphic cards.

This is the first Nimiq GPU miner that utilizes the standard Nano mining protocol and is compatible with all of the available nimiq pools.

Links

Github All Releases

Website: http://noncer.pro

Discord server : https://discord.gg/mXnyrHa

releases: https://github.com/NoncerPro/noncerpro-nimiq-cuda/releases

AMD miner: https://github.com/NoncerPro/noncerpro-nimiq-opencl/releases

Donation addresses

Please consider supporting this project by donating to these addresses (EhssanD):

  BTC  : 15h2QmsRwwwEdNNC6HbYHJU9mpbLrjUdDK

  NIM  : NQ52 7TL5 RA6B SSAS FULC P3SR D88B X2CK 0VTX

Options

  -a, --address         Nimiq wallet address                 [string] [required]
  -n, --name            Device name                                     [string]
  -s, --server          Pool server address
                        multiple addresses or a single string
                        Example: -s=eu.nimpool.io nimiq.icemining.ca
                        Default: eu.nimpool.io                           [array]
  -p, --port            Pool server port
                        Default: 8444                                    [array]
  -d, --devices         Active GPUs
                        Example: -d=0 1 3
                        Default: All available GPUs                      [array]
  -t, --threads         Number of threads per GPU
                        Example: -t=2 or -t=2 2 4
                        Default: 2                                       [array]
  -b, --batchsize       batchsize per thread.
                        Example: -b=100 or -b=100 120 200
                        Default: auto based on available device memory   [array]
  -i, --api             Enable/Disable API
                        Default: Enable                                [boolean]
  -o, --apiport         API port
                        Default: 3000                                   [number]
  --opt, --optimizer    Run parameters optimizer
                        Miner will ignore gpu.json file and rerun the optimizer.
                        Default: false                                 [boolean]
  --diff, --difficulty  Start difficulty
                        Default: Disabled                               [number]
  --ao, --autoOptimize  Auto run the optimizer if there is no gpu.json file.
                        Default: true                                  [boolean]
  -h, --help            Show help                                      [boolean]
  -v, --version         Show version number                            [boolean]

Config file

In addition to the above command line options, there is a miner.conf file which has the exact same options. If you uncomment any config option, it will be used as the default value for that option in the next run. Config options can still be overridden by command line options.

Requirements

Binary packages have been built with CUDA 10.0. You need to update your driver to the latest version in order to use them.

The minimum supported Compute Capability is SM3.5. You can check your cards' SM support in the following link: https://developer.nvidia.com/cuda-gpus

If you are using Windows, the minimum virtual memory must be set to the total amount of allocated gpu memory. So if you have 6 x 1080 tis and want to use 8GB on each card, you will need 48 GB of virtual memory. This only applies to Windows.

If you are running the noncerpro.exe file directly, make sure you have set UV_THREADPOOL_SIZE to atleast 32;

Usage

Download the binary packge for Windows/Linux, edit mine.bat/mine.sh file and insert your own wallet address and run it.

You can control the amount of memory allocation per thread by using --batchsize option. Here is how it works:

Allocated memory per thread = batchsize * 100 * 512 KB. For example batchsize 200 will try to allocate about 10 GB of memory.

Notice that this allocation is per thread, so if you have more than one thread you must adjust the batchsize accordingly.

Best settings

The default values are pretty good but for the best performance try different values for --threads and --batchsize options. Here are my best parameters for 1080 ti on Linux and Windows:

Linux: --threads=4 --batchsize=50
Windows: --threads=2 --batchsize=92

Overclocking

Argon2d is highly dependent on memory bandwidth and memory clock. Try settings the memory clock as high as possible, usually around +1000. Other parameters must be adjusted based on the specific card and OS.

HiveOS

This is how Flight Sheet must be set: HiveOS

Dev fee

This miner has a fixed 2% dev fee. That means 2 minutes in every 100 minutes, miner will run with the donation wallet address.

Happy mining!