v0.2.1-GPU-0.2

A plugin based release which decouples CUDA and OpenCL from the miner.

The plugins are shared objects that sit in the miner directory. Currently plugins are whitelisted, so only kaspacuda and kaspaopencl are allowed. Each plugin manages its own options, so if plugin fails to load, options will be missing.

Unified CUDA miner to a single kernel, boosted mining by 250%.
OpenCL miner should now work, but is still un-optimized.

Improvements

• 250% improvement in CUDA miner
• miner works with CUDA and OpenCL as plugins
• Miner will still run if libraries are missing, but relevant plugin will be disabled

Breaking changes

• --wrokload and --workload-absolute are now --cuda-workload and --cuda-workload-absolute
• --no-gpu is now --cuda-disable

v0.2.1-GPU-0.1 (Devfund, Performance Boost, Experimental OpenCL)

Improvements

• Optimization of memory and work in kernel - expected 30% improvement
• Devfund is now part of the miner (can be turned off)
• Experimental OpenCL integration. Performance was worse compared to same NVidia card. Specification of GPU device is done via --opencl-device and the index number of the device (starting from 0). The OpenCL capability seem not to work for AMD (was tested and ran successfully on NVIDIA)
• Added --no-gpu option to mine only CPU

Note:
Currently OpenCL and CUDA cannot run together.
OpenCL part was not tested thoroughly, and is subject to possible crashes

v0.2.0-GPU-0.1 (Merging with Elichai's)

Improvements

• Mining only most up-to-date block (reduce red block counts)
• Less crashes due to network disconnects
• CPU miner part has asm code now (Elichai)

Breaking Changes

• --workload now denotes the scale of maximum occupancy. To use the absolute number, add --workload-absolute to the command

v0.1.3-GPU.0.2 (Service Update, Rereleased)

This version's default workload is quite conservative. People reported good results for ~5x the default workload

Improvements

This version improves the stability of the miner, and issues more informative error messages. The following improvements are included:

• After submitting a block, wait for a response from kaspad or 500 miliseconds (reduces number of red blocks, while not hurting the effective hash rate)
• Estimate the number of concurrent hashes each GPU can handle, and use this value as the default workload
• Added support for older GPUs, starting from compute capability 3.0 (run deviceQuery to check your device)
• The miner no longer crashes if a single thread fails

Breaking changes

• --gpu-threads was replaced by --cuda-device which allow to specify which GPU devices by id. Use multiple --cude-device for multiple GPU. By default, the miner will work on all GPUs available to CUDA.
• GPU threads are now counted separately from CPU threads. The default number of CPU threads is #cores - #gpus. Setting -t 0 will disable CPU.
• --workload can appear multiple times, setting the workload for each corresponding --cuda-device. Usage is discouraged - workloads is now calculated automatically.
• The default workload was reduced considerably. If you see reduction in performance, try to set it to 5x the reported workload

Linux Binary

Due to issues with GitHub workflows I did not resolve yet, the auto-compiled version kaspa-miner-v0.1.3-GPU.0.2-rerelease-linux-amd64 does not work for everyone. In case of library not found error, please use kaspa-miner-v0.1.3-GPU.0.2-rerelease-linux-amd64-v2

v0.1.3-GPU

• Initial release of the GPU support

SHA256

b206ae089fabf9ee44fad664681ba59ee26bc789286ae099e8dfcaa8a5c47dfa  kaspa-miner-v0.1.3-GPU-linux-amd64
81d1c1cfca4accfdce5f8f8d119e1b029c31cbd80bc5bda58321ec8bda79f88f  kaspa-miner-v0.1.3-GPU-win64-amd64.exe