CPPminer: An Open-source Cross Platform Pearl Miner

[z1640675651]

An Open-source Cross Platform Pearl Miner

A cross platform pearl miner. Currently optimized for CPU and legacy GPU platforms, also aiming for wide device compatibility. Github: https://github.com/1640675651/CPPminer

Supported Devices & Features:
* Nvidia Pascal GPUs (GTX 10 series) via CUDA. Accelerated with DP4A instructions (1 instruction, dot product 4 int8 values). Almost zero overhead compared to pure CUTLASS GEMM.
* AMD/other GPUs via OpenCL. Accelerated with DP4A instructions if available.
* CPUs accelerated with AVX2 instructions.

Hashrate results:
Nvidia:
GTX 1070 ~8.5TH/s

AMD:
Radeon pro 5500M ~5.0TH/s

CPU:
Core i9 9980HK 2.4GHz ~440GH/s
Core i5 12490F 4.0GHz ~640GH/s

Dev fee:
1%

Supported pools:
Tested with luckypool. Support for other pools are planned.

[z1640675651]

V0.2 Release

*Use rank=128 to avoid pearl rank penalty (https://github.com/pearl-research-labs/pearl/pull/275). This adds a bit more overhead compared to rank=256, may hurt hashrate slightly.
*SSE fallback for non-AVX CPU.
*Add CPU thread affinity and prioritize physical cores than SMT threads.
*Use 8x8 tile as default for the opencl worker, lowering register pressure for integrated GPUs. Use 8x16 tile if detects AMD GPU.

[z1640675651]

V0.2.1 Release

Refactor CUTLASS GEMM main loop to reduce XOR overhead. This buys back the lost performance in the last version due to more frequent XOR boundaries. 8.0TH -> 9.1TH on a GTX 1070.

CUTLASS default GEMM produces 18 TOPs. Since 1 hash = 1 MAC (Multiply-Accumulate), and 1MAC = 2 OPs, 9 TH corresponds to 18 TOPs. This is real zero overhead compared to CUTLASS default GEMM.