follow dev70_1, referring to PyQCU/examples/pyqcu/conftest.clover.bistabcg-v20260505.log, I've implemented a bistabcg implementation using CUDA C++ as the matvec and Python as the main loop. Its single iteration time is about 20% slower than the pure CUDA C++ version, but the total number of iterations is about twice as many. I'm currently comparing the differences in the main loops of the two implementations.