So I found in the CUDA Toolkit documentation that
CUBLAS_STATUS_ARCH_MISMATCH: cublasGemmEx is only supported for GPU with architecture capabilities equal or greater than 5.0
My K40 and K80 GPUs are 3.5
Would downgrading CUDA and CuPy to a lower version help, or should I just kill myself with a 5-year old $30K machine that is useless now?