3 ms·

Our GPU implementations are a lot more optimized. This is pretty far behind MKL etc for matrix multiplication