My paper, “Provably Correct, Asymptotically Efficient, Higher-Order Reverse-Mode Automatic Differentiation” explains why this is slow, in section 8.3. Basically, the vector storing the gradients has no hope of being efficiently packed, let alone vectorised. There is some work on improving that over at GitHub - Mikolaj/horde-ad: Higher Order Reverse Derivatives Efficiently - Automatic Differentiation library based on the paper "Provably correct, asymptotically efficient, higher-order reverse-mode automatic differentiation", but I don’t know what their state of progress is.