This exact multiply-then-horizontally-reduce pattern is what BLAS's sdot/ddot routines implement at the core of the matrix multiplies underlying PyTorch and TensorFlow inference. The multiple-accumulator trick you use to shorten the dependency chain is a real technique production numerical libraries apply specifically because a single running sum would bottleneck throughput on modern superscalar CPUs.