4 ms·

A lot of it is also backlog: modern GPUs are optimised for these workloads in a way old models weren’t