“The time required to read the entire memory capacity of a GPU (capacity divided by bandwidth) tends to be around 20 milliseconds for several generations of High Bandwidth Memory (HBM).”