even with a delayed carry, you still need to pump the data through the CPU. For a basic place where this becomes a problem, consider taking exp of each element

even with a delayed carry, you still need to pump the data through the CPU. For a basic place where this becomes a problem, consider taking exp of each element in a vector. A vectorized version of this code will spend most of it's time computing a polynomial which is just 8 fmas (in 2 chains of 4). With a dedicated fma instruction, the cpu can do each of those fmas in 4 cycles, and overlap the chains, leading to a total time of 17 cycles (assuming 4 cycle fma which is fairly standard). Doing the same computation with quires would require moving around 16x as much data, which is impossible to do as quickly.