This loop is mostly compute bound but may also be memory bound

The bottleneck depends heavily on the fused multiply-add (FMA) computational unit. The bottleneck depends greatly on the accessed computational unit. The performance of the loop is also bounded by the DRAM bandwidth.
The performance of the loop is also bounded by the bandwidth of the shared cache and DRAM.
The performance of the loop is also bounded by the private cache bandwidth.
The performance of the loop is also bounded by the L1 bandwidth.
The performance of the loop is also bounded by the L2 bandwidth.
The performance of the loop is also bounded by the L3 bandwidth.
The performance of the loop is also bounded by the L4 bandwidth.
The performance of the loop is also bounded by the DRAM bandwidth.
The performance of the loop is also bounded by the MCDRAM bandwidth.
To improve performance: Switch from %c_isa% to %b_isa% - the highest available instruction set architecture (ISA). Eliminate inefficient memory access patterns. %traits% instructions might degrade performance. The loop is scalar. To fix: Vectorize the loop.

Read More