So by adding ILP/SIMD, in the range between 1/2 to 2 on the x-axis, the limitation which used to be on computation becomes a bandwidth limitation?
This comment was marked helpful 0 times.
kayvonf
The program without SIMD/ILP optimizations ("TLP only") is bandwidth limited (assuming "unit stride accesses only" until the operational intensity hits 1/2 flops/byte. By increasingly compute throughout via SIMD/ILP optimizations, the program becomes bandwidth limited at 2 flops/byte.
So by adding ILP/SIMD, in the range between 1/2 to 2 on the x-axis, the limitation which used to be on computation becomes a bandwidth limitation?
This comment was marked helpful 0 times.
The program without SIMD/ILP optimizations ("TLP only") is bandwidth limited (assuming "unit stride accesses only" until the operational intensity hits 1/2 flops/byte. By increasingly compute throughout via SIMD/ILP optimizations, the program becomes bandwidth limited at 2 flops/byte.
This comment was marked helpful 0 times.