The FFT uses many memories for input, output, internal scratchpad, twiddles and sysmem (heap and stack). Should any two of these be placed on the same bank, then memory collisions can occur, causing wait states and a loss of performance. If performance is not meeting your needs, consider the placement of memories. This may be improved using the BufferOptLevel argument to aiecompiler, or an improvement may be possible by applying location constraints to buffers.