The filterbank has a total of 4096 channels, each with 36 taps of type int32. The sampling rate of each channel is 2e9/4096 = 488.28125 Ksps.
Based on the specified data and coefficient types, AI Engine should be able to perform 8 cint16
x int32
MACs every cycle in a single tile.
For more information, refer to Table 1 of the Versal Adaptive SoC AIE-ML Architecture Manual (AM020).
Assuming we use part xcve2802-vsvh1760-2MP-e-S
, AI Engine can be clocked at 1.25 GHz, as described in Versal AI Core Series Data Sheet: DC and AC Switching Characteristics (DS957).
A general rule of thumb is to reserve some margin for processor overhead in the range of 20-25%.
The number of tiles required based on compute-bound analysis = 2e9 x 36 / 8 / 1.25e9 x 1.25 = 9 tiles.