The vector unit in the first generation AI engine architecture supports the datatypes shown in Fig. 6.
Fig. 6: Datatypes supported in the vector unit of the 1st generation AIE architecture
Note that the rightmost column shows the number of MAC (multiply-accumulate) operations that one tile can perform in one cycle. Thus, with 8-bit operands, one AIE tile can perform 128 MACs/cycle. Fig. 7 shows what calculation is actually performed in one cycle.
Fig. 7: Int8 MACs/cycle for 1st generation AIE architecture
Thus, applications utilizing matrix-matrix or matrix-vector multipliations such as polyphase filters (also called channelizers), FIR/IIR filters, FFTs, beamforming, MIMO signal processing and many others can benefit greatly from using AI Engines.
Other tasks like image, video, audio procecssing, scientific simulations, data compression, networking, speech recognition, machine learning, cryptography and others may also be able to take advantage of the SIMD capabilities of the AI Engine.
Note that “peak theoretical compute capability” mentioned in Fig. 7 is an upper bound, and can never be realized when solving a practical problem. There are periods during which no calculations can be performed to allow data ingress and egress, or when the compute units need to wait for the results of a previous operation to become available.