Evaluation of the Versal Intelligent Engines for Digital Signal Processing Basic Core Units

In this work, the novel Intelligent Engines of the Versal Adaptive Compute Acceleration Platform (ACAP) from AMD (formerly Xilinx) are evaluated for Digital Signal Processing (DSP) applications. Different configurations of Finite Impulse Response (FIR) filters and Fast Fourier Transforms (FFTs) serve as benchmarks. The Intelligent Engines comprise the new DSP blocks generation (DSP58) and the AI Engines (AIEs). As FIR filters and FFTs are core units in almost all DSP algorithms, this benchmark is of great value for understanding the best way to implement DSP algorithms on these new devices. A comparison of several implementations using the Programmable Logic (PL) and DSP blocks of Zynq Ultrascale+ and Versal families, along with the scalar and vector units of the Versal's AIEs, is provided. Throughput, latency, power consumption and hardware cost are measured as key performance indicators for the algorithms under test. Results highlight the Versal architectural advantages in optimizing compute-intensive applications and its potential in terms of throughput, power and resources efficiency. The outcomes also provide valuable insights into the optimal configurations for DSP algorithms, depending on the used data types, filter taps, FFT length, as well as the latency, throughput, power and resources constraints of the target application.