Towards a Methodology to Leverage Alveo Versal System Usability And Parallelization

Versal system arises to deliver an energy-efficient on-field reconfigurable fabric with the parallel floating-point computations of AI Engines (AIE). Despite the success of various implementations, the steep learning curve and the lack of a well-defined programming model make them inaccessible to non-experts. Therefore, we present a methodology and a parallelization strategy automatized through an automation framework. Applying the proposed methodologies to the case of similarity metrics computations, we attain over 13× speedup with respect to a single AI Engine-based solution, leading to 6× speedup and 8× energy efficiency improvement against software solutions.