An Open Source Custom K-means Generator for AWS Cloud FPGA Accelerators

• Lucas Bragança UFV
• Michael Canesche UFMG
• Jeronimo Penha UFV / CEFET-MG
• Westerley Carvalho UFV
• Giovanni Comarela UFES
• José Augusto M. Nacif UFV
• Ricardo Ferreira UFV

Abstract

Nowadays, FPGAs play an essential role in domain-specific hardware accelerators due to their power efficiency and flexibility. Nevertheless, FPGA programming and deploying are still a challenge. We present an accelerator generator targeting Amazon Web Services (AWS) Cloud FPGAs. We validate our generator with the K-means clustering algorithm as a case study. The main contribution of this work is to present an opensource full-stack generator using AWS EC2 F1 cloud FPGAs. Our framework is parameterizable and generates Verilog code for the entire design, including DDR memory and FPGA communication. Furthermore, we use a modular design approach, allowing extensions to address other applications. Compared to the Intel/Altera Harpv2 cloud FPGA, our AWS accelerator is on average twice faster for the K-means case study.

Keywords: Codes, Web services, Software algorithms, Programming, Parallel processing, Systems engineering and theory, Generators, K-Means, FPGA, AWS, Accelerators