Presentation
Skew-oblivious Data Routing for Data Intensive Applications on FPGAs with HLS
Time
Location
SessionScaling the Data Wall
Event Type
Research Manuscript
Hosted in Virtual Platform
SoC, Heterogeneous, and Reconfigurable Architectures
Design
DescriptionAlthough HLS eases the programming of FPGAs, data buffering with BRAMs and multiple PEs need to be expressed explicitly for irregular data-intensive applications.
Data routing based designs embrace exclusive buffers for PEs for high BRAM usage efficiency. However, skew datasets may cause the workload imbalance issue.
This paper proposes a skew-oblivious data routing architecture that schedules secondary PEs to help overloaded primary PEs. With it, we further propose a framework to ease the mapping of data-intensive applications.
The results of five applications show implementations are robust to skew datasets and outperform state-of-the-art designs in both throughput and BRAM usage efficiency.
Data routing based designs embrace exclusive buffers for PEs for high BRAM usage efficiency. However, skew datasets may cause the workload imbalance issue.
This paper proposes a skew-oblivious data routing architecture that schedules secondary PEs to help overloaded primary PEs. With it, we further propose a framework to ease the mapping of data-intensive applications.
The results of five applications show implementations are robust to skew datasets and outperform state-of-the-art designs in both throughput and BRAM usage efficiency.