ARIES Paper Accepted to FPGA 2025 and Nominated for Best Paper!
ARIES tackles the growing complexity of programming AMD’s AI Engine (AIE) architectures by introducing a flexible and unified MLIR-based compilation flow. It dramatically boosts programming productivity by enabling seamless mapping of tasks across AIE cores and FPGA fabric, all while supporting fine-grained instruction-level and tile-level parallelism through a clean abstraction. With impressive throughput on GEMM and up to 22.58× speedup on ResNet layers compared to state-of-the-art, ARIES sets a new standard for performance and portability in reconfigurable computing.
Congratulations to the authors!
Check it out on GitHub: https://github.com/arc-research-lab/Aries.
Yixiao Du
publication