Hardware-Aware Acceleration of VLAs on the Spatial NPUs (abstraction)
Developing hardware-aware performance models and optimization methods for mapping matrix multiplication, attention, and transformer workloads onto spatial AI accelerators. The work focuses on analytical cost modeling, tiling, dataflow selection, spatial mapping, and scheduling under hardware resource and latency constraints. Contributions to AMD's open-source MLIR-AIR compiler include optimized GEMM and FlashAttention kernels, hardware-aware tiling, 2D attention mapping, A-stationary GEMM lowering, and compiler improvements. These compiler improvements significantly accelerate LLMs and SmolVLA; as tracked in PR #2085, they deliver 53.0% lower vision-stage latency and 43.0% lower end-to-end latency compared to AMD's original MLIR implementation, while outperforming both Ryzen AI Software and FastFlowLM.


