Workshop: 2026 International Workshop on Performance, Portability, and Productivity in HPC
Authors: Dzmitry Malyshau (Independent Researcher)
Abstract: Meganeura explores Vulkan and Metal as a common foundation for GPU inference and training. Implemented in Rust, it combines computational graph compilation with bounded, automatic kernel tuning, aiming to keep the engine small and general while adapting to different hardware. We examine this approach across NVIDIA, AMD, Intel, and Apple GPUs, using PyTorch as a reference. Our evaluation considers not only steady-state performance, but also preparation time, numerical validation, memory use, and practical deployment barriers. We discuss how graphics APIs can support machine learning alongside rendering, where their execution model helps or limits performance, and what observability is needed to understand those trade-offs. The work asks how much performance portability a compact engine can achieve without maintaining separate implementations for each vendor.
Back to 2026 International Workshop on Performance, Portability, and Productivity in HPC Archive Listing
Back to Full Workshop Archive Listing