PASTA: A Modular Program Analysis Tool Framework for Accelerators

Researchers Mao Lin and Hyeran Jeon have introduced PASTA (Program Analysis Tool Framework for Accelerators), a breakthrough framework addressing the growing complexity of hardware accelerators in modern computing systems. The tool provides a modular, low-overhead solution that abstracts over diverse deep learning frameworks and low-level profiling APIs, offering researchers and practitioners a unified interface to capture and analyze runtime events across multiple levels. This comes at a critical time as AI workloads increasingly run on heterogeneous hardware environments spanning NVIDIA and AMD GPUs.

PASTA's technical innovation lies in its GPU-accelerated backend that delivers performance analysis with dramatically reduced overhead—up to 13,000x faster than conventional analysis tools. The framework has been extensively evaluated on mainstream deep learning workloads tested across both single- and multi-GPU scenarios, demonstrating practical applications including deep learning workload characterization and Unified Virtual Memory (UVM) optimization tools. Its extensible design enables rapid prototyping of custom analysis tools while maintaining the efficiency needed for production environments, striking an optimal balance between usability, extensibility, and performance for modern accelerator-based computing.