Profiling Concurrent Vision Inference Workloads on NVIDIA Jetson -- Extended

TL;DR

This paper analyzes resource utilization in concurrent vision inference workloads on NVIDIA Jetson devices, revealing underutilized GPU components and CPU bottlenecks, and offers insights for hardware-aware optimizations.

Contribution

It provides a comprehensive profiling methodology and detailed analysis of GPU and CPU resource sharing in edge vision inference workloads, highlighting bottlenecks and optimization opportunities.

Findings

GPU utilization can reach 100% with optimizations

SMs and tensor cores often operate at 15-30% utilization

CPU-side events frequently cause performance bottlenecks

Abstract

The proliferation of IoT devices and advancements in network technologies have intensified the demand for real-time data processing at the network edge. To address these demands, low-power AI accelerators, particularly GPUs, are increasingly deployed for inference tasks, enabling efficient computation while mitigating cloud-based systems' latency and bandwidth limitations. Despite their growing deployment, GPUs remain underutilised even in computationally intensive workloads. This underutilisation stems from the limited understanding of GPU resource sharing, particularly in edge computing scenarios. In this work, we conduct a detailed analysis of both high- and low-level metrics, including GPU utilisation, memory usage, streaming multiprocessor (SM) utilisation, and tensor core usage, to identify bottlenecks and guide hardware-aware optimisations. By integrating traces from multiple profiling tools, we provide a comprehensive view of resource behaviour on NVIDIA Jetson edge devices under concurrent vision inference workloads. Our findings indicate that while GPU utilisation can reach $100\%$ under specific optimisations, critical low-level resources, such as SMs and tensor cores, often operate only at $15\%$ to $30\%$ utilisation. Moreover, we observe that certain CPU-side events, such as thread scheduling, context switching, etc., frequently emerge as bottlenecks, further constraining overall GPU performance. We provide several key observations for users of vision inference workloads on NVIDIA edge devices.