Evaluating Heterogeneous Computing Platforms for AI-RAN

Summary

The evolution of Artificial Intelligence-enabled Radio Access Networks (AI-RAN) is transforming traditional cellular base stations into distributed computing platforms that integrate real-time radio communication, network optimization, and AI-driven edge services. These platforms rely on heterogeneous hardware, including CPUs, GPUs, specialized accelerators, high-bandwidth memory, and high-speed interconnects. However, the coexistence of latency critical radio functions and dynamic, compute and memory intensive AI workloads present significant challenges in performance predictability, resource utilization and energy efficiency.

This internship aims to characterize representative AI-RAN workloads on existing CPU/GPU/accelerator platforms to identify performance bottlenecks and derive hardware architecture requirements for future AI-RAN nodes. Using open-source software and experimental performance profiling, the student will investigate workload execution, resource utilization, memory behaviour, data movement, and interference between concurrent workloads. The collected measurements will support a data-driven assessment of existing hardware configurations, highlighting their limitations and potential improvements. The expected outcome is a workload-to-architecture framework that translates observed performance bottlenecks into recommendations for future heterogeneous computing architectures, supporting more efficient, predictable, and scalable AI-RAN infrastructure.

Desirable skills

• Programming: CUDA, C/C++.Python, GPU execution models,
• Telecommunications: Understanding about 4G/5G RAN architecture, PHY processing, radio resource management, or O-RAN concepts. (Exposure to Open-source RAN software like OpenAirInterface, srsRAN, or related platforms is a plus point.)
• AI/ML workloads: PyTorch, neural network inference, model optimization, or AI acceleration.
• Performance analysis: Linux perf, hardware counters, memory-bandwidth analysis, latency profiling, or power measurement, NVIDIA Nsight Systems, or Nsight Compute