CTTC Accelerates 5G Network Simulation from Hours to Minutes with NVIDIA

GPU-accelerated 3GPP channel models in 5G-LENA

GPU-accelerated 3GPP channel models cut large-scale 5G-LENA simulation runtimes by up to two orders of magnitude

The Centre Tecnològic de Telecomunicacions de Catalunya (CTTC) brings NVIDIA accelerated computing to 3GPP-compliant system-level simulations, dramatically reducing the time required to run large-scale 5G network simulations.

The joint work is based on the implementation of 3GPP TR 38.901 spatial channel models in CUDA (a parallel computing platform and programming model created by NVIDIA) and WebGPU (a vendor-neutral and portable GPU API) and integrating them into ns-3 5G-LENA (CTTC’s open-source full-protocol stack 5G NR system-level simulator), as a drop-in GPU-based replacement for the existing CPU-based implementation. Statistical fidelity has been confirmed by a two-phase TR 38.901 calibration procedure and an end-to-end network performance comparison with the CPU baseline in ns-3 5G-LENA. As the integration preserves the standards-compliant behaviour of the original models, existing 5G-LENA users can benefit from the GPU acceleration without changing their simulation methodology. Also, GPU-inspired ns-3/5G-LENA code re-architectures are being analyzed.

The collaboration targets the most computationally demanding simulation regimes, i.e., scenarios with many cells, high numbers of users, and large antenna array sizes, where channel modelling and channel update times dominate runtime. By offloading these calculations to the GPU, the collaborative work aims to make 5G-LENA an even more capable platform for the demanding evaluations expected in future 3GPP studies. Extensive tests with various GPU platforms have also been carried out.

Early results are already significant. In a representative scenario with 21 macrocells, 210 users, and 8×8 (gNB) / 2×1 (UE) dual-polarized uniform planar antenna arrays, the GPU-accelerated implementation reduced simulation runtime from approximately 3 hours to 3 minutes, with 70x speedup.

The GPU-accelerated channel models and its integration are made available to the community through the NVIDIA CUDA repository and the ns-3 5G-LENA project.

5G-LENA: https://gitlab.com/cttc-lena/nr CUDA: https://github.com/NVIDIA/aerial-cuda-accelerated-ran/tree/main/testBenches/gpu3GPPChan