Sionna: GPU-Accelerated Wireless Simulation

• Sionna is an open-source, GPU-accelerated framework for simulating wireless physical layers using fully differentiable TensorFlow models.
• It features a modular design with Sionna RT, a differentiable ray tracing engine that enables gradient-based optimization of transceiver and environmental parameters.
• The framework integrates with digital twin and network simulators, offering precise channel modeling and facilitating advanced machine learning research in wireless communications.

Sionna is an open-source, GPU-accelerated simulation framework for next-generation wireless physical layer research, developed with a modular architecture on top of TensorFlow. The framework enables link-level and end-to-end system simulations with native support for differentiable building blocks, enabling joint design and optimization of communication systems and environments. The introduction of Sionna RT, a differentiable ray tracing engine for electromagnetic wave propagation, marks a paradigm shift in the simulation of radio channels by enabling gradient-based optimization of both transceiver parameters and physical environments. Sionna has been widely adopted in the academic and industrial research community for its extensibility, performance, and ability to interface with state-of-the-art machine learning tools (Hoydis et al., 2022, Hoydis et al., 2023, Aoudia et al., 30 Apr 2025).

1. Design Principles and Architecture

Sionna is structured as a highly modular, extensible, and fully differentiable link-level simulator built on TensorFlow and Keras. All core components are implemented as tf.keras.layers.Layer, allowing seamless composition in functional or sequential graphs. The framework is optimized for GPU acceleration with support for mixed precision (tf.complex64 by default), vectorized batch processing, and integration of XLA JIT compilation for further speedup. Multi-GPU execution is compatible via standard TensorFlow distribution strategies (Hoydis et al., 2022).

The major modules include:

• Signal sources and mappers (bit generation, modulation)
• Forward error correction (5G LDPC, Polar, convolutional, Turbo codes)
• Channel models: AWGN, 3GPP stochastic, ray tracing via Sionna RT
• MIMO blocks (custom antenna arrays, ZF/MMSE, linear detectors)
• OFDM (with flexible frame/slot structure, pilot abstraction)
• Advanced link simulation, metrics (BER, BLER, mutual information)
• Neural network components (can replace or augment any chain element)
• End-to-end differentiable TensorFlow graphs for joint optimization

The "Link API" connects these blocks into full-system simulations, propagating gradients automatically through the entire communication chain for end-to-end learning.

2. Sionna RT: Differentiable Ray Tracing Engine

Sionna RT integrates a GPU-accelerated, differentiable ray tracing engine for physical modeling of site-specific radio propagation, supporting joint optimization of system and environmental parameters (Hoydis et al., 2023, Aoudia et al., 30 Apr 2025). Its architecture is characterized by:

• Scene Manager: Imports 3D city or indoor models (OBJ, CityGML/PLY), assigning material properties (relative permittivity $\epsilon_r$, conductivity $\sigma$) to polygons.
• Ray Launcher: Employs shooting-and-bouncing rays (SBR) along a Fibonacci lattice; supports up to configurable reflection depth and single diffraction per path.
• Path Processor: Computes per-path delay $\tau_p$, complex gain $a_p$ (including multi-bounce reflection and transmission coefficients), angles of arrival/departure, and Doppler.
• Channel Synthesizer: Produces the channel impulse response (CIR)

$$h(\tau, t) = \sum_{p=1}^{P} a_p\,e^{j2\pi f_{D,p} t}\,\delta(\tau-\tau_p)$$

• Differentiable interface: Uses TensorFlow and Dr.Jit to support automatic differentiation with respect to any environmental or system parameter.

CIR computation for link-level use leverages SBR and the image method, with a hashing-based mechanism for path de-duplication, while coverage maps use pure SBR. All field transformations are handled as TensorFlow ops, ensuring GPU compatibility and participation in the compute graph.

3. Mathematical and Algorithmic Foundations

Sionna RT models the radio channel deterministically based on physical scenes and electromagnetic material properties. The core models are:

• Per-path complex gain:

$$a_p = \prod_{i=1}^{N_p} \Gamma_i \times \frac{\exp(-j2\pi f_c \tau_p)}{4\pi\,d_p/\lambda}$$

where $\Gamma_i$ are reflection/transmission coefficients at bounce $i$, $d_p$ is the path length, and $\lambda$ the wavelength (Pegurri et al., 2024).

• Reflection coefficient (for a planar surface):

$$\Gamma(\epsilon_r, \sigma, \theta_i) = \frac{\eta_2\cos\theta_i - \eta_1\cos\theta_t}{\eta_2\cos\theta_i + \eta_1\cos\theta_t}$$

with $$\eta_k = \sqrt{\mu_k/(\epsilon_k - j\sigma_k/\omega)}$$ (Hoydis et al., 2023).

• Channel Impulse Response (CIR):

$$h(t; \theta) = \sum_{p=0}^{P-1} \alpha_p(\theta) e^{j\phi_p(\theta)} \delta(t-\tau_p(\theta))$$

with gradients analytically tractable via TensorFlow automatic differentiation.

Environmental shadowing, multipath, Doppler, and spatial/temporal correlation are computed from first-principles ray tracing, deterministically reflecting the 3D geometry and material assignments. Unlike statistical models, this approach ensures spatially consistent, temporally smooth channel realizations under arbitrary mobility (Zubow et al., 2024).

4. Main Simulation and Learning Workflows

Sionna supports a broad spectrum of workflows, from classical Monte Carlo link-level evaluation to advanced end-to-end machine learning–based air interface design. Representative capabilities include:

• Classical evaluation: Source $\rightarrow$ FEC $\rightarrow$ Mapper $\rightarrow$ (MIMO, OFDM) $\rightarrow$ Channel (Sionna RT or 3GPP) $\rightarrow$ Equalizer $\rightarrow$ Decoder, exposing all physical and coded-layer metrics.
• Neural optimization: Any component can be replaced by a custom Keras layer for, e.g., learned modulations, demapping, or receiver design; the framework natively supports backpropagation through the entire chain.
• Differentiable channel/environment design: Gradients can be efficiently backpropagated through Sionna RT, enabling automatic calibration of material properties, orientation of transmitter arrays, or optimization of reconfigurable intelligent surfaces (RIS) (Hoydis et al., 2023, Aoudia et al., 30 Apr 2025).
• PyJama integration: Enables end-to-end differentiable simulation and learning of adversarial jamming/anti-jamming strategies, including stochastic gradient descent–based optimization over the OFDM resource grid within a full link-level simulation (Ulbricht et al., 2024).

GPU acceleration ensures high throughput, with typical CIR computation (hundreds of rays, three bounces) requiring a few milliseconds per snapshot, and gradient evaluation incurring only modest additional overhead.

5. Integration with Network and Digital Twin Simulators

Sionna RT has been directly integrated into network-level simulators, notably ns-3 and the OpenAirInterface-based digital twin platforms, to bridge the gap between accurate physical channel modeling and system/network performance analysis (Zubow et al., 2024, Pegurri et al., 2024, Iye et al., 15 Mar 2025). The integration follows a split-process architecture:

• Sionna RT runs externally (Python/TensorFlow), maintaining 3D geometry and responding to channel requests over inter-process communication (ZeroMQ, UDP).
• ns-3 or OAI act as clients, providing current (and predicted) positions, frequencies, and simulation parameters, with Sionna RT providing per-link CIR, CFR, path loss, and delay.
• Mobility models employ physically accurate “block-bounce” updates based on geometric ray tracing.

Caching mechanisms (temporal coherence-based), batch parallelization, and speculative computation are employed to amortize the significant computational cost of deterministic ray tracing, achieving scalability to small/medium node counts and real-time or near-real-time operation in digital twin scenarios. Fine-grained CSI is made accessible to higher layers (PHY, MAC, application), enabling data-driven protocol optimization and network management.

6. Applications and Empirical Validation

Sionna and Sionna RT have enabled a wide range of applications and evaluation campaigns:

• Digital Twins: Slot-accurate, site-specific emulation of urban 5G mobility, with key performance indicators (RSRP, MCS, BLER, throughput) tracked in real time in OAI/O-RAN emulators (Iye et al., 15 Mar 2025).
• Network Stack Validation: Observed up to 65% disagreement in application-layer packet delivery (PRDR metric) between standard stochastic models and Sionna RT–backed deterministic channel models in high-mobility, urban, multi-RAT scenarios (Pegurri et al., 2024).
• Jamming/Anti-Jamming Research: With the PyJama extension, gradient-based learning yields jamming strategies that outperform uniform barrage jammers by up to 20 dB in BER-based power gain; learned jammers induce BLER on LDPC links from 7% to >90% in certain SNR regimes (Ulbricht et al., 2024).
• Channel-aware ML: Sionna provides the basis for environment-in-the-loop learning, facilitating research on physical layer beam management, channel prediction, and RIS design.

Hybrid stochastic/deterministic evaluation is supported by direct replacement, blending, or validation of 3GPP models with imported Sionna RT CIRs (Hoydis et al., 2022).

7. Limitations and Future Directions

Sionna's physical fidelity brings significant computational demand, particularly for large scenes or high mobility. Ns3Sionna and its variants mitigate this via coherence-based caching, batch parallelization, and GPU offloading, but remain best suited to systems with small-to-medium mobile populations or for offline/emulation studies (Zubow et al., 2024).

Current limitations include:

• SISO-only support in some network-level integrations (ns-3); ongoing work targets full MIMO and RIS abstraction.
• Mobility models are limited to specular, block-bounce; more realistic models (non-specular scattering, rough ground, detailed vehicles) remain under development.
• Scene/model size is constrained by (GPU/CPU) memory and per-batch computational requirements.
• Unmodeled phenomena: refraction is not yet included, nonstationary scatterers are under consideration.

Planned and ongoing directions include distributed multi-GPU execution, full-stack digital twin integration (PHY to application), RIS/channel data-driven optimization, and support for MIMO and spectrum coexistence in network simulation.

---

References:

• (Hoydis et al., 2022) Sionna: An Open-Source Library for Next-Generation Physical Layer Research
• (Hoydis et al., 2023) Sionna RT: Differentiable Ray Tracing for Radio Propagation Modeling
• (Ulbricht et al., 2024) PyJama: Differentiable Jamming and Anti-Jamming with NVIDIA Sionna
• (Zubow et al., 2024) Ns3 meets Sionna: Using Realistic Channels in Network Simulation
• (Pegurri et al., 2024) Toward Digital Network Twins: Integrating Sionna RT in ns-3 for 6G Multi-RAT Networks Simulations
• (Iye et al., 15 Mar 2025) Open Wireless Digital Twin: End-to-End 5G Mobility Emulation in O-RAN Framework
• (Aoudia et al., 30 Apr 2025) Sionna RT: Technical Report