Sionna RT: Differentiable GPU Ray Tracing

• Sionna RT is a fully differentiable, GPU-accelerated ray tracing framework for radio wave propagation, integrating deterministic electromagnetic modeling with auto-diff capabilities.
• It employs a hybrid shooting-and-bouncing rays method combined with image-based techniques to accurately compute channel impulse responses and path-specific parameters.
• Leveraging TensorFlow/PyTorch integration and on-GPU batching, Sionna RT enables efficient, site-specific wireless simulation and gradient-based optimization of system parameters.

Sionna RT is a fully differentiable, GPU-accelerated ray tracing (RT) framework for accurate and efficient modeling of radio wave propagation in complex environments. It is a core module within the open-source Sionna library for physical-layer research. Sionna RT integrates deterministic electromagnetic ray tracing with modern auto-differentiation and hardware acceleration, enabling site-specific, physically consistent wireless channel modeling and seamless gradient-based optimization of scene, array, and protocol parameters. The framework is deeply integrated with TensorFlow/PyTorch, Dr.Jit, and the Mitsuba 3 renderer, and is designed for both link-level physical simulations and system-level digital twin applications relevant to 6G wireless networks and beyond (Hoydis et al., 2023, Aoudia et al., 30 Apr 2025, Zubow et al., 2024, Pegurri et al., 2024).

1. Architecture and Ray-Tracing Algorithms

Sionna RT is composed of several layers that mediate between high-level scene and device definitions and low-level, parallelized electromagnetic field computation:

• Scene Modeling: Users describe 3D environments using XML or glTF, referencing OBJ/mesh files for geometry and assigning per-object radio materials (parameterized by relative permittivity $\epsilon_r$, conductivity $\sigma$, etc.).
• Device and Array Configuration: Transmitters and receivers are characterized by array geometries, orientation, and potentially full dual-polarized antenna patterns.
• Ray-Tracing Engines: Two solver types are provided:
  • PathSolver (CIR/streaming paths): Hybrid shooting-and-bouncing rays (SBR) plus the image method is used to compute propagation paths up to a user-specified reflection depth $L$. For each valid path, the framework accumulates all physical interactions (specular, diffuse, transmission) and computes total delay $\tau_n$, complex gain $a_n$, and angles of arrival and departure.
  • RadioMapSolver (power/coverage maps): Pure SBR-based Monte Carlo estimation projects power onto 2D/3D grids, efficiently accumulating partial fields for each measurement cell.

GPU acceleration is achieved through Dr.Jit's vectorized ray–scene intersection, on-GPU batching of all per-ray interactions, and minimal host-device synchronization. In Sionna RT version 1.0, all acceleration structures, path buffers, and hash tables for duplicate-path elimination reside in GPU memory, delivering significant reductions in runtime and resource consumption compared to prior releases (Aoudia et al., 30 Apr 2025).

The ray-tracing output for each transmitter–receiver pair is a set of $L$ discrete paths, each annotated as follows:

| Path Index $\ell$ | Delay $\tau_\ell$ | Complex Amplitude $a_\ell e^{-j\phi_\ell}$ | AoD $\theta^{\text{TX}}_\ell$ / AoA $\theta^{\text{RX}}_\ell$ |

The channel impulse response materializes as

$$h(t) = \sum_{\ell=0}^{L-1} a_\ell e^{-j\phi_\ell} \delta(t - \tau_\ell)$$

and frequency response for $N_{sc}$ OFDM subcarriers $k$ as

$$H[k] = \sum_{\ell=0}^{L-1} a_\ell e^{-j\phi_\ell} \exp(-j2\pi k \Delta f \tau_\ell)$$

(Zubow et al., 2024, Aoudia et al., 30 Apr 2025).

2. Differentiability and Auto-Diff Integration

A distinctive feature of Sionna RT is full end-to-end differentiability with respect to practically all system and environmental parameters (Hoydis et al., 2023, Aoudia et al., 30 Apr 2025). All geometric transformations, field computations, and material interactions are expressed as Dr.Jit and TensorFlow/PyTorch primitives, permitting direct back-propagation of gradients via standard auto-diff mechanisms.

Gradients can be computed for:

• Material permittivities and conductivities ($\epsilon_r$, $\sigma$)
• Antenna patterns and array geometries
• Device positions and orientations
• RIS (reconfigurable intelligent surface) parameters and programmable phase shifts

This enables learning of unknown physical parameters (e.g., from measurement data), end-to-end neural system optimization, and integrated waveform/runtime environment adaptation.

The gradient pipeline extends through CIR and coverage map computations:

$$\frac{\partial h}{\partial \theta} = \sum_{r}[ (\frac{\partial A_r}{\partial \theta})e^{-j2\pi f_c \tau_r} + A_r (-j2\pi f_c \frac{\partial \tau_r}{\partial \theta}) e^{-j2\pi f_c \tau_r}] \delta(t-\tau_r) + \ldots$$

where $\theta$ is an arbitrary differentiable parameter (Hoydis et al., 2023, Aoudia et al., 30 Apr 2025). No manual derivation is required as the computation graph is automatically constructed during forward execution.

3. Computational Efficiency and Hardware Acceleration

Sionna RT exploits extensive parallelization and GPU vectorization:

• Shooting and Bouncing Rays: Massive ray sample generation on a Fibonacci lattice ensures angular uniformity for both direct-path and higher-order multipath components.
• Hybrid SBR–Image Method: SBR proposes candidate paths, after which specular chains are verified and completed by analytic image construction and intersection replay.
• Hashing-Based Duplicate Elimination: Rolling hash tables on GPU eliminate repeated path candidates at negligible additional cost.
• Batch Processing: All Tx–Rx pairs, all rays, and all Rx or measurement grid points are processed in parallel.

Key performance metrics (RTx 4090, v1.0) include sub-second CIR computation (max depth 5, $10^6$ samples, $\sim 0.4$ s, $\sim 0.8$ GB memory) and 5–10× runtime/memory improvement over prior versions (Aoudia et al., 30 Apr 2025).

Pseudocode for the SBR + Image Method is provided for reference in the technical report (Aoudia et al., 30 Apr 2025). As confirmed by ns-3 integration studies, this productivity translates to rapid scene updates and support for dynamic environments once appropriate caching, TTL, and predictive precomputation strategies are employed (Zubow et al., 2024, Pegurri et al., 2024).

4. Integration with ns-3 and Digital Twin Networks

Sionna RT is directly integrated with the ns-3 network simulator through message-based, disaggregated architectures (Zubow et al., 2024, Pegurri et al., 2024):

• Data Exchange: ns-3 sends transmission events, scene and mobility updates, and channel queries via serial IPC (ZeroMQ/Protocol Buffers or UDP). Sionna RT responds with full channel descriptions for the requisite Tx–Rx pairs.
• Propagation Models: Sionna RT-derived CIR/CFR data populate custom PropagationLossModel and PropagationDelayModel instances in ns-3. The path loss and delay are returned as the RMS of frequency response and minimal delay over available paths.
• Mobility Modeling: Sionna RT enforces 3D scene boundaries and specular “bounce” reflections for mobile node trajectory updates, replacing the standard 2D random walk and enabling intrinsically physical mobility dynamics.
• Caching and Coherence-Time Optimization: Channel state is cached and re-used for the duration of the channel coherence time $T_C = \frac{9}{16\pi}\frac{c}{vf_c}$, and speculative prefetching is employed to amortize per-scene intersection overhead.

Quantitative application-layer performance comparisons in vehicular urban digital twins reveal substantial discrepancies (up to 65% difference in delivered messages) between ray-traced and conventional stochastic models, emphasizing the importance of site-specific, physically consistent channel representations for 6G networking research (Pegurri et al., 2024).

5. Applications: Advanced Channel Modeling, ML, and RIS

Channel Output and Derived Metrics

Sionna RT provides granular channel state information:

• Path delays $\{\tau_\ell\}$, path gains $\{a_\ell e^{-j\phi_\ell}\}$, AoA/AoD, and LoS/NLoS flags for each link.
• Derived channel metrics such as RMS delay and angular spread.

This output supports:

• ML-based channel estimation and beamforming relying on spatial consistency and fine-structure multipath (Zubow et al., 2024).
• Sensing and radar applications, leveraging multipath delays/angles for environmental awareness or change detection.
• Protocol design and evaluation under realistic fading, shadowing, and non-stationary propagation, beyond the capabilities of parametric stochastic models.

Differentiable ML Pipelines and Jamming Studies

Sionna RT is fully compatible with differentiable simulation workflows such as PyJama, enabling stochastic gradient descent-based learning of jamming and anti-jamming strategies. Trainable parameters (e.g., jamming power allocations) are optimized directly through the entire link chain, inclusive of the RT channel, in a hardware-accelerated TF graph (Ulbricht et al., 2024).

Reconfigurable Intelligent Surfaces (RIS) and Smart Environments

Ray-tracing-based channel estimation is crucial for evaluating RIS control algorithms in complex digital twin urban scenarios. Sionna RT supports integration with various RIS optimization strategies based on channel estimation, Hadamard LS, compressive sensing (OMP), and iterative coordinate-search (Güneşer et al., 10 Jan 2025).

Key findings from urban digital twin studies include:

RIS Configuration	Center Path Gain	RIS-Enhanced Gain	Computation Time (GPU T4)
RIS-off	$\approx -90$ dB	N/A	N/A
RIS-on (Hadamard/OMP/Iter)	$\approx -60$ dB	$\approx$30 dB gain	0.019s/1.219s/25.861s

Validation against SDR/antenna hardware platforms shows that real-world gains are lower due to hardware mismatches but support the use of digital twins for pre-deployment validation (Güneşer et al., 10 Jan 2025).

6. Limitations and Future Directions

Sionna RT version 1.0 introduces extensive optimizations but still has several active limitations:

• Diffraction effects (UTD) are not implemented in v1.0 but planned (v1.1); only reflection/transmission and scattering are modeled (Aoudia et al., 30 Apr 2025).
• RIS and mobility extensions via point scatterer updates are not yet fully realized in the new solver.
• Path candidate generation is constrained for extremely long specular chains unless sampling budgets are significantly increased.
• Caching/duplicate elimination is probabilistic and may discard rare chains under severe hash collisions, requiring careful tuning of table size and sampling density.

Active development targets support for fully differentiable UTD models, scalability through hybrid neural–ray surrogates, and dynamic/time-variant scene modeling.

7. Summary Table: Key Technical Differentiators

Feature	Sionna RT Capability	Citation
Path-level CIR, CFR, AoA/AoD	Yes (multi-path, full EM modeling)	(Aoudia et al., 30 Apr 2025, Zubow et al., 2024)
End-to-end Differentiability	Full auto-diff through Dr.Jit, TF, PyTorch	(Hoydis et al., 2023, Aoudia et al., 30 Apr 2025)
GPU-native Ray Launch/Batching	Yes, highly parallelized across scene/grid	(Aoudia et al., 30 Apr 2025)
ns-3/Digital Twin Integration	Disaggregated IPC, full-stack DNT support	(Zubow et al., 2024, Pegurri et al., 2024)
ML/Joint Optimization Support	Differentiable pipeline for ML and design	(Ulbricht et al., 2024, Hoydis et al., 2023)
RIS/Smart Surface Evaluation	Native and extended via custom solvers	(Güneşer et al., 10 Jan 2025)

Sionna RT establishes a reproducible, modular foundation for modern, physically accurate wireless simulation, capable of both forward prediction and inverse optimization. Its integration into the Sionna ecosystem brings differentiable, GPU-accelerated propagation modeling to the broader research and engineering community.