Sionna: GPU-Accelerated Wireless Research

• Sionna is an open-source, GPU-accelerated library that supports end-to-end differentiable wireless simulation combining link-level modules, neural integration, and ray tracing.
• The library integrates coding, modulation, OFDM, MIMO, and learning-based receivers with a differentiable ray tracer for scene-consistent channel modeling.
• Sionna facilitates design and optimization in 5G NR, digital twin creation, and hardware-in-the-loop deployment through gradient-based methods.

Sionna is an open-source, GPU-accelerated library for wireless physical-layer research that combines link-level simulation, differentiable signal-processing blocks, and—since release $v0.14$—a differentiable ray tracer for radio propagation modeling. In the literature, the name denotes both the core simulation framework for coding, modulation, OFDM, MIMO, and learning-based receivers, and the Sionna RT subsystem for scene-based channel impulse responses and radio maps. Across recent work, Sionna serves as a common substrate for end-to-end learning, standards-aligned 5G NR experimentation, digital-twin construction, and hardware-in-the-loop deployment (Hoydis et al., 2022, Hoydis et al., 2023, Aoudia et al., 30 Apr 2025).

1. Origins, scope, and evolution

Sionna was introduced as a GPU-accelerated, TensorFlow-based, open-source library for link-level simulation of digital communication systems, with a particular emphasis on next-generation and 6G research (Hoydis et al., 2022). Its initial scope already included a broad collection of physical-layer algorithms, a modular signal-processing chain built from Keras layers, native neural-network integration, and a design explicitly oriented toward rapid prototyping, reproducibility, extensibility, and batch-parallel execution on GPUs. The same account places Sionna in a lineage distinct from traditional link-level simulators by emphasizing end-to-end differentiability, native neural-network support, GPU-centric tensor processing, and integration with physically based channel modeling (Hoydis et al., 2022).

A major expansion occurred when Sionna incorporated Sionna RT, described as a differentiable ray tracer for radio propagation modeling. In that phase, the framework was presented not merely as a forward simulator of channel realizations, but as a system in which channel impulse responses, coverage maps, and related quantities can be differentiated with respect to material properties, antenna patterns, array geometries, and transmitter or receiver poses (Hoydis et al., 2023). This extended Sionna from a link-level library with stochastic channel models into a scene-aware platform for geometry-consistent wireless simulation.

The later technical report on Sionna RT describes the Sionna 1.0 release as a complete overhaul of the ray tracer, with substantial improvements in speed, memory efficiency, and extensibility, and with solver designs specialized for channel impulse responses and radio maps (Aoudia et al., 30 Apr 2025). That report also clarifies that Sionna had become interoperable with major deep-learning frameworks while retaining differentiability through Dr.Jit and Mitsuba 3 in the ray-tracing stack, which suggests a broader role as a general differentiable wireless-simulation infrastructure rather than a narrowly TensorFlow-bound toolkit (Aoudia et al., 30 Apr 2025).

2. Core software model and differentiable physical-layer simulation

At the core of Sionna is a tensor-centric architecture in which major communication-system functions are represented as composable Keras layers. The library exposes modules such as BinarySource, LDPC5GEncoder, Mapper, AWGN, Demapper, and LDPC5GDecoder, and it is designed so that user-defined neural layers can replace analytical blocks without breaking the computational graph (Hoydis et al., 2022). This architecture is coupled to automatic differentiation, making the complete chain amenable to gradient-based optimization.

The generic baseband model used throughout the ecosystem is the familiar linear form

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n},$

with $\mathbf{x}$ the transmit vector, $\mathbf{H}$ the possibly time- and frequency-selective channel matrix, and $\mathbf{n}$ additive noise (Hoydis et al., 2022). On top of this abstraction, Sionna provides OFDM modulation and demodulation, cyclic-prefix insertion and removal, arbitrary pilot patterns, least-squares channel estimation, MIMO processing, and forward-error-correction modules including 5G LDPC and 5G Polar coding with rate matching (Hoydis et al., 2022). For MIMO equalization, the library includes standard linear receivers such as MMSE equalization, exemplified by

$$\hat{\mathbf{x}} = (\mathbf{H}^\mathrm{H}\mathbf{H} + \sigma_n^2\mathbf{I})^{-1}\mathbf{H}^\mathrm{H}\mathbf{y}.$$

Sionna’s differentiability is central to the way later papers use it. A representative example is the training of a 5G NR-standard-compliant neural receiver in which Sionna supplies differentiable modulation, OFDM, LDPC, resource-grid abstractions, DMRS-based channel estimation, and channel models such as TDL or CDL; the neural receiver then replaces or augments equalization and demapping while retaining compatibility with conventional OFDM demodulation and LDPC decoding (Cammerer et al., 19 May 2025). The same differentiable chain enables end-to-end optimization of custom decoders. In transformer-based LDPC decoding, Sionna provides the 5G NR LDPC code definitions, rate matching, AWGN channel, QAM mapping and demapping, and reference BP decoders, while custom transformer layers are inserted between demapper output and bit decisions and trained by backpropagation through the entire chain (Hernandez et al., 23 Jan 2025).

This software model also supports custom CUDA-backed operations when tensor composition alone is inefficient. The original Sionna paper cites custom C++ and CUDA ops for algorithms such as Polar SCL decoding, some 3GPP channel routines, min-sum BP decoding, and time-varying convolution, while preserving the tensor interface expected by the rest of the framework (Hoydis et al., 2022). A plausible implication is that Sionna’s architecture deliberately balances high-level modularity with low-level acceleration for operations that are structurally awkward in pure Python or TensorFlow.

3. Sionna RT: scene representation, path solvers, and radio maps

Sionna RT extends the framework from stochastic or standards-based channel models to geometry-aware propagation over explicit 3D scenes. In early descriptions, scenes are triangle meshes managed through Mitsuba 3, optionally created in Blender and enriched from OpenStreetMap via Blender-OSM, with materials defined by relative permittivity, conductivity, and scattering parameters (Hoydis et al., 2023). Transmitters and receivers are objects with positions and orientations, while arrays are specified through abstractions such as PlanarArray, supporting shared transmitter and receiver array configurations as well as synthetic-array approximations for computational acceleration (Hoydis et al., 2023).

The path-based channel model is expressed as a superposition of multipath components. For a path $p$ with complex gain $a_p(f)$ and delay $\tau_p = d_p/c$, the contribution to the impulse response is approximately

$h_p(t;f)=a_p(f)\,\delta(t-\tau_p),$

and the total response is

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n},$0

In OFDM form,

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n},$1

These quantities are available as differentiable outputs through APIs such as paths.cir(...), and Doppler can be injected through apply_doppler(...) (Hoydis et al., 2023).

The later technical report provides a more explicit algorithmic account. For channel impulse responses, Sionna RT combines shooting and bouncing of rays with the image method and a hashing-based mechanism to eliminate duplicate paths efficiently (Aoudia et al., 30 Apr 2025). For radio maps, it uses a purely SBR-based approach. The report formalizes the channel as

$\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{n},$2

and defines the channel gain as a noncoherent sum over path magnitudes (Aoudia et al., 30 Apr 2025). It also details material interaction models based on Fresnel coefficients for reflection and refraction, slab thickness, scattering coefficients, and polarization-dependent transformations, all implemented in a differentiable manner through Dr.Jit and Mitsuba 3 (Aoudia et al., 30 Apr 2025).

An important version-specific nuance concerns supported interactions. The 2023 Sionna RT paper described the then-current model as including specular and diffuse reflections and first-order diffraction, with refraction and RIS support planned (Hoydis et al., 2023). The 2025 technical report, by contrast, states that version 1.0.2 supports specular reflection, diffuse reflection, and refraction, while diffraction and RIS support are not in 1.0.2 and are planned to return (Aoudia et al., 30 Apr 2025). This version history is significant because claims about Sionna RT’s physical fidelity depend on the release under discussion.

4. Learning, optimization, and model-based design

Because Sionna couples differentiable wireless blocks with automatic differentiation, it is used not only to simulate systems but also to optimize them. In the original formulation, trainable constellations and neural demappers were already natural use cases, with losses defined directly on coded bits or decoder outputs (Hoydis et al., 2022). Later work generalizes this pattern to more specialized design problems.

In 5G NR receiver design, Sionna is used to train a neural receiver end to end on realistic channel realizations and noise, after which the trained TensorFlow model is exported and compiled into a TensorRT engine for deployment in a live 5G NR cell (Cammerer et al., 19 May 2025). In channel decoding, it provides the standard-compliant LDPC chain and BP baselines against which transformer and linear-time transformer decoders are trained and benchmarked, with syndrome-based inputs and bit-wise cross-entropy objectives implemented as differentiable layers inside Sionna’s PHY graph (Hernandez et al., 23 Jan 2025).

Sionna’s differentiable ray tracing also enables direct optimization of environment-aware parameters. In physically consistent RIS design, Sionna RT supplies differentiable path gains and coverage maps as functions of reradiation mode amplitudes, and these amplitudes are optimized with Adam inside TensorFlow for fairness and minimum-coverage objectives over target areas (Shabanpour et al., 2024). In airborne base-station deployment, Sionna RT and TensorFlow are used for back-propagation-based optimization of ABS location, antenna azimuth and tilt, and transmit power through losses defined on coverage and SIR maps, including smooth minimum and AOI-weighted objectives (Belgiovine et al., 17 Aug 2025). In resource allocation, Sionna and Sionna RT are embedded in a closed-loop system-level environment where hierarchical cooperative MARL agents learn joint PRB and power allocation under ray-traced, mobility-aware channels, with adaptive modulation and coding, HARQ, and fairness-aware rewards (Dorcheh et al., 4 May 2026).

The framework has also been extended into adversarial settings. PyJama adds differentiable jamming and anti-jamming blocks to Sionna, including time-domain and frequency-domain jammers, pilot patterns with silence for jammer CSI estimation, and anti-jamming receivers such as POS and interference-aware LMMSE equalization. Within that combined graph, jammer power allocation over an OFDM resource grid is optimized by stochastic gradient descent (Ulbricht et al., 2024). This suggests that Sionna’s differentiability is not limited to improving receivers or codes; it can also be used to study adversarial optimization in realistic PHY stacks.

5. From simulation to testbeds, emulators, and digital twins

A recurring theme in the literature is the use of Sionna as a bridge between simulation and operational systems. The Sionna Research Kit exemplifies this pattern. It combines Sionna-based training with a Jetson AGX Orin, TensorRT, OpenAirInterface, a USRP B210, and a commercial Quectel RM520N-GL 5G NR modem, thereby moving a Sionna-trained, 5G NR-standard-compliant neural receiver into the real-time gNB PHY path of a live standards-compliant cellular link (Cammerer et al., 19 May 2025). In that workflow, Sionna acts as the design and training layer, while the Research Kit, OAI, and RF hardware provide deployment and real-time execution.

Another line of work embeds Sionna RT into full-stack network simulation. One approach integrates ns-3 and Sionna RT to build an open-source digital network twin for multi-RAT 6G simulation, with ns-3 and Sionna RT as separate processes communicating over UDP and with the ray tracer supplying deterministic, geometry-based channel parameters to replace stochastic channel models (Pegurri et al., 2024). A closely related framework, Ns3Sionna, uses ZeroMQ and Protocol Buffers to connect ns-3 to a Sionna RT server, adds a mobility model based on ray tracing, and exploits coherence-time-based pre-caching and point-to-multipoint channel calculations to make realistic channels tractable for small to medium numbers of nodes (Zubow et al., 2024). Both accounts stress that deterministic ray tracing materially changes higher-layer behavior relative to stochastic models.

Sionna RT has also been used to drive channel emulation and systematic hardware testing. In a 5G AoA framework, Sionna RT generates time-varying, multi-antenna channel impulse responses for urban-canyon trajectories; a custom converter transforms these into Keysight PROPSIM .asc files; and the resulting emulator-driven RF paths feed an OAI-based 5G stack with a USRP N310 array and SRS-based AoA estimators such as MUSIC and ESPRIT (Ceresoli et al., 20 Oct 2025). Here Sionna RT functions as a controllable, repeatable virtual environment for real hardware rather than as a runtime inference engine.

Digital-twin usage extends beyond transport stacks and PHY validation. Sionna RT is used in a production-network case study for London to produce site-specific large-scale channel gains from real geometry, which then feed Bayesian optimization of antenna tilts and half-power beamwidths (Benzaghta et al., 1 Apr 2025). It is also coupled with ns-3 and vehicular mobility to generate cross-layer datasets in urban network-digital-twin frameworks (Stenhammar et al., 14 Apr 2026). A plausible implication is that Sionna’s role in contemporary wireless research increasingly includes calibrated, site-specific digital twinning rather than only algorithm prototyping.

6. Application domains, misconceptions, and open issues

The breadth of Sionna’s applications is unusually wide. It appears in neural-receiver deployment for AI-RAN (Cammerer et al., 19 May 2025), transformer-based LDPC decoding (Hernandez et al., 23 Jan 2025), differentiable jamming and anti-jamming (Ulbricht et al., 2024), RIS optimization and physically consistent reradiation design (Shabanpour et al., 2024), large-scale cellular antenna optimization (Benzaghta et al., 1 Apr 2025), network-digital-twin construction (Pegurri et al., 2024), site-specific QoS-dataset generation (Stenhammar et al., 14 Apr 2026), and large-scale interference and direction-finding dataset creation (Wielenberg et al., 3 Jul 2026). This range supports a correction to a common misconception: Sionna is not only a link-level Monte Carlo simulator, and Sionna RT is not only a visualization tool. The published work treats the framework as a differentiable design environment, a scene-based channel engine, a training substrate, and, in some settings, a bridge to emulation and deployment.

At the same time, the literature is explicit about limitations. Ray tracing remains computationally demanding, especially for large scenes, high reflection depth, or dynamic multi-node settings (Zubow et al., 2024, Stenhammar et al., 14 Apr 2026). Version-dependent support for diffraction, refraction, and RIS features complicates the interpretation of results across releases (Hoydis et al., 2023, Aoudia et al., 30 Apr 2025). Several works emphasize the engineering complexity of integrating Sionna-derived models into live or full-stack systems, including TensorRT pipelines, memory management, channel conversion into emulator formats, and synchronization with higher-layer simulators (Cammerer et al., 19 May 2025, Ceresoli et al., 20 Oct 2025).

The most substantive controversy surfaced in recent RIS work concerns practical viability rather than simulation methodology. A calibrated digital twin powered by Sionna ray tracing was used to evaluate automated RIS deployment across 4G, 5G, and hypothetical 6G frequencies, and the study concluded that meaningful coverage enhancement requires dense, large-aperture RIS deployments, raising questions about the practicality and cost of large-scale RIS adoption (Beyraghi et al., 10 Oct 2025). This does not invalidate Sionna-based RIS studies; rather, it shows that site-specific, calibrated use of Sionna can sharpen negative as well as positive conclusions.

Taken together, the literature presents Sionna as a research platform whose defining property is not a single algorithmic feature but the coupling of differentiable wireless-system modeling with physically grounded propagation. The framework’s significance lies in that coupling: communication chains, propagation environments, and optimization objectives are expressed in one computational graph, permitting a continuum from analytical baselines to learned components, from synthetic channels to digital twins, and from simulation to deployed 5G NR prototypes (Hoydis et al., 2022, Aoudia et al., 30 Apr 2025).