NVIDIA Sionna PHY/SYS Simulation Library

• NVIDIA Sionna PHY/SYS is an open-source, GPU-accelerated simulation library that enables rapid prototyping and evaluation of physical-layer and system-level wireless algorithms with differentiable TensorFlow components.
• It offers modular Keras-based layers for core functions such as channel coding, MIMO, OFDM, ray tracing, and integrates neural network components for end-to-end optimization.
• The library supports both link-level performance metrics and large-scale system simulations, driving research in next-generation wireless (5G, 6G) and AI-native Open RAN applications.

NVIDIA Sionna PHY/SYS is an open-source, GPU-accelerated simulation library designed for rapid prototyping and rigorous evaluation of physical-layer (PHY) and system-level (SYS) communication algorithms. Built atop TensorFlow and Keras, Sionna exposes modular, differentiable components that span channel coding, MIMO, OFDM, modulation/demapping, and advanced channel models such as ray tracing. It is explicitly engineered for both link-level (bit error rates, block error rates) and large-scale, system-level simulation, allowing integration of neural networks and end-to-end learnable air interfaces. Key extensions such as Sionna RT enable physically-accurate, scene-driven ray tracing, while complementary libraries like PyJama enable differentiable jamming research. Sionna’s deep integration with GPU compute backends and availability of custom CUDA kernels make it a standard tool for next-generation wireless research, including 5G, 6G, and AI-native Open RAN evaluations.

1. Architectural Framework and Design Principles

Sionna implements a pure Python front-end leveraging TensorFlow 2.x and the Keras API; every PHY/SYS component is a subclassed Keras Layer, ensuring composability and full differentiability throughout the signal chain. Core subpackages include:

• sionna.fec: 5G LDPC, Polar, CRC, Reed–Muller, convolutional codes; decoders (belief propagation, SC/SCL, Viterbi).
• sionna.channel: AWGN, flat-fading, 3GPP TDL/CDL/UMa/UMi/RMa models, and interfaces for ray-traced impulse responses.
• sionna.ofdm: IFFT/FFT, cyclic prefix support, flexible frame and pilot patterns, LS estimators.
• sionna.mimo: MIMO primitives, ZF/MMSE/MRC precoding & equalization, antenna correlation models.
• sionna.rt: Ray tracing, 3D scene loading, multipath computation using exact geometry/material parameters.

Sionna is natively compatible with TensorFlow batch shapes ([batch_size, .., fft_size]) and exploits eager or graph mode execution, JIT compilation (XLA), mixed-precision, and multi-GPU scaling via MirroredStrategy. For non-TensorFlow-expressible kernels, Sionna dynamically loads custom C++/CUDA ops, maintaining automatic differentiation where feasible (Hoydis et al., 2022).

2. Physical Layer and System-level Capabilities

Channel Models: Sionna provides a hierarchy of channel models:

• AWGN ($y = x + n, \; n\sim \mathcal{CN}(0,N_0)$)
• Flat/MIMO fading ($y = Hx + n$)
• 3GPP TDL/CDL models (multipath sum with parametrized path gains $\alpha_{k\ell,p}$ and delays $\tau_{k\ell,p}$)
• Ray traced (Sionna RT): Loads 3D scenes, computes multi-path propagation directly from scene geometry/material via scene.compute_paths(), then transforms to CIR (Hoydis et al., 2022).

Modulation/Coding: M-QAM, soft demapping (bitwise LLR), 5G LDPC encoding/decoding (belief propagation), Polar and CRC codes, rate-matching, and block-interleaving facilities.

MIMO/OFDM Support: Keras layer abstractions exist for precoding (ZF, MMSE, MRC), multiuser scenarios using block diagonalization, OFDM resource grids (flexible FFT size, cyclic prefix, arbitrary pilot placement), and bit-level realistic error evaluation.

Neural/End-to-End Integration: Arbitrary chain segments—constellation, demapper, decoder, scheduler—may be replaced with neural layers, supporting gradient-based optimization and fully differentiable end-to-end learning (Hoydis et al., 2022).

System-level Simulation: Multiple parallel links can be instantiated, each with distinct channel realization and scheduling, facilitating Monte Carlo sweeps and multi-cell evaluations directly on multi-GPU clusters.

3. Sionna RT: Ray Tracing-based Channel Modeling

Sionna RT enables physically accurate channel modeling via 3D ray tracing, using Mitsuba-3 as the geometric optics engine. Key features:

• Multi-path Channel Realization: Computes all propagation paths (LOS, reflection, diffraction) between arbitrary Tx/Rx pairs in static or dynamic scenes.
• Impulse Response Modeling: Each snapshot (static within a channel coherence time $T_C$) is realized as

$$h(t) = \sum_{n=0}^{N-1} a_n e^{-j\phi_n} \delta(t - \tau_n)$$

with amplitude $a_n$, phase $\phi_n$, delay $\tau_n$, and per-path angles.

• OFDM Frequency Response:

$$H(f_k) = \sum_{n=0}^{N-1} a_n e^{-j\phi_n} e^{-j2\pi f_k\tau_n}$$

• GPU Vectorization: By batching up to hundreds of Rx positions, all rays are traced and responses aggregated in a single GPU kernel invocation. Mitsuba’s BVH acceleration structures ensure that triangle intersection and path accumulation are efficient, yielding nearly flat execution time for $R \lesssim 500$ receivers per batch on a Tesla V100 (Zubow et al., 2024).
• Integration with Network Simulators: Ns3Sionna demonstrates bi-process coupling between ns-3 and Sionna RT (ZeroMQ+protobufs), allowing ns-3 to obtain spatially and temporally realistic path loss, delay, and CFRs at packet or mobility event granularity.

4. Advanced PHY Applications and Extensions

High-performance LDPC and Link-Level Benchmarks: Sionna’s built-in LDPC5G encoder/decoder supports realistic 5G NR chains. Benchmarks on DGX Spark platforms show 6× speedup for LDPC decoding (k=512, n=1024) on Blackwell GB10 GPU versus Grace CPU, with per-codeword latency as low as 0.03 ms (I=5 iterations), always remaining inside the 0.5 ms NR slot for reasonable iteration counts. Full-batch CPU utilization for Sionna-based LDPC decoding peaks at ≈17 Grace cores, while GPU decoding only incurs ≈10–15 W incremental load, enabling resource reallocation to higher layers (Barker et al., 4 Feb 2026).

Differentiable Jamming/Anti-Jamming: PyJama extends Sionna via Time/FrequencyDomainJammer, POS/IanLMMSE anti-jamming receivers, and differentiable interference models. Complete OFDM + FEC + MIMO stacks can be constructed; jamming power allocation $\rho$ is trained via SGD to maximize BER/BLER under a given power budget, with backpropagation traversing all Sionna layers. Learned jamming strategies severely outperform uniform jammers, demonstrating up to 20 dB power-equivalent gains and full system denial for coded links (Ulbricht et al., 2024).

5. Optimization, Performance, and Caching Mechanisms

• Caching and Coherence: To amortize the high cost of ray tracing, Sionna implements coherence-time caching. Channels are computed and retained within their $T_C$, reciprocally reused; when cache-misses occur, requests for current and near-future Rx positions are batched to maximize parallel GPU throughput.
• Distance Pruning: A Friis-based power check skips ray tracing for channel pairs under the receiver noise threshold.
• Predictive Batching: At cache-miss points, small “M” time horizons of future positions are submitted in a single GPU batch, balancing memory pressure and spatial/temporal correlation fidelity.
• Trade-offs: Large $M$ boosts GPU efficiency, stressing memory; short $T_C$ (high mobility, high carrier freq.) reduces cache lifetime and increases batch frequency (Zubow et al., 2024).

Component	GPU Throughput*	CPU Utilization*	Power Delta*
LDPC5G Decoder (GB10)	≤0.12 ms/cw @ I=20	≈10–17 Grace cores	+10–15W (GPU)
Ray tracing (V100)	Flat <500 Rx/batch	Linear in #STAs	>80% GPU usage

*Values as reported in (Zubow et al., 2024) and (Barker et al., 4 Feb 2026).

6. Validation, Benchmarks, and Research Impact

• ** Fidelity:** Sionna RT (and Ns3Sionna) recovers free-space path loss and delay (vs. ns-3 Friis) to <0.01 dB, <1 ns. Physically accurate realizations in complex indoor/outdoor scenes capture frequency selectivity, shadowing, and spatial correlation absent in analytic models.
• ** Benchmark Scaling:** Sionna achieves $10^5$ LDPC blocks/s for short codes; complete FEC+MIMO+OFDM chains scale linearly to $10^6$ trials on multi-GPU workstations. Compared to Vienna 5G LLS or AFF3CT, GPU-accelerated Sionna yields 5–20× throughput at reference accuracy (Hoydis et al., 2022).
• ** Extensibility:** The Sionna/PyJama ecoystem is fully open-source (Apache 2.0), supports custom TensorFlow Layers and CUDA ops, and provides standardized benchmarking (BER/BLER/EXIT chart routines). Planned directions exist in THz, semantic/JSCC, and integrated sensing domains.

7. Summary and Significance

NVIDIA Sionna PHY/SYS delivers a unified software architecture for reproducible, high-performance physical-layer and system-level research. Its modular, differentiable structure supports both classical and learning-based algorithm development, full-stack end-to-end simulation, and rigorous physical realism through physics-driven channel models and GPU-accelerated computation. Sionna and its extensions (RT, PyJama, DGX benchmarks) have become foundational in domains spanning large-scale MIMO, OFDM, Open RAN benchmarking, and adversarial wireless security, with empirical evidence of both performance and research flexibility (Hoydis et al., 2022, Zubow et al., 2024, Barker et al., 4 Feb 2026, Ulbricht et al., 2024).