This is the second of six (2/6) posts serializing the FAQ report from the most recent NVIDIA 6G Developer Day (sessions playback here). All questions and answers are directly from attendees on the day and the answers provided to them.
The Sections are:
1/6: Introducing NVIDIA 6G Research & Development Portfolio
2/6: Prototyping 6G with the NVIDIA Sionna Research Kit
3/6: Designing a Modular 6G System Using NVIDIA Aerial Framework
4/6: Aerial Framework – Performant GPU-Accelerated Code From Python
5/6: Optimizing 6G Performance With CUDA and NVIDIA DOCA
6/6: Simulating 6G with NVIDIA Aerial Omniverse Digital Twin
The full FAQ report is here:
2025_6GDD_NVIDIA AI Aerial_FAQ.pdf (374.9 KB)
Prototyping 6G with the NVIDIA Sionna Research Kit
Q 2.1: If we do have access to a pre-existing 5G RAN. Will we still need to use DGX Spark to utilize the cloud resources?
A: The Sionna Research Kit runs the full 5G RAN stack on a single DGX Spark. There is no need to have a pre-existing 5G RAN.
Q 2.2: What is the reference you use to assess the performance of your ray tracer?
A: The SionnaRT ray tracer. Its performance and accuracy have been validated by various academic publications; however, keep in mind that ray tracing remains an abstraction of the real-world.
Q 2.3: You mentioned x-apps – so is it OpenRAN compliant, e.g. we can simulate a number of splits?
A: The Sionna Research Kit is built on the OpenAirInterface 5G stack. Please see the OAI documentation for details on splits.
Q 2.4: What measurement framework did you use to report UE for the target cell RSRP, SINR, TA for performing inter-rat HO?
A: For the Tech-demo, the xApp framework is used to access UE reports such as the current MCS or BLER. This can be extended to other metrics as well.
Q 2.5: Can we use SionnaRK on other hardware? Also, is it possible to use the real-time channel simulation using simulated UEs and 5G Core?
A: SionnaRK is open-source and based on OpenAirInterface; in principle, it can be compiled on other platforms, however, channel emulation and some tutorials require CUDA. Yes, the OAI soft-UE is supported.
Q 2.6: How many UEs can be simulated in Sionna Research Kit with DGX Spark?
A: The channel emulator is currently limited to a single UE. Without channel emulation, multiple UEs can be connected.
Q 2.7: What RAN related function requires AI response in <1 msec? These are all built into protocols already.
A: The entire PHY layer must run in the sub-ms regime; this has implications on many PHY layer algorithms such as neural channel estimation or neural receivers.
Q 2.8: Regarding hardware, how about portable hardware to evaluate end-to-end MIMO setup with Sionna RK running on both UE and BS?
A: SionnaRK supports the NVIDIA DGX Spark and NVIDIA Jetson platforms; both are compact and portable.
Q 2.9: Are we able to work end-to-end on both UE and Base station with Sionna running on NVIDIA hardware?
A: SionnaRK can run an end-to-end 5G network on a single DGX Spark (including software-defined UE).
Q 2.10: Is it possible to run Aerial software on a general desktop computer? For example, in Linux? Or do we have to run it on NVIDIA hardware and there is no other option?
A: SionnaRK is open-source and based on OpenAirInterface; you can in principle compile it on other platforms, however, channel emulation and some tutorials require CUDA.
Q 2.11: How can we enable real time communication, given the huge computing requirements, and hence huge latency of AI models?
A: SionnaRK benefits from GPU acceleration and the unified memory of DGX Spark. This enables real-time AI-RAN inference. However, the neural network model must be also carefully optimized for low latency. See https://developer.nvidia.com/blog/real-time-neural-receivers-drive-ai-ran-innovation/ for further details.
Q 2.12: Is SionnaRK about democratizing research? How ready is this to start working to make solutions for RAN now?
A: SionnaRK and Aerial Testbed democratize 6G/AI-RAN research; for commercial solutions please see NVIDIA Aerial RAN Computer.
Q 2.13: How many USRPS and modems can DGX support at a time?
A: The channel emulator is currently limited to a single UE. Without channel emulation, multiple UEs can be connected.
Q 2.14: Is the inference latency an average or a maximum?
A: Both metrics (avg. and max.) are important and must be considered. The number in the 6G Developer Day presentation is the avg. latency.
Q 2.15: Sionna Research Kit website says, “The NVIDIA Jetson Thor as well as the NVIDIA DGX Spark will be supported soon.” Is there a specific time?
A: SionnaRK v1.1 now supports DGX Spark and Jetson Thor.
Q 2.16: What is the model size used in the 6G Developer Day presentation?
A: The neural receiver has approximately 140k trainable weights; see https://developer.nvidia.com/blog/real-time-neural-receivers-drive-ai-ran-innovation/ for further details.
Q 2.17: What are the main architectural challenges if we attempt to build SRK on discrete GPU servers (PCIe-based) rather than the recommended Unified Memory Jetson platforms? Is real-time performance still feasible?
A: This depends on the exact system configuration and how strict the real-time constraints are; one may need to reduce the effective bandwidth (i.e., number of PRBs).
Q 2.18: What GPU is used for nationwide simulation?
A: 5G coverage across the continental US was simulated in under five minutes by tracing more than 35 trillion rays on 96 NVIDIA L40S GPUs.
Q 2.19: Can the new NVIDIA updates also work with legacy RU or hardware platform in case operators wants to migrate to 6G from 5G or those hardware needs to be replaced with NVIDIA HW stack?
A: SionnaRK and Aerial Testbed democratize 6G/AI-RAN research; for commercial solutions please see NVIDIA Aerial RAN Computer.
Q 2.20: As most researchers in the physical layer used MATLAB for simulation, can I ask what the main advantage of Sionna compared with MATLAB toolboxes?
A: Sionna is fully differentiable, which allows for an easy integration with neural networks. It is Python-based, fully open source, and runs very efficiently on GPUs.
Q 2.21: Can you use any map for the channel emulation? And in which format it should be?
A: We can import any 3D scene; in our case we use OpenStreetMaps; see for details https://www.youtube.com/watch?v=7xHLDxUaQ7c
Q 2.22: What are the channel models used in each region of the US coverage calculation? Is it a data-driven model?
A: The 6G Developer Day presentation is a tech-demo to show what is possible with enough compute. The 3D environment comes from OpenStreetMaps and the cell tower positions are taken from OpenCellID.
Q 2.23: For the national coverage map, where is the terrain and clutter data coming from? Do you have this for every site today?
A: The 6G Developer Day presentation is a tech-demo to show what is possible with enough compute. The 3D environment comes from OpenStreetMaps and the cell tower positions are taken from OpenCellID.
Q 2.24: How about training NRX on data generated for different channel models e.g. UMi and UMa. How does that compare to site-specific NRX?
A: Please see Sec. IV in https://arxiv.org/pdf/2409.02912 for details
Q 2.25: Do you plan to implement diffraction of more than first order (diffraction happening even in NLOS cases)? Will Sionna also support frequencies below 1GHz?
A: SionnaRT supports frequencies below 1 GHz.
Q 2.26: Are you emulating the channel after you receive I/Q samples from USRP? You have mentioned 1-2 microseconds latency. Is it for the lowest PRB/bandwidth configuration?
A: The demo uses 24 and 51 PRBs (30kHz subcarrier-spacing).
Q 2.27: What is USRP?
A: USRPs are a software-defined radio from Ettus Research / National Instruments; see USRP Software Defined Radio (SDR) online catalog - Ettus Research | Ettus Research, a National Instruments Brand | The leader in Software Defined Radio (SDR)
Q 2.28: Is there a cloud-based version of Sionna, the ray tracers with GPU setup, and similar tools for researchers something like Google Colab?
A: No; the USRPs must be connected locally. Note that SionnaRT runs in GoogleColab.
Q 2.29: Is the real time tracing demo available in GitHub?
A: Not yet
Q 2.30: Can Sionna be integrated with macro radios?
A: Please see NVIDIA Aerial RAN Computer for commercial-grade systems.
Q 2.31: Does Sionna have 5G DL RX? I remember it had only UL RX, which prevents DL simulation
A: SionnaRK uses the OpenAirInterface 5G stack which supports DL and UL; however, SionnaPHY (i.e., the Python link-level simulator) only supports the 5G PUSCH.
Q 2.32: To connect with commercial 5G model we need implementation of all 5G layers (L2/L3). Does Sionna provide all the layers?
A: No, Sionna only implements a very small part of the standard. If you want to test your algorithms in a full 5G system, you can either rely on Sionna RK (based on the OpenAirInterface) or use Aerial Testbed.
Q 2.33: Where is the CORE part of the network running, in GPU or outside?
A: The 5G core network runs on the ARM CPU.
Q 2.34: Which are other open-source emulators it is compatible with? srsRAN?
A: SionnaRK is based on OpenAirInterface.
Q 2.35: What opensource CORE is being used here?
A: Sionna PHY/SYS are based on TensorFlow. Sionna RT is based on Dr.Jit and Mitsuba3. Sionna RK is based on the OpenAirInterface.
Q 2.36: Are xApps running on FlexRIC?
A: Yes, xApps run on FlexRIC.
Q 2.37: What is the limit on the number of hops in the ray tracing before a receiver is reached?
A: The max number of reflections can be configured in SionnaRT.
Q 2.38: Is there a support for terrestrial and satellite network integration, one of the main 6G features?
A: SionnaRK is based on OpenAirInterface; there are a few NTN features. However, SionnaRK has no dedicated NTN tutorials.
Q 2.39: Do we have any sample demo code which can be used?
A: All examples shown are open source; see Research Kit (RK) — Sionna 2.0.1 for details.
Q 2.40: Can we run Sionna code on any laptops having RTX GPU(s)? Are there any hardware requirements?
A: Yes, Sionna PHY/SYS/RT run on almost all modern GPUs (even on CPU only, as well as Apple Silicon). Sionna RK is only ensured to work on DGX Spark and Jetson.
Q 2.41: Do the limitations on OAI, like lack of Full E2 interface, also exist in Sionna research kit?
A: SionnaRK is based on OAI and has the same features as OAI.
Q 2.42: Can we also simulate in the situation if we use a Massive MIMO (64T/64R) in the network to evaluate the Network Performance?
A: Yes. However, it depends on the framework being referred to. In Sionna PHY/SYS/RT, the only limitation is GPU memory. In Sionna RK, it would be very challenging to evaluate a massive MIMO setup because (1) channel emulation becomes prohibitively complex and (2) OAI has currently no multiuser MIMO support.
Q 2.43: Does Sionna use Inline or Lookaside acceleration?
A: NVIDIA Aerial is based on inline acceleration.
Q 2.44: Is it possible to replace the DGX with a laptop?
A: Yes, Sionna RK also runs on other hardware platforms. However, there is no guarantee that it will work out-of-the-box.
Q 2.45: Is it possible to use SRK on other SDR devices brands?
A: Yes, but there is no guarantee that it will work out-of-the-box.
Q 2.46: Can you expand on Cloud Adoption?
A: There is currently no cloud version; keep in mind that the SDR must be physically connected.
Q 2.47: If we run raytracing for a local region, we can obtain signal coverage only within that region. However, many rays travel much farther than the simulated boundary, and antennas from neighboring regions also contribute signal into this area. So, splitting the map into smaller tiles becomes inaccurate, while using one extremely large map makes the computation infeasible. How can this problem be solved when the geographic area is too large for raytracing in a single model, but dividing it into smaller pieces is also unrealistic?
A: Ray tracing is physically grounded simulation; however, it remains a model/abstraction of the real-world. Keep in mind that rays with long paths and/or many reflections carry relatively little energy, i.e., contribute little to the CIR.
Q 2.48: To model the coverage area of the United States, a large 3D map would be required, which makes the signal-coverage data extremely heavy. How can such a vast coverage area be represented in practice?
A: The 6G Developer Day demo used tiles whose size depend on the base station density.
Q 2.49: How does message passing decoding for LDPC codes compare to neural decoding? Does the neural decoder achieve near capacity results with linear complexity?
A: The neural receiver replaces channel estimation/equalization and demapping. The LDPC decoder is a standard BP implementation with the same performance as in the baseline. See https://developer.nvidia.com/blog/real-time-neural-receivers-drive-ai-ran-innovation/ for details.
Q 2.50: How do we ensure the AI workloads seamlessly integrate AI RAN and Core?
A: Integrating AI-RAN algorithms in the 5G/6G stack is indeed a challenging task for researchers. Thus, SionnaRK provides several tutorials on how to integrate such algorithms; see Tutorials — Sionna 2.0.1 for details.
Q 2.51: Can you think of a particular scenario or setup where Sionna RK would not be the suitable tool?
A: SionnaRK democratizes 6G/AI-RAN research and targets academic and industrial researchers; if commercial deployment is targeted, please consider Aerial Testbed and AODT (for simulations).
Q 2.52: How can custom scenes be loaded into Sionna? What requirements do you have to consider for high fidelity scenes?
A: Any 3D scene can be imported; in this case, OpenStreetMaps was used; see for details https://www.youtube.com/watch?v=7xHLDxUaQ7c
Q 2.53: How important is Time Synchronization between RU and DU? Can I use COTS O-RU and focus on development for O-DU?
A: SionnaRK is based on the OpenAirInterface 5G stack which in principle supports O-RUs. However, O-RUs have not being tested. Please consider Aerial Testbed for such cases.
Q 2.54: Can we run the backend on consumer hardware (e.g. 4090 RTX)? What are the min VRAM requirements for real-time simulation?
A: Most of the tutorials run on consumer hardware (without unified memory). The requirements strongly depend on the 5G system configuration (e.g. number of PRBs).
Q 2.55: Is it possible to test propagation of different MHz waves, signal strength and material density?
A: In Sionna RT, the carrier frequency, material properties and transmit powers can be freely configured.
Q 2.56: What are the hardware requirements and capabilities for running Sionna, including support for custom scenes, User Equipment capacity, and compatibility with various GPUs and systems like NVIDIA Jetson and DGX Spark?
A: SionnaRK runs on DGX Spark and on the Jetson platform. The tech demo presented in the talk uses single UE configuration with 24-51 PRBs. For more capable systems please see Aerial Testbed Aerial Testbed — Aerial Testbed
Q 2.57: What are the main advantages of using Sionna over MATLAB toolboxes for physical layer simulations, and how are measurements planned to be performed in Sionna for specific events?
A: Sionna is a differentiable simulation toolbox which allows to replace arbitrary signal processing blocks in the transceiver signal processing pipeline, e.g., the channel estimator can be replaced by a trainable neural network, or an entire neural receiver can be trained end-to-end. See https://developer.nvidia.com/blog/real-time-neural-receivers-drive-ai-ran-innovation/
Q 2.58: Can Sionna support advanced diffraction, <1GHz frequencies, integration with various HW and macro radios, real-time channel simulation with 5G Core, full 5G layer implementation, 5G DL RX, and terrestrial-satellite network integration?
A: For such advanced 3GPP compliant end-to-end simulations, please consider Aerial Omniverse Digital Twin (AODT). Aerial Omniverse Digital Twin 1.4.0 — Aerial Omniverse Digital Twin
Q 2.59: What are the architectural challenges and performance implications of building SRK on discrete GPU servers compared to NVIDIA’s recommended Unified Memory Jetson platforms and which solution does NVIDIA prefer?
A: Real-time AI inferencing with sub-ms latency requires a carefully optimized software stack to avoid unnecessary memcopy overheads (e.g., using unified memory) and to reduce kernel launch overhead (e.g., via CUDA graphs).
Q 2.60: What are the compatibility options with other open-source emulators, different SDR device brands, and the potential for hardware offset emulation?
A: SionnaRK is open source and all components can be used as blueprints with other SDRs or software frameworks and all tutorials can be extended according to the user’s need. However, please understand that support may not be provided for such custom implementations.
Q 2.61: What open-source core is being used, where is it running (GPU or otherwise), and what GPU is utilized for the nation-wide simulation?
A: We simulated 5G coverage across the continental US in under five minutes by tracing more than 35 trillion rays on 96 NVIDIA L40S GPUs. The code is available as open source GitHub - NVlabs/sionna-large-radio-maps: Large-scale simulation of wireless network coverage maps using Sionna RT. · GitHub