Neural receivers: How a neural network replaces the classic 5G radio decoder

A neural receiver is a neural network that decodes the radio signal arriving at a cell tower. It takes over a job normally done by a chain of hand-designed math steps. The aim is to recover the bits a phone sent, with fewer errors or less computing.

Nvidia’s open 5G and 6G tools use the idea. SoftBank tested a neural receiver in Nvidia’s Aerial software, and ETH Zurich researchers tuned neural receivers on local measurements, as covered in our report on Nvidia’s open 6G tools.

Reading a smudged message at the other end of a noisy room

Picture a friend shouting a message across a crowded hall. Echoes blur the words, other people talk over them, and you only hear part of it. A radio receiver faces the same problem. A phone sends bits as radio waves. Those waves bounce off buildings, fade, and mix with signals from other phones.

The receiver must work out what was sent. A classic receiver does this in stages, like a team passing notes along a line. A neural receiver replaces several of those stages with one trained network.

Three hand-built steps that become one network

In the uplink (phone to tower) of 5G, the receiver typically runs three steps:

• Channel estimation: guess how the air distorted the signal, using known reference symbols called pilots.
• Equalization: undo that distortion, and separate users who transmit at the same time.
• Demapping: turn the cleaned-up symbols into the likelihood of each bit being 0 or 1.

Each step rests on a mathematical model with simplifying assumptions. A neural receiver feeds the raw received signal, plus the pilots, into one network. It outputs those bit likelihoods directly. The network is trained on many simulated transmissions, where the true bits are known, so it learns to reduce errors without being told the formulas.

The network still sits inside a standard chain. Its output goes to the usual error-correcting decoder, so the rest of the system stays unchanged.

From a research idea to a 5G-compatible design

Learning-based receivers were studied for years before 5G products. Nvidia researchers, including Sebastian Cammerer and Jakob Hoydis, published “A Neural Receiver for 5G NR Multi-user MIMO” with colleagues from Rohde & Schwarz in 2023 (arXiv:2312.02601, IEEE Globecom 2023). In this context, MIMO means multiple antennas serving several users at once, and NR means New Radio, the 5G air interface.

That receiver handles a whole 5G time-frequency slot at once. It uses convolutional layers to exploit how the channel changes smoothly across time and frequency. A graph neural network handles several users. According to the paper, it works with varying numbers of subcarriers, layers and users without retraining. It came within 1 dB of a strong classic baseline, with lower computational complexity, and was checked in hardware-in-the-loop tests with 3GPP conformance scenarios.

A follow-up in 2024 (Wiesmayr et al., arXiv:2409.02912) targeted real deployment. It cut inference time to under 1 millisecond on an Nvidia A100 GPU using TensorRT, Nvidia’s inference engine.

Training on the real site with a digital twin

A receiver trained for average conditions is a generalist. Nvidia’s researchers describe a two-stage approach. First, pre-train on randomized urban scenarios, with varied signal strength, speed and number of users. Then fine-tune on data from a ray-traced digital twin of the actual site. A digital twin here means a simulation of the real buildings, in which radio paths are traced.

In Nvidia’s September 2024 write-up, about one minute of fine-tuning on one GPU improved error performance by up to 2.2 dB over the pre-trained version. It also reported that a smaller fine-tuned network could match a universal model about four times larger. These are Nvidia’s own results from simulation and lab setups. The Leeds, SoftBank and ETH Zurich work in our news story follows the same logic of adapting to local conditions, with numbers that Nvidia reports and that were not accompanied by a paper in the announcement.

Where it falls short

Speed is the first limit. A cell tower must decode each slot within a fraction of a millisecond. The Wiesmayr paper reports that shrinking the network to meet the under 1 ms target cost less than 0.7 dB compared with a larger, non-real-time design. Nvidia’s authors also argue that counting operations or layers is a poor guide, so they measured latency on real hardware.

Generalization is the second. A network can overfit to the channels it saw in training. The 2023 paper stresses that careful training is what makes a receiver work on conditions it has never met. A receiver fine-tuned to one site may also need retraining when that site changes.

Third, most results come from simulation or controlled tests. Real networks add hardware imperfections and operator requirements for predictability. A black-box network is harder to analyze than a formula, which matters for standards and certification. Nvidia describes much of this work as research-stage.

A related idea goes further. Nvidia’s researchers also explored a pilot-free scheme with custom constellations, which could remove the overhead of reference symbols. Those constellations are not compliant with the 5G NR standard.

What to watch as 6G standards take shape

The open question is how much of the radio chain will be learned, and how much will stay hand-designed. Neural receivers are a low-risk entry point. They need no change to what phones transmit, and they run on the base station side, where GPUs are already being added.

Three things are worth following. One is whether operators publish field results on commercial networks, not only lab tests. Another is how well receivers hold up as sites change. The third is whether 6G standards allow designs, such as pilot-free signaling, that only a learned receiver can exploit. Open tools such as Sionna, for simulation and training, and Aerial, for running on real radio hardware, let any university lab test these claims without a vendor partnership.