Description de l'entreprise
Free Mobile (Iliad group) designs and deploys its 5G network infrastructure, from the O-RAN radio unit and distributed unit (DU) to the user plane.
The DU physical layer (L1) is where AI and GPU computing now make the difference. Neural receivers, learned channel estimation and GPU-accelerated signal processing must all run within a 0.5 ms slot.
This internship sits at the intersection of: real 5G radio signals, hard real-time constraints, and AI models that must beat classical algorithms on measurable metrics such as BLER, throughput and latency.
Description du poste
INTERNSHIP SUBJECT
You will design, implement and benchmark AI-assisted physical-layer algorithms for an O-RAN 7.2x DU, running in real time on GPU and DSP/SIMD targets.
The focus is the uplink receiver chain. The DU receives frequency-domain IQ samples from the radio unit over eCPRI, then estimates the channel, equalizes, detects the MIMO layers and decodes. You will compare classical algorithms (least squares, MMSE, linear MIMO detection) with learned ones (neural channel estimators, neural receivers), then port the best candidates to GPU under real-time constraints.
Possible topics, adjusted to your profile:
AI-based channel estimation and interpolation on DMRS pilots
Neural MIMO detection and soft-output (LLR) computation for LDPC decoding
PRACH preamble detection and timing estimation with machine learning
Beamforming weight computation for massive MIMO
Fronthaul IQ compression (block floating point) and its impact on model accuracy
MISSIONS :
The work runs from mathematical equations to measured, optimized code.
Learn the 5G NR physical layer (3GPP TS 38.211, 38.212, 38.214), the O-RAN 7.2x fronthaul, and the state of the art on AI for the physical layer
Build a link-level and maybe simulate in Python with 3GPP TDL/ CDL channel models and a classical baseline receiver
AI design: Design, train and evaluate a model. Compare BLER and throughput versus SNR against the baseline
Implement the baseline and the AI inference in C/C++ and CUDA, or with SIMD intrinsics. Quantize (FP16, INT8), profile, and meet the slot latency budget
Integrate and validate: Plug the algorithm into the DU test bench, fed with IQ samples captured from the fronthaul. Measure latency, accuracy and GPU load
Engineering practice: Code reviews, documentation, with final report and a presentation. An open-source contribution or a publication is possible depending on results
Qualifications
MATHEMATICS AND PHYSICS REQUIRED :
A strong applied-mathematics background is the core requirement: these algorithms are mathematics before they are code. Linear algebra, probability and signal processing are mandatory; the other areas are a strong plus and can be deepened during the internship.
Mathematics
Linear algebra: Key concepts: matrix products and inverses, QR, Cholesky, SVD, eigendecomposition, pseudo-inverse, Kronecker products · Used for: MIMO detection, MMSE equalization, beamforming, precoding
Probability and statistics: Key concepts: random variables, complex Gaussian distributions, covariance matrices, Bayes’ rule, hypothesis testing · Used for: noise modeling, soft-output (LLR) computation, detection thresholds
Estimation and detection theory: Key concepts: least squares, LMMSE, maximum likelihood, MAP, Cramér-Rao bound, false-alarm and detection probabilities · Used for: channel estimation, PRACH detection, time and frequency synchronization
Signal processing: Key concepts: Fourier transform, DFT/FFT, convolution, sampling theorem, filtering, windowing, interpolation, correlation · Used for: OFDM modulation, pilot interpolation, timing and frequency offset estimation
Information and coding theory: Key concepts: entropy, mutual information, channel capacity, LDPC and polar codes, belief propagation · Used for: decoding, soft-output design, performance bounds
Optimization: Key concepts: gradient descent (SGD, Adam), convex optimization, Lagrange multipliers, loss function design · Used for: model training, beamforming weights, power allocation
Multivariable and complex calculus: Key concepts: gradients, Jacobians, chain rule, Wirtinger derivatives · Used for: backpropagation through complex-valued layers
Numerical analysis: Key concepts: fixed-point and floating-point arithmetic (FP32, FP16, BF16, INT8), quantization error, conditioning, numerical stability · Used for: GPU and DSP implementation, model quantization, IQ compression
Physics :
Wave propagation: Key concepts: path loss, multipath, delay spread, coherence bandwidth · Used for: channel models, pilot density
Mobility: Key concepts: Doppler shift and spread, coherence time, time-varying channels · Used for: channel tracking across OFDM symbols
Antenna arrays: Key concepts: array factor, steering vectors, spatial correlation · Used for: beamforming, massive MIMO
RF impairments: Key concepts: phase noise, carrier frequency offset, IQ imbalance, amplifier nonlinearity · Used for: robust receiver design
TECHNICAL SKILLS
You should be ready to deep dive and learn one of the two acceleration targets, GPU or DSP/SIMD.
Then of course, you’ll practice:
C on Linux, with attention to memory layout and performance
Python with NumPy and SciPy, and PyTorch or TensorFlow for model training
Basics of parallel computing and performance measurement For GPU and DSP acceleration:
GPU programming: CUDA kernels, memory hierarchy, warps, streams; cuBLAS, cuFFT, TensorRT for inference
DSP and SIMD: fixed-point arithmetic, AVX2/AVX-512 or Arm NEON/SVE intrinsics, DSP processor architectures
Background, but not mandatory:
5G NR physical layer (3GPP TS 38.211, 38.212, 38.214)
O-RAN fronthaul (7.2x split, eCPRI) and DPDK
Link-level simulation tools: NVIDIA Sionna, MATLAB 5G Toolbox, OpenAirInterface, srsRAN
Git and profiling tools (NVIDIA Nsight, perf)
Informations supplémentaires
Le groupe Iliad s’engage pleinement en faveur de ses collaborateur·rice·s, à travers une politique stricte de non-discrimination et de lutte contre le harcèlement. Toutes les candidatures sont les bienvenues, quels que soient le parcours, l’âge, l’origine, le genre ou l’orientation sexuelle.
Summary
Design, implement and benchmark AI-assisted 5G physical-layer algorithms for an O-RAN 7.2x distributed unit uplink receiver, comparing classical and neural algorithms and porting them to GPU and DSP/SIMD targets. Requires a strong applied-mathematics background (linear algebra, probability, signal processing), C on Linux, Python, and CUDA or SIMD programming.
Responsibilities
Design, implement and benchmark AI-assisted physical-layer algorithms for an O-RAN 7.2x DU uplink receiver; compare classical and neural algorithms; port candidates to GPU/DSP in C/C++, CUDA or SIMD intrinsics; integrate into DU test bench; code reviews, documentation, final report and presentation.
Qualifications
Strong applied-mathematics background: linear algebra, probability and signal processing mandatory; ready to learn GPU or DSP/SIMD acceleration; C on Linux; Python with NumPy/SciPy; PyTorch or TensorFlow; basics of parallel computing and performance measurement.