chance.tarver
Projects

2017–Present · Lead, platform & testbed

GPU-Based AI-RAN

A through-line in my work since 2017: running RAN workloads on commodity GPUs. Started with LDPC decoding for vRAN, now driving AI-RAN testbed development on a cuBB DU with a proprietary MIMO RU.

The disaggregated-RAN bet has always been that you can run a base station’s L1 on a general-purpose accelerator and still hit the timing budget. The work has moved with the question: in 2017–2021 the binding workloads were LDPC decoding for vRAN (high-throughput, low-latency, codec-rate flexibility) and massive-MIMO array linearization (per-antenna DPD that scales). The contribution on the LDPC side was a GPU decoder integrated into the OpenAirInterface NR stack; on the linearization side, a GPU implementation that made per-antenna DPD across a MIMO array tractable. Today the binding question is how do you bring AI features into the RAN end-to-end — develop in simulation, train against representative data, drop into a real DU/RU stack OTA, validate in the field. That is what the current Samsung AI-RAN testbed work is about: GPU-based AI-RAN on a cuBB-based DU with a proprietary MIMO RU and an end-to-end “sim-to-field” workflow.

Same fundamental bet (GPU-on-RAN), seven years apart, different binding constraints.

Selected publications

AI-RAN (2024–Present)

GPU-Accelerated RAN Workloads (2017–2021)