Demo: Power-Aware Compute Allocation for Real-Time Multi-Cell mmWave Baseband Processing

Abstract

5G New Radio (NR) in the millimeter-wave (mmWave) Frequency Range 2 (FR2) imposes stringent baseband processing deadlines that must be met with minimal compute resources. As virtualized Radio Access Networks (vRANs) move baseband processing onto shared, general-purpose servers, a central question emerges: how many computational resources does each cell actually need, and how can a single server serve many cells at once without missing deadlines? We demonstrate Nexus, the first system to realize real-time, virtualized multi-cell mmWave baseband processing on a single server with heterogeneous compute, i.e., CPU cores and an Intel ACC100 eASIC virtualized into up to sixteen virtual functions (VFs). We demonstrate Nexus's random forest (RAF)-based scheduler with arbitrary cell configurations and observe the predicted CPU-core and VF allocations and their associated power cost. We then demonstrate a live end-to-end (E2E) emulation in which one server streams full-load uplink traffic while a distributed-unit (DU) server processes multiple concurrent FR2 cells and meets the three-slot (0.375 ms) PHY-processing deadline.