Beaver enables fast GPU sharing for AI and critical radio access tasks
Beaver: Elastic GPU Sharing between ML and Latency-Critical vRAN Workloads
Distributed, Parallel, and Cluster ComputingNetworking and Internet Architecture
Summary
Some computer systems use GPUs (powerful processors) to run different kinds of workloads at the same time. But mixing tasks that need quick responses, like radio signals for phones, with slower tasks, like machine learning, is tricky. The authors designed Beaver to share GPUs between these tasks while making sure the fast radio work always meets strict timing deadlines. Beaver adjusts how much GPU power each task gets and changes machine learning programs to avoid slowing down the radio work. Their tests show that Beaver keeps up with tight deadlines for radio tasks while still doing a lot of machine learning work.
What this means in practice
- •For network infrastructure operators: Run latency-sensitive vRAN workloads and machine learning services together on GPUs without missing strict radio timing deadlines.
- •For cloud ai service providers: Deploy large machine learning model serving alongside real-time communication workloads on shared GPUs, maintaining high throughput and low latency.$Commercial implications: Allows offering combined AI and vRAN GPU services with guaranteed latency, enabling new cloud products for telecom and AI customers.
Authors
Yuncheng Yao, Zhenzhou Qi, Junyao Zheng, Chung-Hsuan Tung, Danyang Zhuo, Tingjun Chen
Abstract
Within the shared industry vision of AI-RAN, AI-and-RAN seeks to co-locate virtualized radio access network (vRAN) workloads and AI services on shared GPUs. This sharing is inherently asymmetric: vRAN workload is latency-critical, whereas the machine learning (ML) workload is a throughput-oriented, best-effort co-tenant. We present Beaver, a GPU sharing system that jointly manages compute and memory resources while protecting the vRAN's strict processing deadline. Beaver sizes the vRAN's streaming multiprocessor (SM) allocation from each slot's scheduled workload, repartitions SM allocations at slot granularity, and rewrites compiled ML kernels to yield HBM bandwidth during the vRAN's memory-critical phases. We implement Beaver and evaluate it using NVIDIA Aerial with heterogeneous multi-cell workloads, real-world cellular traces, production inference kernels, and full-stack LLM serving. On an H200 GPU, Beaver keeps the vRAN's p99.9 latency within its 1.5ms uplink deadline while retaining 74% of Llama-3.3-70B serving throughput. It also incurs no observed deadline misses under replayed cellular traces, protects a 375us downlink deadline, and generalizes to other GPUs including A100, GB10 and GH200.