Beluga: A CXL-Based Memory Architecture for Scalable and Efficient LLM KVCache Management
Open Access
- 2 April 2026
- journal article
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (1) , 1-29
Abstract
The rapid increase in LLM model sizes and the growing demand for long-context inference have made memory a critical bottleneck in GPU-accelerated LLM serving. Although high-bandwidth memory (HBM) on GPUs offers fast access, its limited capacity necessitates reliance on host memory (CPU DRAM) to support large KVCache. However, the maximum DRAM capacity is constrained by the limited number of memory channels per CPU socket. To overcome this limitation, current systems often adopt RDMA-based disaggregated memory pools, which introduce significant challenges including high access latency, complex communication protocols, and synchronization overhead. Fortunately, the emerging CXL technology introduces new opportunities in KVCache design. In this paper, we propose Beluga , a novel memory architecture that enables GPUs and CPUs to access a shared, large-scale memory pool through CXL switches. By supporting native load/store access semantics over the CXL fabric, our design delivers near-local memory latency, while reducing programming complexity and minimizing synchronization overhead. We conduct a systematic characterization of CXL-based memory pool and propose a set of design guidelines. Based on Beluga , we design and implement Beluga -KVCache, a system tailored for managing the large-scale KVCache for LLM inference. Beluga -KVCache achieves an 89.6% reduction in Time-To-First-Token (TTFT) and 7.35x throughput improvement in vLLM compared to RDMA-based solutions. To the best of our knowledge, Beluga is the first system that enables GPUs to directly access large-scale memory pools through CXL switches (Marvell XConn XC50256), marking a significant step toward low-latency, shared access to vast memory resources by GPUs.Keywords
All Related Versions
This publication has 38 references indexed in Scilit:
- Enabling Efficient Large Recommendation Model Training with Near CXL Memory Processing2024
- A CXL- Powered Database System: Opportunities and Challenges2024
- Distributed GPU Joins on Fast RDMA-capable NetworksProceedings of the ACM on Management of Data, 2023
- GPU-Initiated On-Demand High-Throughput Storage Access in the BaM System Architecture2023
- Pond: CXL-Based Memory Pooling Systems for Cloud Platforms2023
- Exploiting CXL-based Memory for Distributed Deep Learning2022
- Enabling CXL Memory Expansion for In-Memory Database Management Systems2022
- ScaleStore: A Fast and Cost-Efficient Storage Engine using DRAM, NVMe, and RDMA2022
- Rethinking Stateful Stream Processing with RDMA2022
- Scalable RDMA RPC on Reliable Connection with Efficient Resource Sharing2019