Understanding Bottlenecks for Efficiently Serving LLM Inference With KV Offloading
Fuente:
arXiv
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866917226858151936 |
|---|---|
| author | Meng, William Lee, Benjamin Wang, Hong |
| author_facet | Meng, William Lee, Benjamin Wang, Hong |
| contents | KV cache offloading enables long-context LLM inference by storing caches in CPU DRAM, but PCIe bandwidth limitations create severe bottlenecks. In this paper, we develops an analytical framework that derives $κ_{\text{crit}}$, the critical cached-to-prefill token ratio where execution becomes memory-bound and show typical workloads exceed this threshold by orders of magnitude. Empirical characterization reveals 99\% of latency spent on transfers and serving offloaded requests results in GPU's consuming only 28\% of their rated TDP, motivating our proposed optimizations for hardware interconnects, model architectures, and scheduling algorithms. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_19910 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Understanding Bottlenecks for Efficiently Serving LLM Inference With KV Offloading Meng, William Lee, Benjamin Wang, Hong Hardware Architecture Distributed, Parallel, and Cluster Computing KV cache offloading enables long-context LLM inference by storing caches in CPU DRAM, but PCIe bandwidth limitations create severe bottlenecks. In this paper, we develops an analytical framework that derives $κ_{\text{crit}}$, the critical cached-to-prefill token ratio where execution becomes memory-bound and show typical workloads exceed this threshold by orders of magnitude. Empirical characterization reveals 99\% of latency spent on transfers and serving offloaded requests results in GPU's consuming only 28\% of their rated TDP, motivating our proposed optimizations for hardware interconnects, model architectures, and scheduling algorithms. |
| title | Understanding Bottlenecks for Efficiently Serving LLM Inference With KV Offloading |
| topic | Hardware Architecture Distributed, Parallel, and Cluster Computing |
| url | https://arxiv.org/abs/2601.19910 |