Mind the Memory Gap: Unveiling GPU Bottlenecks in Large-Batch LLM Inference
This addresses resource inefficiency for practitioners deploying smaller language models in inference, offering incremental improvements through optimized memory allocation.
The paper tackles the problem of inefficient GPU resource utilization in large-batch LLM inference, revealing that it remains memory-bound due to DRAM bandwidth saturation, and proposes a Batching Configuration Advisor that reduces GPU memory requirements by up to 30% with minimal throughput loss.
Large language models have been widely adopted across different tasks, but their auto-regressive generation nature often leads to inefficient resource utilization during inference. While batching is commonly used to increase throughput, performance gains plateau beyond a certain batch size, especially with smaller models, a phenomenon that existing literature typically explains as a shift to the compute-bound regime. In this paper, through an in-depth GPU-level analysis, we reveal that large-batch inference remains memory-bound, with most GPU compute capabilities underutilized due to DRAM bandwidth saturation as the primary bottleneck. To address this, we propose a Batching Configuration Advisor (BCA) that optimizes memory allocation, reducing GPU memory requirements with minimal impact on throughput. The freed memory and underutilized GPU compute capabilities can then be leveraged by concurrent workloads. Specifically, we use model replication to improve serving throughput and GPU utilization. Our findings challenge conventional assumptions about LLM inference, offering new insights and practical strategies for improving resource utilization, particularly for smaller language models. The code is publicly available at https://github.com/FerranAgulloLopez/vLLMBatchingMemoryGap.