LLM serving should triage by five-resource analytical floors and wall ordering, not grid search; on 16×H20, TP16 is capacity-capped at ~70 while EP+DP attention reaches ~644 concurrent 8K requests.
Mind the Memory Gap: Unveiling GPU Bottlenecks in Large-Batch LLM Inference
5 Pith papers cite this work. Polarity classification is still indexing.
5
Pith papers citing it
abstract
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.
representative citing papers
Watt Counts supplies over 5,000 energy measurements across 50 LLMs and 10 GPUs and shows that hardware-aware selection can reduce server-scenario energy use by up to 70 percent with little effect on user experience.
The paper analyzes CPU bottlenecks in agentic AI serving, selects representative workloads, and demonstrates that CPU-aware scheduling optimizations COMB and MAS can reduce P50 latency by up to 1.7x and total latency by up to 2.49x on two hardware systems.
Learning-augmented LRU achieves 1-consistency and O(k)-robustness for GPU caching with low overhead, implemented in LCR to cut P99 TTFT by up to 28.3% on LLM workloads and raise throughput by up to 24.2% on DLRM workloads.
citing papers explorer
-
Towards Understanding, Analyzing, and Optimizing Agentic AI Execution: A CPU-Centric Perspective
The paper analyzes CPU bottlenecks in agentic AI serving, selects representative workloads, and demonstrates that CPU-aware scheduling optimizations COMB and MAS can reduce P50 latency by up to 1.7x and total latency by up to 2.49x on two hardware systems.