Prism: New Framework Improves Cost Efficiency of Multi-LLM GPU Serving
Researchers have developed Prism, a memory-centric framework that uses GPU memory ballooning to improve the efficiency of serving multiple large language models simultaneously. The system addresses the challenge of maintaining availability for many LLMs while managing dynamic, bursty usage patterns that existing approaches struggle to handle. The framework has been deployed in production across 10,000+ GPUs and offers practical benefits as token prices decline and resource efficiency becomes increasingly important.
Prism is a new LLM co-serving framework designed to optimize GPU resource utilization when running multiple language models. The system applies memory ballooning techniques to dynamically reclaim and reallocate memory across different models, enabling both spatial sharing (running multiple models on one GPU) and temporal sharing (switching between models over time) under a unified approach. The researchers analyzed production traces and identified a dynamic bursty-group pattern where sets of models become active together and shift over time—a variability that existing space- and time-sharing approaches cannot adapt to without sacrificing either service level objectives or efficiency. Prism's core component, a balloon driver called kvcached, has been open-sourced and is currently deployed across 10,000+ GPUs in production environments. The work addresses a growing industry need as inference providers must maintain availability for many models, including low-volume but essential ones, while token prices fall and cost efficiency becomes paramount.
What different sources said
- arXiv cs.AICenter
Prism: Cost-Efficient Multi-LLM Serving via GPU Memory Ballooning
Related
Topology-Aware Thermodynamics Improves DNA Probe Specificity Design
Researchers developed a new framework for designing DNA probes that accounts for the spatial organization of matched sequences, not just overall thermodynamic stability. Traditional methods rely on scalar measures like melting temperature and free energy, which miss how mismatches are distributed along the probe. The approach could improve diagnostic accuracy in applications like HPV detection and gene expression profiling.
Study Identifies Optimal Thermal Dose for Combining Focused Ultrasound with Immunotherapy in Tumors
Researchers used multimodal PET imaging to identify an optimal thermal dose range for focused ultrasound ablation that destroys tumor tissue while preserving conditions for immunotherapy delivery. The study found that excessive heating collapses blood vessels needed for antibody access, while insufficient heating fails to adequately reduce tumor burden. The findings could guide clinical design of combination treatments pairing thermal ablation with immunotherapies.
Plant MSH1 Protein Functions as Mismatch-Directed Nuclease for Organelle Genome Maintenance
Researchers have identified the precise mechanism by which the AtMSH1 protein in Arabidopsis plants recognizes and cleaves DNA mismatches and lesions, preventing mutations in organellar genomes. The protein combines a DNA mismatch recognition module with a nuclease domain that makes staggered cuts at specific positions relative to DNA damage. This discovery explains how plants maintain unusually low mutation rates in their mitochondrial and chloroplast DNA compared to other eukaryotes.