Understanding the Performance Behaviors of End-to-End Protein Design Pipelines on GPUs

TL;DR

This work addresses the underexplored issue of end-to-end GPU performance in protein-design pipelines. It employs a systematic profiling approach at both component and full-pipeline levels across varying inputs and sampling settings, revealing generally low GPU utilization and strong sensitivity to sequence length and sampling strategies. The authors demonstrate meaningful gains from GPU co-location and identify challenges in multi-GPU scaling, guiding future directions in cost-aware metrics, large-scale deployment, and agent-assisted orchestration. By releasing an open-source pipeline and profiling tools, the study provides a practical foundation for researchers and developers to optimize GPU-centric protein-design workflows and informs resource planning for scalable in silico design campaigns.

Abstract

Recent computational advances enable protein design pipelines to run end-to-end on GPUs, yet their heterogeneous computational behaviors remain undercharacterized at the system level. We implement and profile a representative pipeline at both component and full-pipeline granularities across varying inputs and hyperparameters. Our characterization identifies generally low GPU utilization and high sensitivity to sequence length and sampling strategies. We outline future research directions based on these insights and release an open-source pipeline and profiling scripts to facilitate further studies.

Figures (4)

• Figure 1: Four major stages of a protein-design pipeline.
• Figure 2: Profiling of individual pipeline components across varying inputs and hyperparameters. Vina-CPU does not utilize GPU resources. Vina-GPU is implemented using OpenCL, which does not support spatial profiling; thus, only temporal utilization is reported.
• Figure 3: Normalized runtime of pipeline components across varied sampling.
• Figure 4: GPU temporal utilization over execution time.