UiPath Scales Agentic AI with Shared GPU Fleet on Google Cloud

UiPath leveraged the Google Cloud AI Hypercomputer, an integrated system of performance-optimized hardware and flexible software, to support its growing AI scale. This environment minimizes friction between hardware and software, allowing engineering teams to focus on model performance.

UiPath transitioned from isolated GPU clusters to a shared Google Cloud GPU fleet, managed by its machine learning services platform. This approach prioritizes work across teams and time windows, balancing inference during the day with batch training at night to maximize utilization and reduce contention.

To overcome supply bottlenecks for high-end GPUs, UiPath uses Google Cloud's Dynamic Workload Scheduler (DWS) to schedule training runs in advance and secure capacity. This allows for predictable capacity planning rather than reacting to scarcity, ensuring consistent access to necessary compute resources.

UiPath now utilizes both A3 VM instances with NVIDIA H100 GPUs for heavy-duty training and fine-tuning, alongside G4 VM instances with NVIDIA RTX Pro 6000 for cost-effective inference workloads. This mix optimizes for both speed and cost, especially for lighter inference tasks.

UiPath is pioneering agentic AI, requiring significant computational power and reliable infrastructure to deploy autonomous agents for complex business processes. This shift necessitates efficient orchestration of hundreds of GPUs for both massive training jobs and real-time inference without escalating costs or latency.

With consistent access to Google Cloud GPUs, UiPath can bring advanced models into production, enabling high-accuracy data extraction from unstructured documents. Customers like Omega Healthcare and Thermo Fisher Scientific are experiencing significant improvements in processing time and accuracy.