From Data to Decisions
“A showcase of data visualization solutions that help users explore, understand, and act with confidence.”
GPU Demand and Capacity Visulisation
Industry: Cloud Computing & AI Infrastructur I Domain - GPU Capacity Planning
PROJECT BACKGROUND
As demand for AI workloads continues to grow, GPU providers need to efficiently manage limited GPU capacity across multiple customers, models, providers, and regions.
This project focused on improving the GPU management workflow by helping providers:
-
Understand where customer GPU demand is highest
-
Evaluate available capacity against customer requests
-
Make faster decisions on which requests to accept, reject, or counter-offer
MY ROLE
Experience Designer involved in Requirement Gathering, Alignment with stakeholders, Research, AND Design Solution
PROJECT YEAR
2025
The Challenge
The challenge was to simplify complex GPU demand and capacity data into intuitive visualizations that support faster operational decisions.
Challenge 1: Simplifying GPU Demand Hierarchies
Visualization Problem
Operations teams needed to make capacity planning decisions, but customer GPU demand was spread across multiple interconnected dimensions—including GPU models, hosting providers, regions, and geographical markets. The fragmented data made it difficult to identify demand patterns, anticipate capacity needs, and make informed planning decisions. Without a unified way to understand Customer requestrs across differnt regions, moels
Exploration 1 → Table based view
Approach
Started with a table-based view to organize GPU requests across the different hierarchy levels and validate the information structure.
What I Learned
The table provided a complete view of the data and worked well for looking up individual records. However, comparing demand across multiple levels required users to scan and mentally connect information, making it difficult to identify patterns and demand hotspots.
Exploration 2 → Multi-level accordian
Approach
Designed a multi-level accordion structure to represent the hierarchy and provide request visulisation by Geographical Market & GPU Model,
Users could expand and drill down through each level to explore demand distribution.
What I Learned
The accordion improved understanding of the parent-child relationships and made navigation through the hierarchy easier. However, comparing demand across different providers, regions, and markets still required excessive expansion and navigation, slowing down analysis.
Final Visualization → Matrix View
Solution
Designed a matrix-based visualization that consolidated the complete GPU demand hierarchy into a single view. The matrix allowed users to compare GPU models, hosting providers, regions, and geographical markets simultaneously.
How the Design Solved the Problem
-
Single consolidated view - Removed the need to navigate through multiple hierarchy levels.
-
Faster demand comparison - Enabled quick comparison of request volumes across GPU models, providers, and regions.
-
Demand hotspot identification - Helped teams quickly identify where customer demand was highest.
-
Better planning decisions - Supported faster capacity planning and resource allocation decisions.
User value
The final visualization transformed a complex hierarchical dataset into an intuitive comparison tool. Operations teams could quickly identify high-demand areas, understand demand distribution, and make informed decisions about GPU capacity planning.
Challenge 2: Enabling GPU Capacity Planning Decisions
Visualization Problem
Providers needed a better way to understand GPU capacity availability over time. Customer requests varied in duration and often overlapped, making it difficult to visualize available, reserved, and blocked capacity together with incoming requests.
The challenge was to design a scalable visualization that enabled providers to quickly evaluate requests and decide whether capacity should be reserved, declined, or counter-offered.
Exploration 1 - Timeline-Based Capacity View
Approach
Designed a split view with request cards on the left and a capacity graph on the right to visualize GPU capacity alongside selected requests.
What I Learned
The timeline successfully introduced a visual representation of capacity over time and helped users understand available, reserved, and blocked GPU capacity. However, as the number of selected requests increased (15–20 requests), the visualization became cluttered. Long-duration requests also became compressed when viewing larger timeframes, making them difficult to compare.
Exploration 2 - Stacked Capacity View
Approach
Replaced the timeline with a stacked bar graph to better support multiple requests across longer time periods.
What I Learned
The stacked visualization improved readability for larger datasets and longer timelines, reducing visual clutter. However, it sacrificed visibility into individual request durations, making it difficult for providers to compare the longest and shortest requests when evaluating capacity.
Final Visualization – Unified Capacity Planner
Solution
The final design integrated request duration directly within the request cards while using a single capacity timeline to visualize available, reserved, and blocked GPU capacity. Bringing related information together reduced context switching and created a more unified decision-making experience.
How the Design Solved the Problem
-
Capacity at a glance - Providers could immediately distinguish available, reserved, and blocked GPU capacity across the selected timeframe.
-
Request duration in context - Displaying request duration within each request card eliminated the need for a separate duration visualization and made comparisons easier.
-
Scalable visualization - Multiple requests could be viewed together without overwhelming the interface, even across longer time periods.
-
Reduced cognitive load - Combining request details, duration, and capacity into a single view minimized context switching and simplified analysis.
User value