Project
GPUVision
GPUVision is a configurable C++17 GPU simulator that models warp execution, streaming multiprocessor activity, memory delays, and scheduling algorithm performance. It includes a benchmarking framework for configurable workloads, exports performance results to CSV and JSON, and provides a Next.js dashboard for visualizing warp states, SM activity, execution timelines, and scheduler metrics.
Overview
GPUVision is a configurable simulator for studying GPU warp execution, streaming multiprocessor activity, memory delays, and scheduler behavior.
Problem
GPU scheduling behavior can be difficult to reason about from raw metrics alone. A simulator needs clear workload configuration, comparable scheduling runs, and visual output that makes warp and SM state changes easier to inspect.
Solution
The project uses a C++17 simulation engine with benchmark exports to CSV and JSON, then surfaces execution timelines and scheduler metrics through an interactive Next.js dashboard.
Key Features
- Configurable GPU warp and instruction simulation
- Scheduler performance comparisons
- IPC, total-cycle, SM-utilization, and warp-completion metrics
- CSV and JSON benchmark exports
- More than 14 Google Test unit tests
- Interactive Next.js performance dashboard
What I Learned
- How scheduler policy affects throughput and completion timing under configurable workloads
- How to validate simulation behavior with focused unit tests
- How structured benchmark exports make systems data easier to analyze in a web dashboard
Future Improvements
- Add more scheduler policies for comparison
- Expand workload presets for different memory-delay patterns
- Add richer dashboard filters for benchmark analysis