GPU Roofline Model Study for a Computer Vision Inference Workload
Overview
What this challenge is about.
Construct empirical roofline plots for the current GPU (A100-class) and next-gen GPU (H100-class) using vendor-published peak FLOPs and measured peak memory bandwidth. Profile the ResNet-50 inference workload at batch sizes 1, 8, 32, and 128 with Nsight Compute, capturing arithmetic intensity, achieved FLOPs/s, and achieved memory bandwidth per kernel. Place each (batch, GPU) point on the roofline. Identify compute-bound vs. memory-bound regimes per batch size. Deliver the roofline plots, per-kernel profiling tables, and a 10-page placement recommendation memo.
The Brief
What you'll do, and what you'll demonstrate.
Build empirical roofline models for two GPU generations, profile a ResNet-50 inference workload, and recommend batch-size-and-SKU placement.
Earning criteria — what you'll demonstrate
- Construct an empirical roofline model from measured peak FLOPs and bandwidth
- Profile a real GPU workload with Nsight Compute
- Map workload regimes to GPU SKUs based on arithmetic intensity
- Communicate placement recommendations to a capacity team
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.