Roofline Analysis and Mixed-Precision Gradients for Matmul
Asked at Microsoft · Deep Learning & Architectures · Difficulty: Medium · Last asked Mar 22, 2026
A GPU has a peak FP16 throughput of 312 TFLOP/s and an HBM bandwidth of 1.5 TB/s. Use the roofline model to analyse matrix multiplications on it, then reason about the gradient of a residual connection and about where casting from FP32 to FP16 causes trouble.
Loading the problem…
More Microsoft interview questions
- LLM conceptual questions
- Tiny Next-Token Prediction Model (PyTorch Fill-In)
- DNA Shotgun Sequencing: Chain and Eulerian Path Reconstruction
- Design Search Autocomplete / Typeahead
- Design a Chatbot Personalization / Memory System
- Design a Job Scheduler / ETL Pipeline