The VRAM Shortcut Trap

Senior Computer Vision Engineer interview at Google DeepMind, and the interviewer asks:

“We are passing high-resolution medical images through a deep, 50-layer CNN. To save VRAM on our H100 GPUs, a junior proposes dropping zero-padding on all convolutions, arguing we only lose a tiny 2-pixel border per layer. Do you approve this PR?”

Don’t say: “Yes, it’s a smart micro-optimization. Valid convolutions skip the zero-computation, saving memory bandwidth and FLOPs. A tiny edge crop on a 4K scan is statistically insignificant to the final classification.”