Whenever you're ready, walk me through how you'd approach this.
Explain the trade-offs between memory bandwidth and compute utilization in your previous AI projects.
Discuss how you identified whether workloads were memory-bound or compute-bound, which profiling data you collected, and what optimizations you applied. Cover at least one training or inference example, such as changing tensor layouts, fusion, precision, batching, parallelism, or model architecture. Distinguish improvements in raw throughput from improvements in end-to-end latency, cost, and energy efficiency. State how you validated that an optimization improved production behavior without changing model quality.