As AI-generated code becomes the default rather than the exception, developers inherit a new problem: that code is written to be correct, not to be fast on the specific hardware in front of them. Our NVIDIA GTC 2025 session was about closing that gap.
The gap AI code leaves behind
A model generating code has no measurement of your machine. It cannot know your memory bandwidth, your occupancy limits, or which of two equally readable implementations wins on your GPU. It produces something plausible. Whether it is efficient is an empirical question nobody asked.
Context-aware optimisation
The talk covered techniques for adapting code to exploit NVIDIA's parallel processing characteristics: profiling to find where runtime actually goes, generating candidates aware of the target architecture, and validating each one on real workloads rather than trusting a static heuristic.
Beyond raw speed
Runtime is the headline metric, but the same loop improves energy efficiency and scalability, and removes the bottlenecks that quietly cap throughput as workloads grow.
