At ISC 2026, we will be discussing performance at length in many of the technical sessions during the week. Meanwhile, on the show floor, one exhibitor will highlight how much of that performance is being quietly wasted.
Peak performance looks impressive on paper, but becomes irrelevant the moment a real workload starts. Unlike traditional applications, these workloads do not follow neat, stable patterns. Instead, they fluctuate: stalling during communication phases and surging during compute-heavy steps.
Despite these variations, much of the underlying infrastructure operates under the assumption that nothing changes once execution begins. Frequency remains fixed, power is allocated evenly, and systems are designed to run conservatively rather than responsively. Systems are designed to achieve peak performance, but in practice, they often operate far below that potential – not because the hardware cannot deliver, but because the applications don’t fully optimize the use of that hardware at runtime.
The uncomfortable truth about HPC optimization is that a significant portion of the process occurs before the job even starts. Once execution begins, systems typically lock into their initial configuration.
Static Systems in a Dynamic World
Energy Aware Runtime (EAR) at Booth C11 challenges this traditional model by treating execution as a dynamic process – one that should be continuously refined in real time based on actual workload behavior, rather than static expectations. This means adjusting frequency, redistributing power, and responding to workload phases as they unfold. Static configurations are not best practice; they are a constraint we have normalized because they make systems easier to manage.
In many data centers, power caps are treated as hard limits that must not be exceeded. The result is a cautious operational model that prioritizes predictability over efficiency. Systems are deliberately running below their true capacity to avoid risk, even as energy costs continue to rise.
The usual objection is predictability. If systems start adapting during runtime, do we risk losing control? But what we call predictable is often just inefficient. Systems are locked into safe configurations not because they reflect real workload behavior, but because they are easier to manage. But that comes at a cost.
Power is not just a constraint; it is a resource. Right now, that resource is routinely underused because systems cannot react in real time. Data centers end up with the worst of both worlds: conservative settings, underutilized capacity, and growing operational costs. At this point, the bottleneck is no longer hardware efficiency; it is the absence of runtime intelligence.
In practice, this inefficiency is visible in how systems behave. During communication-heavy phases, nodes remain underutilized yet consume near-peak power. During compute-intensive phases, performance is constrained by conservative power settings rather than hardware capability. The workload shifts, but the operation of the hardware does not.
AI Makes This Impossible to Ignore
In the realm of AI, runtime inefficiencies are not merely inconveniences; they determine the cost of doing business. Long-running training jobs, fluctuating utilization, and tightly constrained power budgets mean that even small inefficiencies have enormous effects on the bottom line.
As demand continues to grow, “good enough” configurations are no longer sufficient. Power availability is not scaling at the same pace as compute demand, particularly in Europe, where energy price constraints, regulatory pressure, and sustainability targets are tightening simultaneously. Under these conditions, static optimization ceases to be a compromise; it becomes a liability. ISC 2026 will showcase what comes next in hardware: new architectures, accelerators, and systems. But EAR asks a more fundamental question that often goes unasked: why are we still wasting performance and energy in deployed systems and how long are we willing to accept that?
