Exascale Computing
Exascale systems perform on the order of a billion billion operations per second, a milestone that reshaped how large simulations are written.
What exascale means
An exaFLOP is 10^18 floating-point operations per second, a quintillion. An exascale computer sustains roughly that rate on demanding benchmarks. Reaching it required assembling millions of cores, mostly on GPUs, with commensurate memory, network, and storage, and holding total power to a feasible level.
The hard constraints
- Power: efficiency, measured in operations per watt, is the binding limit, met largely by accelerators
- Parallelism: exploiting millions of threads demands parallelism at every level
- Resilience: with so many parts, failures are routine, so checkpointing and fault tolerance are essential
- Data movement: moving bytes costs more energy than computing on them, so codes minimize communication
Why it is not just bigger
Codes written for petascale often do not simply scale up. The serial and communication fractions that were tolerable become dominant (Amdahl's law), and reliance on CPU-only execution wastes most of the machine. Reaching exascale performance typically meant restructuring around GPUs, reducing global synchronization, and overlapping communication with computation.
Energy per operation
Because moving data dominates energy use, exascale design favors keeping data local, reusing it in fast memory, and using lower precision where accuracy permits. Algorithmic changes that cut communication can matter more than hardware.
What it enables
Exascale supports higher-fidelity science: finer meshes, longer integrations, and coupled multiphysics. For fusion, that means fuller-device turbulence and transport studies. Kronos develops its machine designs in simulation on such large systems; the machines themselves are not yet built.