hardware//compute-bound

A workload is compute-bound when arithmetic execution is the limiting resource. Supplying more memory bandwidth would not materially accelerate it because compute units are already the bottleneck. Its siblings are memory-bound and I/O-bound (waiting on disks, networks or remote services), and roofline-style reasoning tells the two apart by comparing a workload's arithmetic demand with the data it has to move.

A workload is compute-bound when arithmetic execution is the limiting resource. Supplying more memory bandwidth would not materially accelerate it because compute units are already the bottleneck. Its siblings are memory-bound and I/O-bound (waiting on disks, networks or remote services), and roofline-style reasoning tells the two apart by comparing a workload's arithmetic demand with the data it has to move.

Optimization should follow that diagnosis: reduce operations, use faster kernels, exploit specialized instructions or add appropriate compute capacity. The label is conditional on hardware, shapes, precision and batch size. The same model can be memory-bound at small batches and more compute-bound after batching reuses the same weights across requests.

Bottlenecks are relationships between workload and machine, not permanent personality traits of algorithms.