Apple Silicon's Real Legacy: The Death of the Discrete Everything

The integration doctrine

When Apple shipped the M1, skeptics called unified memory a compromise. Five years later, every major chip vendor is racing toward the same architecture: CPU, GPU, NPU, and memory fused into a single package with shared bandwidth measured in terabytes per second.

Why discrete lost

The physics are unforgiving. Moving data between separate chips costs orders of magnitude more energy than computing on it. As AI workloads made memory bandwidth the binding constraint, the discrete GPU's architectural disadvantage became fatal for everything except the highest-end training clusters.

The new landscape

Laptops are already fully integrated. Workstations are following. Even datacenters are moving to integrated superchips for inference. The discrete component is not dead — but it is becoming what the mainframe became: a specialized tool for the few workloads that genuinely need it.