Data Movement Is Architecture

Accelerator performance is not only core throughput. SerDes links, memory movement, packetization, lane count, latency, topology, and synchronization can dominate the useful shape of a system.

A design with powerful compute islands connected by expensive data highways asks the software to pay for distance. The programmer has to know where data lives, how often it crosses a link, and whether the work can remain local long enough to justify the transfer.

Meshes, Islands, and Locality

Mesh and lattice-style thinking keeps locality visible. Island-style accelerator systems can still work well, but the interconnect becomes part of the programming model. A kernel that looks efficient in isolation may be poor system design if every useful step moves data across a high-cost boundary.

The useful question is measurable: how much energy, latency, and engineering complexity are spent moving data compared with computing on it?

Useful Scope

This note is not a vendor critique. It is a topology critique. Large-scale computation needs source code and hardware layout to preserve locality, causality, and failure boundaries instead of treating data movement as an invisible detail.