This note describes the architecture behind the Volantis optical interconnect and the measurements we use to characterise it. It is intended for architects evaluating memory-bound serving at rack scale.

Architecture

Optical engines are co-packaged alongside the accelerator die. Waveguides route off-package to a memory shelf, extending usable reach from millimetres to roughly one metre while holding latency within the budget that SRAM-class access requires.

Results

Across our current test fleet we measure roughly 100x memory reach relative to an electrical interposer baseline, with 6x better energy per token and 15x better tokens per second per dollar on models above ten trillion parameters.