Introducing Photon1

Optical die-to-die · two GPUs, one link optical die-to-die · GPU to GPU two dual-die GPUs · one Photon¹ link
In Development

Light between dies

Photon1 is an optical die-to-die interconnect: a chiplet that sits beside a GPU or memory controller and carries its die-to-die traffic over light, so reticle-sized dies and their memory work as one system.

It links GPU to GPU and GPU to memory in the package, across the board, or across the rack, with the host’s protocol layers running unchanged.

Interface Drops in beside the host with an optical link beyond the package, and the host’s protocol layers run unchanged
Energy per bit Well under a picojoule at any reach, as a design target, against 5 to 10 pJ for serial links that leave the package
Reach In the package, across the board, or across the rack, over the same link
Status In development, sold as an independent chiplet beside any GPU or memory controller
Energy per bit · by reach 0 1 pJ 10 pJ PER BIT Serial link, off the package 5 to 10 pJ Electrical, in the package well under 1 pJ Photon¹, in the package Photon¹, across the board Photon¹, across the rack well under 1 pJ at every reach · design target Link energy only, bars to scale. Electrical links are cheap inside the package, and jump once they leave it. Photon¹ holds a sub-picojoule figure at any reach.
Reach · by link type 1 mm 1 cm 10 cm 1 m 10 m 100 m Electrical, in the package a few centimetres Serial link, board and cable a few metres Photon¹, one link PACKAGE BOARD RACK Reach on a log scale. Electrical links end within centimetres in the package, or a few metres of board and cable. Photon¹ is one link from the package to the rack.

Energy that holds with distance

Electrical links are cheap only inside the package. Past it, every lane becomes a serial link at several picojoules per bit. Photon1 keeps its electronics at logic speed, so the energy per bit barely moves whether the far die is a millimetre or a rack away.

One link, every scale

The same link joins dies inside a package, GPUs across a board, and boards across a rack: one optical interconnect for scale-in and scale-up, instead of a different technology at each step.

Drop-in for the host

Photon1 sits beside the host as an independent chiplet. The host’s protocol layers run unchanged, so partners place it next to their own GPU or memory controller without redesigning the die.

Explore Scale1