Nvidia has shown a co-packaged optics prototype it calls optics-on-interposer, presented at SEMICON West under the heading of optical integration for the next generation of AI. The optical engines move off the organic package substrate and onto the silicon interposer that carries the compute dies.
The point of the move is to take the organic substrate out of the electrical path between the switch or accelerator die and the optics. What remains is a short, fine-pitch connection across silicon, which costs less energy per bit and allows more connections in the same width. Nvidia cites higher bandwidth density and lower power as the two benefits. A separate wavelength-multiplexing demonstration shown at the same event carried eight wavelengths at up to 32 Gb/s each, or 256 Gb/s on a single fibre.
It is research, not a product. No commercialisation date has been given, and the company describes it as under evaluation for possible future use. The difficulties are the predictable ones: a laser is temperature-sensitive and an interposer sits directly above the hottest silicon in the system, and fibre attachment has to survive assembly, reflow and field thermal cycling. Nvidia’s own researchers put thermal and packaging challenges at the top of the list.
The direction, though, has been consistent for two years. Optics started in a pluggable module at the faceplate, moved onto the package, and are now being tried on the interposer — each step shortening the copper run because copper is what stops scaling first.
For systems engineers the lesson is not the prototype but the trajectory: every generation moves the electrical-to-optical boundary closer to the die, and every move makes the optics harder to service and the thermal design harder to get right.
Source: TrendForce