Silicon Photonics

Co-packaged silicon photonics is the networking technology that makes rack-scale AI compute physically possible. In the Vera Rubin platform, NVIDIA's Spectrum-6 Ethernet switch integrates silicon photonics directly into the switch package, delivering 102.4 Tb/s per switch with a 64x improvement in signal integrity over conventional electrical interconnects.

Traditional copper-based interconnects hit fundamental limits at the bandwidth densities required for rack-scale AI. Signal degradation, power consumption, and thermal challenges all compound as data rates increase. Silicon photonics solves this by converting electrical signals to light at the chip package boundary, transmitting optically within the rack, and converting back at the destination. The result is dramatically higher bandwidth at lower power per bit, with signal integrity that scales rather than degrades with distance.

For the NVL72 rack, silicon photonics enables the 260 TB/s aggregate bandwidth across 72 GPUs. Without optical interconnects, the cableless modular tray design — where trays connect via blind-mate connectors rather than cables — would not be feasible at these data rates.

The power figure, in watts (added 2026-09-24)

The concept's missing number was the power saving itself. NVIDIA's technical blog gives it, as a vendor claim about switches not yet shipping on its date (2025-08-18): pluggable optics draw "often 30W per interface" and co-packaged optics "as low as 9W". The electrical path from switch chip to optics loses up to 22 dB at 200 Gb/s per channel with pluggables and about 4 dB with co-packaged optics (NVIDIA CPO blog).

Three things follow, all our arithmetic on the vendor's numbers:

• 30 W to 9 W is about 3.3x, not the 3.5x NVIDIA and Broadcom both headline. Broadcom's "70% reduction" is also 3.3x (Broadcom). The claims converge at about 3.3x on the numbers each vendor actually prints.
• The 64x signal-integrity figure below is the 22 dB to 4 dB gap converted to a power ratio (18 dB ≈ 63x). It compares the electrical path to a pluggable module with the path to an on-package optical engine, not optics with copper cabling.
• Per bit, if an "interface" is one 800 Gb/s port, that is about 37 pJ/bit against 11 pJ/bit [unverified: the blog does not define "interface"].

Key Claims

• 102.4 Tb/s per Spectrum-6 switch — Co-packaged silicon photonics integrated at the switch level. Evidence: strong (NVIDIA Vera Rubin)
• 64x signal integrity improvement — Over conventional electrical interconnects. Evidence: strong (NVIDIA Vera Rubin) Clarified 2026-09-24: the same vendor's blog shows this is electrical loss of up to 22 dB (to a pluggable module) against about 4 dB (to a co-packaged optical engine), a best case. Evidence: moderate (vendor-stated) (NVIDIA CPO blog)
• Optics power: "often 30W per interface" for pluggables against "as low as 9W" co-packaged — typical against best case, about 3.3x. Evidence: moderate (vendor-stated, pre-shipping, 2025-08-18) (NVIDIA CPO blog)
• Two vendors headline about 3.5x lower optical power for co-packaged optics. Evidence: moderate (two vendors, both self-reported) (Broadcom; NVIDIA release)
• Enables cableless rack design — Optical interconnects make blind-mate tray connectors viable at rack-scale bandwidth. Evidence: moderate (inferred) (NVIDIA Vera Rubin)

Open Questions

• What are the yield challenges for co-packaged silicon photonics at scale production?
• How does the cost premium compare to conventional electrical interconnects, and when does it cross over?
• Can competing rack-scale architectures (hyperscaler ASICs) adopt silicon photonics independently of NVIDIA?
• What is the power saving per bit, independently measured? The vendor figures are 30 W against 9 W per "interface", undefined and unmeasured by anyone else.

Related Concepts

• Rack-Scale AI Compute — The system architecture that silicon photonics enables (held on the hardware rung)

Backlinks

Pages that reference this concept:

• NVIDIA Vera Rubin Platform
• Networking index and frontier

Changelog

• 2026-04-09 — Created from the NVIDIA Vera Rubin platform blog (held on hardware).
• 2026-09-24 — Added the watts: NVIDIA's CPO blog gives 30 W (pluggable, typical) against 9 W (co-packaged, best case) and 22 dB against 4 dB of electrical loss. Clarified that the 64x signal-integrity figure is that loss gap, measured against pluggable modules. Linked the two CPO switch sources on this rung.