Overview
- Nvidia detailed CPO-based interconnects for rack-scale AI, committing to Quantum‑X InfiniBand in early 2026 and Spectrum‑X Ethernet in the second half of 2026.
- Quantum‑X switches are specified at 115 Tb/s with 144×800 Gb/s ports, include 14.4 TFLOPS of in‑network processing with SHARP support, and use liquid cooling.
- Spectrum‑X will use the Spectrum‑6 ASIC, with the SN6810 at 102.4 Tb/s (128×800 Gb/s) and the SN6800 at 409.6 Tb/s (512×800 Gb/s), both liquid‑cooled.
- Nvidia argues CPO cuts electrical loss and per‑port power versus pluggables (example figures: ~22 dB to ~4 dB loss and ~30W to ~9W), and claims 3.5× better power efficiency, 64× signal integrity, 10× resiliency, and ~30% faster deployment.
- The plan aligns with TSMC’s three‑stage COUPE path (OSFP engines → CoWoS/CPO at motherboard → in‑package photonics), as industry rivals including Broadcom, AMD (via Enosemi), and IBM pursue their own photonics efforts.