Teralynx T100 sparks back-of-envelope math on the path to one-petabit AI switches with 640 MMC connectors
jwt0625 · x · 2026-09-09
Discussing Marvell's Teralynx T100 AI data center switch (16×4 OSFP cages, 512×200G lanes), the author sketches what a one-petabit switch would take: doubling lane rates implies 5x the OSFPs, or 40 XPOs, or 640 MMC connectors with 5,120 active fibers plus CPO.
He notes Nvidia's planned SN6800 already uses 512 MMCs (with 4 ASICs), showing the path isn't sci-fi. With 8+8 DWDM lighting 8 fibers per MMC, you'd approach one MMC per scale-up ASIC.
Related event: Marvell Teralynx T100 and Petabit Switches Aim to Break the Network Wall(2 posts)→
More from Infra
- Semantic Cache Verification Measured: 30-40ms on the Hot Path vs 1.7s for an LLM Judge — Reasonable_Royal_621 · 2026-09-09
- Agents researching agents: Prime Intellect's open stack eases RL environments — seanwbren · 2026-09-09
- Open-source standalone RTX VSR upscaler hits 150-200fps on GPU — gerryn · 2026-09-09
- LARRI: open-source CLI that rents a GPU and tunnels open models to a local endpoint — gethackteam · 2026-09-09
- Fuzzing their Rust Postgres rewrite surfaced 20+ bugs in Postgres itself — DanielLockyer · 2026-09-09
- Singapore switches on biological data center with 20 CL1 units of living human neurons — Olivier__OG · 2026-09-09