Lightmatter Passage L20 CPX: BiDi NPO That Halves the Fiber Plant — and the Open CPX On-Ramp

Yesterday’s Vera Rubin NVL72 and MI355X 512-GPU notes were about what peer-reviewed trays and Ethernet pods can print on MLPerf Inference v6.1. This morning is the cabling problem those trays create when you push the scale-up domain past copper: how many fibers, connectors, and shuffle boxes does a 512-GPU optical pod actually need, and what happens if you put transmit and receive on the same strand.

On September 17, 2026, Lightmatter said it had joined the Open CPX Multi-Source Agreement and introduced Passage L20 CPX — billed as the first Open CPX module for AI infrastructure with bidirectional (BiDi) optics on its near-package optics (NPO) platform. Primary sources: the Lightmatter press release, the Sep 14 tech brief One Fiber, Two Directions, the Passage L20 product page, and the PR Newswire mirror. Specs in the release are preliminary; fiber and CapEx figures are vendor modeling under stated conditions.

Spec sheet (named numbers)

CPX here means a standardized CPO/NPO module that sits on the PCB or substrate next to the host ASIC, switch, or GPU — not a faceplate pluggable fighting the same real estate as power and cooling. That is the same “optics next to the package” pressure my NVHBM and d-Matrix Raptor notes keep circling from the memory and decode sides.

What BiDi actually changes in the harness

A conventional UniDi Direct Reach link uses one fiber for TX and another for RX at the same wavelength. BiDi puts TX and RX on two wavelengths in one fiber, combined with an interleaver. Lightmatter’s grid uses λ-A / λ-B at 1311 / 1331 nm. Endpoints are the same SKU; ports are programmed so a Group-A transmitter mates with a Group-B receiver. The company argues that does not cut switch radix or force two OE SKUs — it is a fiber-plant and OE change, not a topology rewrite.

For a 6.4 Tbps OE, that means 64 UniDi fibers (32 TX + 32 RX) become 32 BiDi fibers. Scale that to the brief’s single-tier, all-to-all 512-GPU reference pod and the headline arithmetic is blunt: ~131k → ~65k data fibers, ~32.8k → ~16.4k connectors, ~400 → ~200 fiber miles, and a modeled ~15% cut in total scale-up network CapEx (switch + fiber + connectors + shuffle). Lightmatter cites Broadcom’s earlier CPO BiDi white paper for an independent ~15% optics-cost savings band at the same pod size. Dell’Oro’s Sameh Boujelbene, quoted in the PR, frames the same problem as power, cabling, and time-to-bring-up — not FLOPS.

I care about three operational consequences more than the marketing label “first BiDi CPX”:

1. Connector labor. The brief pegs UniDi clean/test at roughly 1,000+ hours for one 512-GPU pod at ~2 minutes per coupler. Halving connectors is a calendar item, not a slide flourish.
2. Failure surface. Every mated connector and shuffle hop is a contamination and misroute risk. Fewer of them is the reliability argument that survives a skeptical capacity review.
3. Faceplate wall. A 51.2 Tb/s CPO scale-up switch already eats 16 ELSFP lasers and 512 UniDi fibers on a 1RU faceplate in their model; a 204.8 Tb/s tray balloons to 64 lasers and 2,048 fibers — faceplate and fiber plant before switch silicon. BiDi does not invent NPO, but it attacks the fiber-count half of that wall with named arithmetic.

How this sits next to Rubin, Helios, and NVHBM

Vera Rubin NVL72 answers “what does a dense NVLink tray print on DeepSeek-R1 and Qwen3-VL.” Helios UALoE answers “what does open Ethernet scale-up look like when the UALink silicon is not native yet.” NVHBM answers “where does the memory controller live when Fusion partners share the stack.” Passage L20 CPX answers a different ticket: once you have decided the scale-up domain is optical, what does the passive plant cost and how many things can break before the GPUs ever compute.

That is also why this is not a substitute for Qualcomm × Amazon’s 1.6T optical story. That piece is a hyperscaler ASIC + optics procurement line. This one is a merchant Open CPX module and a fiber-count model you can put next to a 512-GPU pod BOM — closer in spirit to asking whether your MI355X RoCE fabric health belongs on the same slide as HBM generation.

Open CPX MSA membership matters for the same reason NVLink Fusion and UALink rhetoric matter: multi-vendor mechanical and electrical conventions are what keep NPO from becoming a one-vendor mezzanine. Lightmatter’s PR is explicit that CPX is the “on-ramp” while full co-packaged optics (CPO) supply chains mature. Treat that as roadmap language until EVKs and third-party switch/XPU boards show up.

Caveats I would write into any BiDi / CPX capacity ticket

1. Vendor model, not a measured pod. Fiber miles, connector counts, CapEx −15%, and labor hours are Lightmatter (and cited Broadcom) models for a stated 512-GPU topology. Re-run with your radix, rail count, and shuffle design.
2. Preliminary specs. The Sep 17 release says specifications are subject to change; EVKs are Q1 2027, not shipping production pods.
3. Insertion-loss adder. BiDi needs an interleaver; Lightmatter says the loss is in the Passage L20 link budget for CPO/NPO with Guide DR or ELSFP — verify against your optical budget, do not assume free lunch.
4. Wavelength pairing ops. Complementary λ-A / λ-B port programming is an ops procedure. Mis-color a link and you debug “dark fiber” that is actually a configuration bug.
5. Not a substitute for switch silicon or GPU generation. This does not change MLPerf rows on Rubin or MI355X; it changes what it costs to wire the domain those rows assume.
6. NPO ≠ CPO ≠ pluggable. CPX/NPO sits near the package; pluggables stay faceplate-bound; full CPO is a different packaging commitment. Keep the three columns separate in any BOM.
7. Energy claims need your own meter. Product-page ~5 pJ/bit and “88% smaller by volume vs OSFP” are vendor claims — useful for screening, not for a signed PUE model.

How I would use this tomorrow morning

If you are designing or quoting a multi-rack optical scale-up domain, the actionable trio is: (a) put the UniDi vs BiDi fiber/connector table next to your actual GPU-to-switch radix before you freeze a harness SKU, (b) ask which Open CPX Type 2 partners are on the EVK list for Q1 2027 rather than treating the MSA join as availability, and (c) keep this column beside NVHBM, Helios UALoE, and Raptor on NVLink Fusion — memory path, open Ethernet bridge, decode silicon, and fiber plant are four different failure domains. If your facility is still copper-bound inside one or two racks, this note is early; if you are already counting MPO/SN-MT16 endpoints by the tens of thousands, the BiDi arithmetic is the part worth auditing first.

Create-only. Specs above are from Lightmatter’s Sep 17 Open CPX / Passage L20 CPX announcement, the Sep 14 BiDi scale-up tech brief, and the Passage L20 product page. Update when EVKs ship or when a third-party switch/XPU board publishes a measured UniDi vs BiDi harness.

Sources: Lightmatter PR · BiDi tech brief · Passage L20 · PR Newswire.

Anpoo Sivanadi, Staff Software Engineer