Manufacturing.ai

Topicsco-packaged-optics

Co-Packaged Optics: Moving Light Into the Package

A sourced reference collection on co-packaged optics — why AI networking is pulling optical conversion inside the chip package, what the power savings actually measure, and which problems still stand between sampling and deployment.

Assertions
7
Sources consulted
11
Read in full
4/11
Cited as evidence
4

11 sources sit behind this page — including any that arrive with a concept this page shares with another collection. 4 were retrieved and read in full, and only those can back an assertion. 4 could not be retrieved, and 3 were surfaced and deliberately set aside. Every one of them is named in the register below, with the reason in view. How we source this.

Timeline newest first · evenly spaced, not to scale

NOWEARLIEROct 8, 2025Broadcom Announces Tomahawk 6 – Davisson, a 102.4 Tbps Co-Packaged Optics Switch
Background

Our own synthesis, written to orient you — not evidence. Every factual statement here is asserted and sourced further down this page.

Every optical link in a data centre converts electricity into light and back again, and for twenty years that conversion happened in a module you could pull out of the front of the switch with your fingers. Co-packaged optics moves it inside the package, next to the switch silicon. The reason is power: an independent engineering account puts co-packaged optics at roughly 5.5 watts per 800 Gb/s port against about 15 for equivalent pluggables — around threefold, or 6 to 7 picojoules per bit. At the port counts modern switches carry, that difference stops being an efficiency argument and becomes a thermal one.

In 2026 this stopped being a sampling technology. Broadcom announced Tomahawk 6 – Davisson on 8 October 2025, a third-generation co-packaged Ethernet switch at 102.4 Tbps and 200 Gbps per channel using TSMC COUPE optical engines. NVIDIA states two lines of its own: a Quantum-X InfiniBand Photonics switch at 144 ports of 800 Gb/s, which it says connects over 10,000 GPUs in a two-level fat-tree, and Spectrum-X Ethernet Photonics at up to 409.6 Tb/s, stated as available in the second half of 2026.

Read the vendor ratios carefully, because this collection does not repeat them. NVIDIA publishes 5x better power efficiency than pluggables, 5x sustained AI application runtime and 1.3x faster time to insight, all without a stated baseline configuration. Those are claims. The independently reported watts-per-port figure above is what this page asserts.

What is given up is the thing that made pluggables popular in the first place. A faceplate module can be swapped in minutes, and an operator can choose the optics per port. Once the optical engine is inside the package, a failure is a package-level failure, and the constraints that remain are argued to be systems-level rather than device-level: standardisation across platforms, serviceability, and thermal-aware photonic-electronic co-design. Those are organisational problems as much as physical ones, which is usually a sign that a technology arrives slower than its performance numbers suggest.

This page is concept-heavy and carries few dated events, and that is a truthful reflection of where the technology sits rather than a gap in the research: announced availability windows are recorded as vendor claims and never promoted into events. What is missing is named — an independent survey of what is actually shipping could not be retrieved, so deployment status here rests on vendor statements, marked as such on every assertion; Broadcom's figures reach this collection through a syndicated reproduction rather than its own site; and the widely-repeated one-million-link-hours reliability figure is deliberately absent, because the release referencing it gives neither test hours nor the testing entity.

Figures

Every number below is asserted and sourced elsewhere on this page.

The power the package saves

Watts per 800 Gb/s port. This is the independently reported comparison, not a vendor ratio — the difference is about threefold, or 6 to 7 picojoules per bit.

Watts per 800 Gb/s port

~15 W

Pluggable optics

~5.5 W

Co-packaged optics

The independent engineering account cited on this page, which gives roughly 5.5 W per 800 Gb/s port for co-packaged optics against about 15 W for equivalent pluggables. Reported measurements, not a vendor comparison; the larger ratios vendors publish carry no stated baseline and are excluded.

Switch bandwidth behind the optics

Stated switch capacity for the two co-packaged platforms this page covers. The second figure is a vendor statement for a product with an announced availability window.

Switch bandwidth, Tb/s

102.4 Tb/s

Tomahawk 6 – Davisson

409.6 Tb/s

Spectrum-X Photonics

Broadcom's Tomahawk 6 – Davisson announcement of 8 October 2025 (102.4 Tbps), reaching this collection through a syndicated reproduction, and NVIDIA's silicon photonics product page (up to 409.6 Tb/s, stated available in the second half of 2026). Both are vendor-stated capacities.

Concepts

The vocabulary this subject is built from, and what we can show about each.

Advanced Packaging

packaging technologyshared from another collection — see its own page for what it asserts

Co-Packaged Optics (CPO)

packaging technology

The binding constraints on CPO are argued to be systems-level rather than device-level: standardisation across platforms, serviceability, and thermal-aware photonic-electronic co-design — matching the operational objection that an embedded optical failure means replacing the whole package rather than swapping a module.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

An independent engineering account puts CPO at roughly 5.5W per 800Gb/s port against about 15W for equivalent pluggables — around threefold, or 6-7 picojoules per bit — while vendors claim larger and less specified improvements: Broadcom 70% and 3.5x, NVIDIA 5x, neither stating a baseline configuration.

ReportedSupported by the sources below, not yet editor-reviewed.
3 sources3 retrieved & read

Co-packaged optics moves electrical-to-optical conversion inside the switch or accelerator package instead of into faceplate pluggable modules, shortening the lossy electrical path — at the cost of per-port configurability, field-replaceable optics, and the need for liquid cold-plate cooling.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

NVIDIA states a Quantum-X InfiniBand Photonics switch at 144 ports of 800 Gb/s connecting over 10,000 GPUs, and Spectrum-X Ethernet Photonics at up to 409.6 Tb/s available in the second half of 2026, both on 200G SerDes with optics co-packaged onto the ASIC and no DSP retimers.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

InfiniBand

interconnectshared from another collection — see its own page for what it asserts

Making Ethernet Fit for AI

interconnectshared from another collection — see its own page for what it asserts

Pluggable Optical Transceiver

component

Pluggable optical transceivers are field-replaceable faceplate modules whose advantages are operational — per-port choice of optics and minutes-long swaps — against a power cost of roughly 15W per 800Gb/s port and a lossy electrical path to the switch silicon.

ReportedSupported by the sources below, not yet editor-reviewed.
1 source1 retrieved & read

Scale-Up Versus Scale-Out

interconnectshared from another collection — see its own page for what it asserts

Silicon Photonics

component

Silicon photonics fabricates optical components in silicon so optical engines can be packaged with logic — Broadcom's third-generation switch uses TSMC COUPE engines via heterogeneous multi-chip packaging — but both Broadcom and NVIDIA keep the lasers outside the package to limit power and improve reliability.

ReportedSupported by the sources below, not yet editor-reviewed.
2 sources2 retrieved & read

Switch Silicon Capacity

componentshared from another collection — see its own page for what it asserts

Timeline

What actually happened, in order, with sources.

Coverage
  • 1 United States

Where this topic’s events took place, as far as our sources establish it. Events with no single location — a standards publication, say — and events we have not yet attributed are both counted as unattributed rather than omitted.

  1. Oct 8, 2025

    Broadcom Announces Tomahawk 6 – Davisson, a 102.4 Tbps Co-Packaged Optics Switch

    technology generation milestoneUnited StatesNorth America

    On 8 October 2025 Broadcom announced Tomahawk 6 – Davisson, a third-generation co-packaged optics Ethernet switch at 102.4 Tbps and 200 Gbps per channel using TSMC COUPE optical engines, claiming a 70% reduction in optical interconnect power; it was sampling to early access customers, not generally available.

    ReportedSupported by the sources below, not yet editor-reviewed.
    1 source1 retrieved & read

Source register

All 11 sources behind this page — what we read, what we tried to read and could not, and what we looked at and set aside, with the reason in view for each. A concept shared with another collection brings its own references with it, so some entries here were surfaced for a neighbouring topic rather than this one.

Cited as evidence
4
Tried, could not read
4
Surfaced, set aside
3
Cited sources 4 distinct links

Original publisher links. Files open on the publisher’s site; we do not host copies. A linked document is not an additional source or an independent verification.

Tried, could not read4

We attempted these and were refused or served nothing. Nothing on this page rests on them; they are published so the gaps are checkable rather than invisible.

Surfaced, set aside3

These came up while researching and were deliberately not used. We do not claim to have read them — each is listed with why it was passed over, so the shape of the survey is visible and not just its conclusions.

Coverage & limits

What this page does and does not claim.

Sixth packet through the generic ingestion pipeline, rebuilt in September 2026 to the reference-collection template and reweighted toward what is shipping now rather than toward the technology's development history. (August 2026 original.) Four sources were directly retrieved and three are recorded as references only: Broadcom's own investor-relations release and EDN's 2026 status survey both timed out repeatedly, and Semiconductor Engineering returned HTTP 403 for the second session running. Because EDN could not be read, the independent survey of what is actually shipping is missing, and deployment status here rests on vendor statements — marked as such on every assertion. Broadcom's figures reach this collection through a syndicated reproduction rather than from Broadcom's own site, which is recorded rather than glossed. The widely-repeated 'one million link hours at Meta' reliability figure is deliberately absent: Broadcom's release references a link flap study without giving test hours or naming the testing entity, so it is not asserted. This collection is concept-heavy and carries a single dated event, which is a truthful reflection of a technology at the sampling stage rather than a gap — announced availability windows are recorded as vendor claims, never promoted into events. Not yet editor-reviewed; every assertion reads as reported.

Source check, 2026-09-17. Numeric-presence checks passed for 7 assertions using available source text, which may be cached. This is not verification of their meaning. What this check does and does not prove →

  • Not editor-reviewed unless labelled. Assertions marked Reported are assembled from the sources shown and have not yet been checked by an editor. Only Primary source and Corroborated mean a human verified them.
  • Disagreements are preserved, not resolved. Where sources conflict, both accounts appear and the assertion is marked Disputed.
  • Retrieval status is disclosed per source. A source we could not open is never counted as evidence for an assertion.

This page is also available as structured data: /api/v1/topics/co-packaged-optics