Co-packaged optics (CPO)

Co-packaged optics (CPO) is a system architecture where optical transmitter and receiver components are integrated directly onto the same substrate package as the host networking ASIC (typically a switch chip or AI accelerator). This contrasts with the dominant pluggable architecture, where the ASIC drives a SerDes electrical interface across a printed circuit board to a separate pluggable optical module.

Why CPO matters. As datacom data rates scale beyond 800G per port, the electrical interface between switch ASIC and pluggable becomes the dominant power consumer. A typical 800G QSFP-DD pluggable in 2024 dissipates $\sim 15$ W; half of that is the electrical-side power (SerDes drivers, retimers) needed to push the signal across the board to the pluggable. As speeds rise to 1.6 Tb/s and 3.2 Tb/s, the electrical interface power scales superlinearly with speed (due to the need for higher-order modulation, more aggressive equalization, and more retimer chips), while pluggable thermal envelopes cannot grow proportionally.

CPO eliminates the long electrical trace by placing the optics directly next to the ASIC. The estimated power saving at 800G+ is 30 – 50 % vs the equivalent pluggable architecture — a critical lever in AI training cluster scaling, where networking power consumption is a substantial fraction of total facility power.

Standard architecture. A CPO module typically includes:

• Host ASIC at center (switch chip or accelerator), $\sim$ 800 mm² die
• 8 – 16 optical engines placed around the periphery of the host ASIC, each containing a silicon photonic chip with multiple lasers, modulators, photodetectors
• Substrate or bridge providing high-bandwidth electrical signaling between optics and host
• Common laser-light source (sometimes external, sometimes integrated) delivering CW laser light to each optical engine via fiber
• Fiber array exiting the package, typically 32 – 256 fibers per package

Industry status (2025-2026). CPO is in the "early production" phase:

• Broadcom announced CPO Tomahawk switches in 2024 – 2025 (51.2T total bandwidth) using silicon photonic optical engines
• NVIDIA uses CPO-like designs in its NVLink fabric for large AI clusters (Hopper and Blackwell generations)
• Marvell and Cisco / Acacia have announced CPO product roadmaps
• Intel has demonstrated CPO prototypes but not yet productized at scale
• Open Compute Project (OCP) has a CPO working group defining standardized interfaces

Key technical challenges.

1. Optical alignment in manufacturing: bonding the fiber array to the silicon photonic engine with sub-μm accuracy at high yield. Active alignment is too slow for CPO production volumes; passive alignment with precision micro-features is the target.
2. Laser placement: lasers are heat-sensitive; placing them next to a 100+ W ASIC creates thermal challenges. Remote laser sources (off-package laser modules feeding the CPO via fiber) are common in production designs.
3. Field replaceability: pluggable modules fail and get hot-swapped in seconds; CPO failure modes require ASIC replacement, dramatically increasing the cost of any single fault. Robust optical engine design with redundancy is essential.
4. Standardization: CPO breaks the pluggable interoperability model. Industry is still negotiating how to specify optical I/O at the package level.

Comparison to pluggable.

Metric	Pluggable (OSFP-DD)	Co-packaged
Electrical reach	$\sim 15$ cm	$\sim 1$ cm
SerDes power (per 100G)	$\sim 5$ pJ/bit	$\sim 2$ pJ/bit
Field-replaceable?	Yes	No (package-level swap)
Interoperability	Vendor-neutral via MSA	Vendor-specific
Per-port cost (2025)	$1000 – $3000 for 800G	similar; cost not yet differentiating
Volume scaling potential	mature, declining cost	early; significant headroom for cost reduction

Light source options. Two architectural choices:

• Lasers on the optical engine — simpler signal path but bigger thermal challenge
• Remote laser source with CW light delivered to the optical engine via fiber — better thermal management, but adds an external module that itself fails and must be hot-swappable

The remote-light approach dominates in 2025 – 2026 production designs because the laser is the most failure-prone component in any photonic system.

Long-term outlook. CPO is widely expected to displace pluggable for the highest-end datacom and AI fabric applications by 2027 – 2029. Pluggable will continue to dominate metro/access/lower-end datacom for the foreseeable future due to operational flexibility advantages.

References: Lightwave magazine special issue on CPO (2024); OCP CPO project documentation; Broadcom Tomahawk 5 CPO whitepapers.