NPO vs CPO Explained: It's Not Distance, It's Board Assembly vs Chip Packaging

Bottom line: A lot of people treat NPO and CPO as the same idea with the optics parked a little closer or a little farther from the chip. They aren't. The real difference is the level of manufacturing: NPO is still board-level system assembly, while CPO crosses into semiconductor packaging. Make that jump and the electrical path drops from centimeters to millimeters, and power, reliability, serviceability and the supply chain all change with it.

This piece expands a post from our Facebook page: NPO vs CPO: it's not distance, it's a different league

What happened

• NVIDIA's Spectrum-X Photonics CPO switch, built on TSMC's COUPE packaging, is shipping to select partners with a ramp in the second half of 2026. NVIDIA claims about 3.5x better power efficiency than pluggable optics.
• TrendForce (July 2026) says Broadcom's 51.2T Bailly CPO switch is in volume production and names optical engine yield and advanced packaging capacity as the main bottlenecks.
• Meanwhile Broadcom is also pitching a VCSEL-based NPO at roughly 1 pJ/bit that stays separately replaceable, so NPO isn't going away quietly.
• Previously: we broke down the CPO receiver side in CPO Germanium Photodetectors Explained.

The cheat sheet:
NPO: optical engine and ASIC packaged separately → same PCB → about 1 to 10 cm of copper → board assembly, field-replaceable.
CPO: optical engine and ASIC in one package (substrate or interposer) → microbumps / TSVs → millimeter-scale electrical path → chip packaging, hard to repair.

Engineering point 1: what really gets shorter is the electrical leg

Fiber can run hundreds of meters without complaint. The pain is the high-speed copper before light turns back into electrons. At 200G-class SerDes rates, every extra centimeter of PCB adds loss, so designers bolt on DSPs and retimers to clean it up. Industry write-ups put NPO channel loss under about 13 dB; NVIDIA says CPO cuts electrical loss to roughly 4 dB. Those few dB decide whether the power-hungry DSP stays on the board.

Engineering point 2: count the savings in pJ/bit

The commonly quoted ranges are 15 to 20 pJ/bit for pluggable modules and 5 to 10 pJ/bit for CPO. Our math (not official): the optics on a 102.4 Tb/s switch at 15 pJ/bit draw about 1,536 W; at 5 pJ/bit, about 512 W. That's roughly a kilowatt per box, and megawatts across a large AI cluster. The original post puts NPO savings at 15 to 25% and CPO at 30 to 50%; treat those as ballpark ranges that shift by generation and architecture.

Engineering point 3: assembly yield and packaging yield are different games

An NPO optical engine can be tested on its own and swapped on its own. CPO needs every engine to be known-good before it goes into the package, because once it's sealed in, one bad engine or one bad row of microbumps can scrap a very expensive package. It's also why the laser stays outside as an external laser source: the part most likely to wear out never goes into the package.

Engineering point 4: CPO competes with AI chips for the same packaging lines

Once CPO depends on 2.5D/3D platforms like COUPE, it competes with GPUs for the same advanced packaging capacity, which TrendForce lists as a top bottleneck. Heat is the other fight: optical parts dislike heat, and they now sit right next to a switch ASIC drawing hundreds of watts, which is why the Spectrum-X Photonics chassis are liquid-cooled. That's also why TSMC matters here. It isn't just fabbing the switch die; it's the one sealing light and electronics into the same package.

Open questions

• There is no agreed public pJ/bit figure for NPO as a category. Broadcom's ~1 pJ/bit VCSEL NPO is a vendor claim for mostly short-reach multimode links, so it isn't apples to apples with single-mode silicon photonics CPO.
• NVIDIA's 3.5x is a network-level figure that includes removing host-side retimers, not just the optics.
• The 1,536 W and 512 W figures are our arithmetic from industry ranges, not measurements.
• How CPO gets serviced in the field (swap the switch or the board?) and its long-term reliability still need large-scale deployment data.

Related videos

Silicon Photonics explained

TSMC CoWoS S, R and L explained

Is NPO just a bridge until CPO matures, or will the two live side by side? If you ran a data center, which would you pick? Tell us in the comments.

Sources: NVIDIA Technical Blog, "Scaling AI Factories with Co-Packaged Optics"; NVIDIA press release, Mar 18 2025; TrendForce, Jul 27 2026; Taiwan News, Jun 4 2026; Broadcom VCSEL NPO article; original Facebook post. Figures marked as our math are not official.
Videos by @TheEngineeringCore-v. Published on The Automation Core.
Original post: https://www.facebook.com/1147074221177116/posts/1156573600227178
This one sparked a lot of discussion on our Facebook page. Come join the conversation: PLC on Facebook