top of page

📢 STT 訂閱專區已上線

免費文章會照常更新,一篇都不會少。訂閱是「加強版」——每週深度週評、財報法說的完整判讀、所有長篇深度報告全包。

免費讓你跟上,訂閱讓你看懂、能做判斷。

月訂 NT$199|年訂 NT$2,000(約 NT$167/月)
👉 立即訂閱: vocus.cc/salon/simpletechtrend

OCP Global Summit 2025 | Lotes | Copper Interconnect in 200G Scale-Up Networks

3 days ago
4 min read

Introduction

In an era when optical interconnect increasingly dominates AI infrastructure discussions, Lotes used its OCP 2025 talk to return to the electrical core: copper interconnect.

Speaker Raymond Law stressed that he is "not an optics person, but a signal integrity engineer", and in a conference full of optical module demos, he wanted to remind everyone:

"In the 200G era, copper still plays a critical role, and in some scenarios it remains the most reliable choice."

The talk can be read as a rational counterpoint to the recent wave of CPO (Co-Packaged Optics) and LPO (Linear-Drive Pluggable Optics). Starting from real cable lengths, loss budgets and system design, it explains why copper has not left the stage.


Content

1. Structure and Challenges of Scale-Up Networks

Lotes first defined the essence of a scale-up network:

It lets multiple GPU accelerators operate with minimal latency so that, from the outside, the whole system looks like one "super GPU".

Such links must offer:

  • High-bandwidth, low-latency accelerator-to-accelerator transport;

  • Reliable connections over short distances (1–10 meters);

  • Low power and high reliability to support long-running AI training jobs.

He noted that today's 200G systems run at about 212.5 Gbps per lane (OIF 224G spec), which already pushes channel design and cable quality very close to physical limits.


2. Why Can Copper Survive at 200G?

Lotes cited research from OCP PCIe Extended Connectivity:

In experiments, a 1-meter passive copper cable had an insertion loss of about 16.3 dB and still met the spec.

This means:

  • With current technology, uplinks and downlinks within a rack can still be done in pure copper;

  • With good design, 1–2 meter links can cover an entire GPU rack.

In other words, copper remains competitive at scale-up (in-rack) distances.


3. Co-Packaged Copper (CPC) and Near-Packaged Copper (NPC)

To extend reach and improve signal integrity, Lotes proposed two package-level architectures:

  1. CPC (Co-Packaged Copper):

    • Places the connector directly next to the ASIC package;

    • Can use "flyover cables" to reduce trace loss.

  2. NPC (Near-Packaged Copper):

    • Routes cables on an interposer around the chip;

    • Allows longer external link distances (1.4 meters or more).

These design concepts are ready for volume production, and physical samples were on display at the Lotes booth.

Raymond emphasized: "CPC and NPC are copper life-extension solutions until optics takes over."


4. Bringing in Active Components: Repeaters, Retimers, Redrivers

Beyond 1.5 meters, active compensation components are required.

Lotes outlined three typical solutions:

  • Redriver: boosts the signal; low cost but requires manual tuning.

  • Retimer: re-times the signal; stable but power-hungry and expensive.

  • Repeater chain: compensates in segments along long copper runs, for links between two racks (2–7 meters).

These designs allow copper to reach 2 meters of active reach (Active Copper Cable, ACC) at 200G–224G.

Raymond added: "With today's technology, 2-meter ACC is not a problem and is already in volume production."


5. Rack Implementation and Cabling Challenges

For real OCP rack environments, he recommended:

  • Pure copper within a single rack (CPC/NPC + redriver);

  • ACC (Active Copper Cable) between adjacent racks;

  • Beyond two racks (>7 meters), switch to optical modules.

He strongly urged the industry to avoid routing rack cabling through ceilings or conduits, which needlessly stretches distances beyond 10 meters and adds cost and power.


6. Where Copper Ends and Optics Begins

At the end of the talk, Raymond posed a core question:

"When exactly should we move from electrical to optical?"

His view:

  • Copper still has three major advantages: low cost, low power and high reliability;

  • But optical solutions (CPO/NPO) will gradually replace copper as bandwidth density and power constraints tighten;

  • The ideal future is a hybrid interconnect architecture that picks the best medium by distance and application.

He called on the OCP community to create "design guidelines for the electrical-to-optical boundary" to help vendors decide when to bring in optics.


Conclusion

Lotes' message at OCP 2025 was clear and pragmatic:

"In the 200G era, copper is not obsolete; it is the practical foundation of high-performance interconnect."

Within racks, between adjacent servers, or at the GPU module level,

copper, with NPC/CPC and redriver technology, still offers a stable and cost-effective solution.

It reminds the industry that while CPO and optical packaging are the future, "until that future arrives, copper is still the most reliable answer today".


Further Perspectives

  1. Technology impact

    • Lotes' talk can be seen as a complement to the Broadcom EON / Ethernet Scale-Up architecture, pointing out from a practical standpoint that "the short-reach layer still belongs to copper".

    • The maturity of CPC/NPC shows that the transition toward "in-package electro-optical hybridization" is already taking concrete shape.

  2. Supply chain observations

    • As a global supplier of connectors and high-speed copper cables, Lotes is actively building out AI rack-level product lines.

    • It forms a complementary triangle with Amphenol, Samtec and Credo: the others focus on standards and packaging, while Lotes is strong in system cabling.

  3. Market trends

    • As 200G–400G technology goes mainstream, the copper/optics boundary is moving down from the switch layer to the GPU module layer.

    • Future data centers will form a hybrid interconnect fabric, switching media based on power, distance and latency.

Recent Posts

See All

Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page