Wired For Success: Enhancing AI Cluster Reliability With Copper Cables

Author: Amit Suri, Credo Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Credo presents the case for active electrical copper cables (AECs) as a reliability-critical alternative to optics for the first link between GPUs and leaf switches in AI clusters. As GPU clusters scale from 100K to 3 million GPUs, the mean time between link flaps across the network drops to minutes or seconds, making link reliability a first-order concern. The backend (scale-out) network uses UDP-based RDMA, which unlike TCP has no packet retransmission capability -- a single link flap can bring down the entire distributed training job, wasting thousands of dollars in GPU time.

Credo's AEC technology places retimers inside both ends of a passive copper cable, breaking the link into three manageable segments. This extends copper reach to 7 meters at 800G (current generation) while maintaining an MTBF of 100 million hours -- 10x that of optics at approximately 10 million hours. Credo has deployed approximately 2 million AEC cables with 6 billion cumulative hours of operation without a link flap. The company's fully integrated stack (DSP, SerDes, firmware, and manufacturing all in-house) enables interoperability testing in their Taiwan system lab.

For the upcoming Rubin (NVL144) architecture at 150 kW/rack, Credo introduces 1.6T AEC cables reaching 6 meters (5 meters with 32-gauge Cat-6-like cable). The NVL144 connects 5 racks (2 GPU racks per side, 1 network rack center) with 1.6T/5m AECs. For the Rubin Ultra/Kyber (NVL576) at 600 kW/rack with 576 GPUs, 100% liquid-cooled, the cable reach shrinks to 1.5-2 meters, enabling even thinner Cat-6-like cables. The AEC power at 1.6T is 13W (half of optics), and at 800G (4x200G lanes) is 8W. GPU densification and liquid cooling work in copper's favor by progressively shortening the required reach.