OCP Global Summit 2025 | Credo & Oracle | The Path to Zero Flap: Reinventing Optical Reliability for Scalable AI Clusters
Updated: 23 hours ago
Introduction
AI data centers are entering the era of “megawatt-scale clusters.” As a single cluster grows from 3MW to 30MW, then 300MW, 1GW, and even 10GW, network stability becomes a lifeline.
In these hyperscale environments, link flaps have become the biggest threat to scaling the AI fabric.
At OCP 2025, Credo and Oracle Cloud Infrastructure (OCI) jointly announced the “Zero Flap” initiative, which aims to use smart optical modules, real-time telemetry, and distributed monitoring
to eliminate link flaps entirely and redefine optical connectivity reliability for AI clusters.
Content
1. The root cause: L1 stability is the bottleneck for every layer
OCI networking engineering lead Steve Manley opened bluntly:
“Link flaps are not a small problem. They make the whole fabric stall and rebuild, and can even crash training jobs running on tens of thousands of GPUs.”
In an AI cluster, an unexpected drop on a single optical link can trigger:
A full reset across L1–L3;
IGP re-convergence;
RDMA job instability and checkpoint thrashing;
Every flap can waste tens of thousands of GPU-hours of compute.
Manley stressed: “If L1 isn’t stable, the stability of every protocol above it is an illusion.”
2. Why link flaps are so hard to cure
Over the past 18 months, Oracle’s engineering team dug into the causes of link flaps and found it is a “systemic problem” rather than a single component defect. The main causes include:
Optical transceiver defects: laser aging, PD sensitivity drift;
Connector and fiber contamination: dust, loose connectors, MPO misalignment;
Environmental factors: construction dust and thermal cycling causing optical module performance drift;
ESD damage: accidental contact by installers causing latent, delayed failures;
Firmware and DSP reliability differences: interoperability issues between optical modules and switches.
These issues can be tolerated in small environments, but when a system scales to millions of links and tens of thousands of nodes,
every flap gets amplified into network-wide congestion and training delays, seriously hurting compute performance and revenue.
3. The limits of traditional fixes
Oracle’s current stopgap is “Accelerated Fabric Grooming”:
When a potentially unstable link is detected, shut that port down;
Trade “sacrificed bandwidth” for stability;
Predict anomalies through energy analysis, error distributions, histograms, and pattern matching;
And collect more data at the host layer to aid diagnosis.
This works, but it’s expensive. As Manley put it: “I’d rather have a few fewer links than a fabric that flaps.”
He showed a real example: a “bucket of optics” in a corner of the data center,
filled with removed, expensive 800G modules: “These should be lighting up in a switch, not gathering dust in a bucket.”
4. The turning point: from centralized monitoring to distributed intelligence
Manley proposed a fundamental shift in perspective:
“The reason we made the internet succeed was that it was distributed. So why are optical networks still relying on centralized monitoring?”
He argued for making the optical module itself an intelligent node, with the module’s built-in DSP performing health monitoring and decision-making directly.
The optical module knows:
Its own temperature, power, SNR, and BER;
The serial number and status of the module at the other end;
And can proactively raise a “check engine light” style alert.
This is the core idea behind “Zero Flap Optics”: give every module the ability to self-diagnose and raise alarms.
5. Credo’s solution: Zero Flap optical modules
Credo CEO Bill Brennan then walked through the technical details.
The Zero Flap module is a smart optical module platform co-developed with Oracle, with three core capabilities:
(1) Hardened optics
Repeatedly broken, fixed, and retested under thermal cycling and environmental stress testing;
Reliability 2–3 orders of magnitude higher than conventional modules;
Applicable to 800G / 1.6T modules.
(2) Built-in telemetry system
Provides real-time eye height, SNR, pre-FEC BER, error histograms, power drift, and more;
Flags potentially degrading links in real time, like an “optical-layer health map”;
Adds ESD damage detection to predict latent, delayed failures.
(3) In-band messaging & self-isolation
Modules can communicate directly and exchange health information;
When a threshold is triggered, unstable links can be isolated automatically without waiting for an external control system;
Status can also be uploaded to a centralized management platform (Zero Flap Agent).
Credo emphasized: “This isn’t just a better module; it’s a module that thinks.”
6. Open standardization: the Optical Reliability Project
Credo and Oracle have submitted the Zero Flap Host Spec to OCP,
formally launching the Optical Reliability Project,
with the goal of establishing a cross-vendor, interoperable standard for optical-layer telemetry and self-diagnosis.
The project will include:
Open host drivers and APIs;
Joint validation with multiple switch vendors (including Broadcom, Cisco, and Arista);
Driving the industry toward a “Zero Flap Ready” certification.
Manley concluded:
“800G is repeating the mistakes of 400G. 1.6T will be even harder. This is where we should draw a new line.”
Conclusion
“Zero Flap” isn’t just a Credo–Oracle collaboration; it’s a turning point for AI data center reliability.
When the AI fabric spans tens of thousands of nodes and millions of links,
only by giving optical modules real-time self-awareness, distributed decision-making, and open telemetry,
can we truly achieve “AI Ready Optical Infrastructure.”
Data center stability will shift from electrical to optical intelligence.
Extended perspectives
Technical takeaways
Zero Flap modules mark the start of an intelligent optical layer; optical modules will evolve toward being “self-healing, distributed-control, and software-defined.”
The DSP’s role in the optical layer will expand from signal processing to an “edge AI monitoring core.”
Industry implications
This standardization move may become the reliability watershed for the 800G / 1.6T generation.
By entering through an “optical module telemetry platform,” Credo could have a disruptive impact on traditional optical module suppliers (such as Coherent, Lumentum, and Innolight).
Supply chain trends
Once Oracle integrates the Zero Flap mechanism into the OCI network architecture, other hyperscalers (AWS, Google, Meta) are bound to follow.
Future optical module specs may require not only power, BER, and temperature range, but also add “Telemetry Capability” and “Self-Isolation Behavior” fields.
Comments