BEP Research · Interactive Explainer

How Zero-Flap Optics Works

A single flapping link can stall a synchronized AI training step and idle thousands of healthy GPUs. Zero-Flap is engineered, in hardware and software, to stop the flap before it cascades. Press play and watch the difference.

1The flap, and why it idles the cluster

Each square is a GPU; the lines are optical links carrying a synchronized all-reduce — every GPU must finish together before the next training step starts. One link down stalls the whole step. Toggle the optics and watch utilization.

GPU computing GPU stalled (waiting) Link healthy / telemetry Link flapped (down)
100%
Cluster utilization
0
Training steps done
0
Link flaps
0
GPU-seconds idle

2Three layers, one system tap a card

Zero-Flap is a system, not a component. It attacks instability at the three layers Credo can control. That integration is the moat — far more than any single part.

Hardware

Fewer lasers, fewer failure points

Silicon photonics (from the Dust Photonics acquisition) integrates the optics onto a chip and uses substantially fewer lasers per link — the laser is the least reliable part. Fewer lasers means better power, lower cost, and less exposure to the laser supply shortage. The DSP is tuned for link stability, not just raw throughput.

Why it matters →
Software

Catch the flap before it happens

The PILOT software, integrated at the switch via the SDK, continuously reads link-health telemetry — bit-error rates, signal-quality histograms, laser-degradation signals. It is the check-engine light for the optical link: it sees the signals that precede a failure and mitigates the instability before the link drops the cluster. Prevention, not cleanup.

Why it matters →
Integration

Own more of the stack, see more of the link

Because Credo controls both the DSP and the photonic integrated circuit (PIC), it gets tighter DSP-to-PIC integration than a vendor selling one piece. Tighter integration means richer telemetry, deeper diagnostics, and system-level optimization a component supplier cannot match. This hardware integration is the durable moat; the software is the contestable layer.

Why it matters →

3The blast radius: what flaps cost you

Credo's CEO has said one customer was losing 20–30% of its productivity to link flaps before deploying the Zero-Flap approach. Reliability is not a feature here; it is the limiting factor on usable compute. Set your fleet and the flap tax.

Zero-Flap aims to recover most of the lost slice. The point is not that flaps are common; it is that at cluster scale a small per-link probability compounds into a large idle fleet.

75,000
effective usable GPUs
25,000
idled by link flaps
Read the full analysis
Why Zero-Flap is Credo's highest-ASP leg — the mix, the ramp, and the bear case.
Read on BEP Research →
BEP Research · Interactive companion to the Zero-Flap optics explainer. Reliability is the constraint. Illustrative model, not a benchmark
The simulation is a teaching abstraction of all-reduce stalls and link-health mitigation, not a measurement of any product. Disclosure: the author holds CRDO. © 2026 Ben Pouladian.