Certification & routing runtime
The runtime that certifies Vector’s output and guarantees a dense fallback.
What it is
Relay wraps Vector with a recall-safe verifier, exact-k routing, and a guaranteed dense fallback. Its certified output is k tuples — not the dense N-spectrum — so it also cuts the bytes that cross the PCIe interconnect on discrete GPUs (and the unified-memory path on GB10). That data-movement win applies to any downstream consumer that can operate on certified sparse tuples instead of the full dense spectrum. Callable via a C ABI + Python SDK.
How it works
Certify, route, fall back
Relay is the runtime around Vector. It decides when the sparse path is trustworthy, certifies the output against dense recall, and routes to the dense baseline whenever certification fails.
Recall-safe verifier
Confirms the sparse output matches dense top-k recall on the same seeds and the same top-k target before it is trusted. Certification is the gate, not an afterthought.
Exact-k routing
Dispatches each call to the sparse path only where it is measured iso-correct, and routes everything else to the dense baseline.
Guaranteed dense fallback
When certification fails — high dynamic range, large k, out-of-envelope — Relay falls back to dense cuFFT, which stays recall-authoritative. No silent overclaim.
Sparse-output data movement
Because the certified payload is k tuples rather than the dense N-spectrum, sparse-aware consumers move far less data across the interconnect. The benefit requires sparse-tuple-compatible consumers.
The data-movement win
MeasuredBecause the certified payload is k tuples instead of the dense N-spectrum, sparse-aware consumers move far less data across the interconnect.
Routing policy
Where Relay routes sparse, and where it falls back
Relay routes to the sparse path only where it is measured iso-correct. Everywhere else it falls back to dense cuFFT, which stays recall-authoritative.
| Regime | Status | Path | Claim |
|---|---|---|---|
| Clean / on-grid, k ≤ 300, N ≤ 4M | supported | sparse | measured iso-correct |
| Noisy AWGN, bounded dynamic range | supported when gates pass | sparse + certification | review gate: recall ≥ 0.97, precision ≥ 0.99, no p95 regression |
| Off-grid (0.25 bin) | supported | sparse | measured, bounded |
| High dynamic range | dense-better | fallback | no sparse claim |
| k ≥ 500 | dense-better / open | dense fallback | no sparse claim (GPU device-resident path) |
| N = 8M, k = 300 | open | fallback or caveat | not headlined |
Application fit
The broader sparse-output pattern
Relay generalizes a runtime pattern: emit certified k-sized candidate tuples when the sparse path is valid, and route to the dense path when certification fails. Many edge workloads process large sensor, signal, telemetry, or memory-state streams where the useful output is a small candidate set. These are application-fit examples, not measured product claims.
AI-RAN / 5G / 6G
Sparse candidates for sync, random access, pilots, interference, beam and sensing, and telemetry.
Robotics / sensors
Sparse events and regions of interest from high-rate sensor streams.
Industrial edge
Fault frequencies, harmonics, anomalies, and equipment-health events.
AI runtimes
Sparse candidate selection: active pages, MoE router top-k (which experts fire), and retrieval candidates.
Guardrails
Not claimed
- Not a universal cuFFT replacement.
- Not a dense-spectrum FFT win.
- Not controller-side / PIM hardware validation.
- Not a high-dynamic-range sparse win.
- Not a k ≥ 500 sparse win on the GPU device-resident path (the CPU sqrtn lane does win at k ≥ 500).
- Not real-data precision 1.0; real-data precision is signal-bound.
- Benefit requires sparse-tuple-compatible consumers; not every downstream consumer benefits.
- The claim is sparse-output path materialization / transfer reduction, not that the original signal never exists.
- Dense fallback stays authoritative when certification fails.
Relay certifies it. SparseDSP Vector computes it.
See the operator layer that produces the certified sparse output Relay verifies — or request an evaluation packet.