Capability Matrix
This document enumerates the truth table of claims each Denpex product tier can and cannot make. It exists so that every marketing claim can be traced to either a passing test or to a documented gap.
Denpex Lite (denpex_lite)
Can Claim
- Crash-log diagnosis from wrapped stdout/stderr.
- First failed rank detection from logs.
- NCCL timeout interpretation (string match).
- CUDA exception diagnosis.
- Timeline reconstruction from process logs.
- Framework-aware recommendations (PyTorch, DeepSpeed, Megatron, FSDP, JAX, TF).
- AI fallback summary (over textual evidence only).
- No-root deployment (pure Python, stdlib-only wrapper mode).
- Stdlib-only Python wrapper mode.
Cannot Claim Alone
- Silent Data Corruption detection (requires canary microbatch via SDK + agent).
- Hardware degradation detection.
- PCIe diagnosis.
- NVLink degradation detection.
- ECC correlation across ranks.
- Predictive hardware failure.
- Per-layer numeric anomaly detection.
- Gray failure detection from hardware signals.
Full Denpex Platform
Inherits all Lite claims, plus
- Hardware degradation detection (SBE / DBE / row-remap / etc.).
- PCIe and NVLink diagnosis.
- ECC correlation across ranks.
- Predictive hardware failure (Weibull hazard, peer-relative z-score).
- Per-layer numeric anomaly (weight / grad / activation L2, NaN/Inf, checksum drift, optimizer asymmetry).
- Gray failure detection from hardware signals (z-score, throttle count, etc.).
- Topology-aware root cause (Ring / Tree / CollNet).
- Automated remediation on a per-policy basis.
- Air-gapped deployment with HSM-rooted KMS.
Cannot Claim
- Worldwide low-latency SLA. The SaaS control plane is single-region primary with globally-replicated reads; worldwide multi-region active-active is not available today (roadmap, not a tier entitlement). Per-region low latency requires a self-hosted / on-prem per-region deployment with a pod in each region.
- FedRAMP authorization (requires Growth or higher).
- Custom on-prem hardware brands beyond the supported matrix.
Tier Comparison Table
| Capability | Free | Team | Scale | Growth | Air-gap |
|---|---|---|---|---|---|
| GPU / Fabric collectors | ❌ | ❌ | ✅ | ✅ | ✅ |
| RAG (sentence-transformer) | ❌ | ❌ | ❌ | ✅ | ✅ |
| K8s / SLURM real remediator | ❌ | ❌ | ✅ | ✅ | ✅ |
| Self-hosted CP | ❌ | ❌ | ✅ | ✅ | ✅ |
| HSM-backed KMS | ❌ | ❌ | ❌ | ✅ | ✅ |
| FedRAMP roadmap | ❌ | ❌ | ❌ | ✅ | ✅ |
| Custom on-prem hardware | ❌ | ❌ | ❌ | ❌ | ✅ |
| Air-gap install | ❌ | ❌ | ❌ | ❌ | ✅ |
| Multi-tenant CP | ❌ | ❌ | ✅ | ✅ | ✅ |
| Audit log retention | 7 d | 30 d | 90 d | 1 y | indefinite |
What is NEVER Claimed (any tier)
- "Detection latency under 1 ms." Unless measured on the customer's site with a published reproducer, we don't make this claim.
- "Zero overhead." Measured overhead is bounded at < 1.5 % throughput delta on a 64-rank job (
docs/agent-overhead-benchmarks.md). - "100 % SDC detection." Per
docs/silent-corruption-detection-limits.md, we are explicit about the limits. - "Compatible with anything not on the matrix." Any not-on-matrix configuration requires a fresh test.
Audit Trail for Claims
Each row above is reproducible from a passing test or benchmark. See the corresponding docs/<topic>.md for the provenance.