Positioning
"Denpex is the reliability control plane for AI training infrastructure."
Not
- Not "a Python script that diagnoses GPUs".
- Not "an LLM that hallucinates a root cause."
- Not "a metrics vendor that lists counters without context."
Yes
Denpex is the seven-layer stack that:
- Wraps training (
denpex_lite) — stdlib Python, 397 signatures. - Hosts telemetry — agent + GPU + fabric + storage collectors.
- Instruments frameworks — per-layer hooks + SDC canary.
- Correlates & diagnoses — deterministic rules + RAG.
- Predicts — Welford / Weibull / peer-relative drift.
- Auto-remediates — K8s + SLURM state machine, policy tiers.
- Runs in your tenant — multi-tenant, air-gapped, FIPS, SOC2.
Architecture (7 Layers)
+-------------------------------------------------------------------+
| Layer 7 | Enterprise Plane | SAML, OIDC, RBAC, KMS, Audit |
| Layer 6 | Control Plane | Correlation, diagnosis, RAG |
| Layer 5 | Remediation | State machine, policy, audit |
| Layer 4 | Training SDK | Per-layer metrics, SDC canary |
| Layer 3 | Fabric Collector | IB, RoCE, NCCL, GPUDirect |
| Layer 2 | GPU Collector | NVML, DCGM, CUPTI |
| Layer 1 | Host Agent | eBPF, PSI, journald, NUMA |
| Layer 0 | denpex_lite | stdlib Python wrapper, signatures |
+-------------------------------------------------------------------+
See docs/architecture-whitepaper.md for the full design.
Capability Matrix (Summary)
See docs/capability-matrix.md for the truth table.
| Capability | Denpex Lite | Full Platform |
|---|---|---|
| First-failed-rank (logs) | Yes | Yes |
| NCCL timeout interpretation | Yes | Yes |
| CUDA exception diagnosis | Yes | Yes |
| SDC canary | - | Yes |
| Hardware degradation | - | Yes |
| PCIe / NVLink diagnosis | - | Yes |
| ECC correlation across ranks | - | Yes |
| Predictive hardware failure | - | Yes |
| Per-layer numeric anomaly | - | Yes |
| Gray-failure from hardware | - | Yes |
| Multi-tenant control plane | - | Yes |
| Air-gapped installation | - | Yes (appliance) |
What We Don't Claim
Per docs/capability-matrix.md#what-is-never-claimed-any-tier:
- "Detection latency under 1 ms" without published measurement.
- "Zero overhead" — measured overhead is bounded at < 1.5 % per
docs/agent-overhead-benchmarks.md. - "100 % SDC detection" — limits are documented in
docs/silent-corruption-detection-limits.md. - "Compatible with anything not on the matrix" — see each
supported-*-matrix.md.
Pricing Tiers
See docs/pricing-architecture.md.
- Free (0–8 GPUs, 7 d retention): Denpex Lite only.
- Team ($415/mo, 8–64 GPUs, 30 d retention): + premium signatures.
- Scale ($2,495/mo, 256–1,024 GPUs, 90 d, multi-cluster): + Host Agent + RBAC.
- Growth ($4,580/mo, 4,096+ GPUs, 1 y, multi-cluster): + RAG + K8s/SLURM remediation + FedRAMP.
- Air-gap appliance (16k+ GPUs, customer VPC): custom contract.
Where Denpex Ships
- SaaS control plane — single-region primary (WNAM) with globally-replicated reads. Multi-region active-active data residency is available only in the self-hosted control plane and air-gapped appliance, not the SaaS tier (see
docs/operations/SLA.md). - Self-hosted control plane — Helm + bare-metal/RHEL/Ubuntu.
- Air-gapped appliance — single signed tarball, HSM-rooted KMS.
Contact
- Website: https://denpex.ai
- Support: support@denpex.ai
- Security: security@denpex.ai
- Federal: federal@denpex.ai