Guard-railed autopilot, not a script that YOLOs your fleet — every action gated, role-scoped, and logged.
Most monitoring reacts to a fixed line — CPU above 90%, ECC rate above X. Fleet Memory reasons against your fleet's own history instead: "this node's HBM ECC rate is climbing the way node 114's did four days before it died." Pattern recognition against real precedent, not a static alarm.
The system tries the smallest fix first, and only reaches for the drastic one when the evidence says so. Every step is gated by policy, scoped by role, reversible until it isn't, and fully audited.
The platform ships with a validation harness built in — no separate procurement, no waiting on a benchmarking team.
| Check | What it proves |
|---|---|
| GPCNeT | Network congestion & interference — whether noisy neighbors are stealing fabric bandwidth. |
| NCCL | Multi-GPU collective bandwidth and latency — the primitives every distributed training job depends on. |
| OSU | MPI point-to-point and collective micro-benchmarks — the classic HPC interconnect health check. |
| HPL | Sustained floating-point throughput — the same benchmark that ranks the TOP500. |
Panacea Ops sits on top of what a site already runs — Slurm for scheduling, Redfish/OpenBMC for hardware control, NVIDIA UFM for fabric telemetry where present — rather than asking anyone to rip anything out. The closed loop is what's missing, not another scheduler.