GPU simulation

nl6 simulates NVIDIA DGX and HGX GPU servers — complete with NVIDIA DCGM OIDs, nvidia-smi SSH output, and DCGM-shaped REST endpoints. This section consolidates the GPU-specific design notes originally kept under plans/.

Pages

• DCGM simulation — the simulator-side plan: new metric OID types, the GPU cycler extension, device profiles, resource file layout, and integration points. This describes how the simulator produces the data the other two pages consume.
• Protobuf model — the GpuDevice protobuf schema in probler/proto/inventory.proto: top-level entity, host system resources, individual GPUs, NVLink / NVSwitch topology, and health.
• Pollaris and parsing rules — the l8parser pollaris definitions and parsing rules that collect SNMP, SSH, and REST data from the simulator and populate the GpuDevice model. Part 1 covers the foundational SNMP polls; Part 2 covers SSH / REST and the gap closure to full coverage.

Supported GPU servers

Device	GPUs	VRAM/GPU	System RAM
NVIDIA DGX-A100	8	80 GB	1 TB
NVIDIA DGX-H100	8	80 GB	2 TB
NVIDIA HGX-H200	8	141 GB	2 TB

See Device types → GPU servers for where these sit in the broader device catalogue.

Design-note provenance

Every page in this section was originally authored as a design plan under the repository's plans/ directory. The plans/ directory has been retired — the pages here are the current reference. History is preserved via git mv; run git log --follow docs/reference/gpu/<page>.md to see the full history including the original plans/ entries.