Where RM Runs: Host RM, GSP-RM, and Firmware Boot
A boundary-oriented view of how NVIDIA moves physical resource management onto GSP, bootstraps the firmware, and reconnects it to host RM over RPC.
02 / notes
Short observations from current experiments, reading, and debugging.
A boundary-oriented view of how NVIDIA moves physical resource management onto GSP, bootstraps the firmware, and reconnects it to host RM over RPC.
Following an NVIDIA driver build across RM, Kbuild, module ABI checks, GSP firmware, UVM, and DRM to see what DKMS does—and what it cannot do.
Tracing a replayable GPU fault through UVM batching, residency policy, Linux HMM invalidation, migrate_vma, GPU page-table updates, and replay.