Release Notes

GPU Operational Events remain a preview feature. This page describes changes in R620 from R615. The public NVML event APIs, category values, severity values, and CPER section layout are unchanged. Consumers must still update their event and context decoders for the changes below.

New Event Coverage

Module	Added Codes	Conditions
GPU-BUS	0x0003	GPU fallen off the bus (Xid 79).
GPU-NVLINK	0x0005–0x0014	Fatal link errors, privilege errors, and SAW, RLW, TLW, TREX, NVLPW, and NETIR errors.
GPU-NVLINK (secure)	0x0001–0x0011	Secure interrupts, containment, encryption, privilege, traffic, and parity errors.
GPU-PCIE	0x0001–0x0003	Fatal timeouts, correctable errors, and reset failures.
GPU-PSHC	0x0001–0x0004	Power-smoothing controller engagement, disengagement, and low or exhausted lifetime.
GPU-RAS	0x0017–0x003D	LTC/LRC and DRAM errors, contained and uncontained poison, SRAM ECC errors, and sparing state changes. SPARING_EVENT (0x003D) reports Xid 185.

The GPU Operational Event Catalog lists each event’s category, severity, scope, log level, and legacy Xid. Event availability depends on the GPU and the reporting path.

New Reporting Capabilities

• The Linux open kernel module exposes GPU Operational Events through nvidia tracepoints, selected by log level. See Linux Tracepoints.

• CPER record flags now reflect the event group’s recovered, previous-error, and simulated attributes. Imported CPER timestamps are marked imprecise.

• Legacy Xid messages for simulated events include [SIMULATED].