Deploy, monitor, and maintain NVIDIA GPUs in production
Reference documentation for driver and toolkit compatibility, GPU management and monitoring interfaces, and diagnosing GPU health, errors, and operational events on deployed systems.
Compatibility, monitoring, and management
Plan upgrades, query GPU state, configure driver behavior, and share GPUs across processes.
CUDA Compatibility
Using new CUDA Toolkit components on systems with older base installations, including minor version and forward compatibility.
NVML API Reference Guide
The NVIDIA Management Library reference: a C-based API for monitoring and managing GPU device states.
nvidia-smi
Man page for the NVIDIA System Management Interface command-line utility.
Multi-Process Service (MPS)
Binary-compatible CUDA API implementation that lets cooperative multi-process applications, typically MPI jobs, share a GPU.
Driver Persistence
Default kernel mode driver behavior and the options for keeping the driver persistent across GPU interactions.
Health and diagnostics
Validate GPU health, interpret errors, and work through hardware failures.
NVIDIA GPU Operational EventsNew
Structured operational event reporting for deployed GPUs.
NVIDIA GPU Debug Guidelines
Error debug and diagnosis guidelines for getting servers back up and running.
XID Errors
What Xid messages mean and how to use them when analyzing and resolving GPU problems.
NVIDIA GPU Memory Error Management
Memory error recovery features introduced in the NVIDIA A100 and A800 GPUs.
Dynamic Page Retirement
Automatic retirement of framebuffer pages containing degrading memory cells, and why it matters for board longevity.
HW Field Diag
Verify GPU hardware integrity in the field. Required as part of the RMA process.
RMA Process
The standardized process for identifying products that qualify for return merchandise authorization.
GDS Support for NVMe Drives
Enabling GPUDirect Storage with NVMe drives using Linux PCI P2PDMA support.