AMD has introduced RDC, a tool designed to help developers monitor GPU performance and health by providing a range of metrics that can be used for application tuning and system validation. The tool is part of the ROCm platform and offers insights into hardware status, reliability data, and runtime profiling metrics. RDC allows developers to check if a workload is throttling, if data is moving across GPU interconnects as expected, and if ECC counters are increasing. These metrics are essential for evaluating whether an application is using GPU resources effectively. This blog post compares RDC metrics with similar NVIDIA DCGM metrics to help developers understand how to interpret them on AMD GPU platforms.
RDC combines data from several ROCm components, including ROCm-SMI for hardware status, RAS for reliability and availability, and RocProfiler for runtime information. Together, these components provide developers with a comprehensive view of GPU behavior without needing to open a full profiler interface. For example, if a workload is slower than expected, RDC can help identify whether the issue is due to power or thermal throttling, inter-GPU traffic, ECC errors, or low compute utilization. This makes RDC a valuable tool for both system-level monitoring and application-level performance analysis.
The post outlines how RDC works, explaining that it uses a daemon (rdcd) and a command-line tool (rdci) to collect and query telemetry data. Developers can use commands like 'rdci dmon -u --list-all' to list all supported metrics or start collecting metrics with 'rdcd -u' followed by specific commands to monitor selected fields. This architecture allows RDC to be used interactively during development or as part of automated monitoring workflows. The tool is particularly useful for quickly checking one GPU while an application is running.
Source: amd
