The clock metrics measure the operating frequencies of various components of the GPU, including Streaming Multiprocessors (SM), memory, graphics engines, and the video engine. Maximum achievable clock speeds are also provided for these components.
| Metric [MHz] | Description |
|---|---|
| sm | Current clock speed of the Streaming Multiprocessors (SM) |
| memory | Current clock speed of the GPU's memory subsystem |
| graphics | Current clock speed optimized for graphics workloads |
| video | Current clock speed of the video engine used for encoding and decoding tasks |
| sm max | Maximum potential clock speed of the SM |
| memory max | Maximum achievable clock speed for the memory subsystem |
| graphics max | Maximum possible clock speed for graphics workloads |
| video max | Upper limit of the clock speed for the video engine |
Clock metrics provide insights into the operating frequencies of GPU subsystems and their maximum performance capabilities.
Memory usage metrics provide insights into the utilization of the GPU's memory subsystems, including the frame buffer, BAR1 memory, and total global memory usage.
| Metric [%] | Description |
|---|---|
| frame buffer | Percentage of GPU memory (VRAM) used for storing textures, frame data, and resources |
| bar1 | Percentage of BAR1 memory aperture used for CPU-GPU direct data transfer |
| global | Overall percentage of total global memory utilized across the GPU |
These metrics help monitor memory usage patterns during workloads. High memory utilization in the frame buffer or BAR1 can indicate intensive resource usage, while global memory utilization provides an overall view of GPU memory demands.
Memory used metrics measure the actual amount of GPU memory allocated in the frame buffer and BAR1 memory aperture.
| Metric [e.g. GByte] | Description |
|---|---|
| frame buffer | Amount of GPU memory (VRAM) currently used for textures, frame data, etc. |
| bar1 | Amount of BAR1 memory aperture currently used for CPU-GPU data transfers |
P-states (Performance States) represent the GPU's power and performance levels. Each P-state corresponds to a specific combination of clock frequencies and voltages, optimized for balancing performance and power consumption based on workload demands.
| Metric | Description |
|---|---|
| p-state | Current performance state of the GPU (e.g., P0, P1) |
The GPU dynamically adjusts its P-state to optimize power and thermal efficiency. For example:
- P0 (Maximum Performance): High clock speeds and voltage for demanding workloads.
- P1 to P15: Intermediate states with progressively lower clock speeds and power consumption.
- P8 (Idle or Low Performance): Low-power state for idle or light workloads.
Monitoring the P-state helps understand GPU power management behavior and workload demands. It is useful for diagnosing power efficiency and performance trade-offs during GPU operation.
Utilization metrics provide an overview of how various components of the GPU are being utilized during workloads. These metrics include overall utilization, as well as specific metrics for the encoder and decoder subsystems.
| Metric [%] | Description |
|---|---|
| util | Overall GPU utilization, including compute and graphics workloads |
| encoder util | Utilization of the GPU's encoder engine (NVIDIA) |
| decoder util | Utilization of the GPU's decoder engine (NVIDIA) |
| multimedia util | Utilization of the GPU's multimedia engine (AMD) |
These metrics help analyze GPU workload distribution and identify performance-critical areas. High overall utilization indicates intensive GPU use, while encoder and decoder utilization metrics are particularly useful for applications involving video processing tasks such as encoding and decoding.