GPU monitoring
The GPU monitor in Site24x7 provides comprehensive visibility into the health, performance, and utilization of graphics processing units (GPUs) deployed on your servers. Monitor GPU utilization, temperature, memory usage, clock speeds, power consumption, and workload distribution to optimize compute performance and prevent thermal or power-related incidents.
GPU metrics are collected by the Full-Stack Agent running on each server with GPUs. The Full-Stack Agent queries the GPU drivers to retrieve real-time and historical metrics.
Supported GPU vendors
The Full-Stack Agent supports GPU monitoring from these vendors:
- NVIDIA
- AMD
Licensing
10 GPU monitors consume one basic monitor license.
Auto-discovery
GPU monitors are automatically created by default when the Full-Stack Agent detects GPUs on a monitored server. If GPUs are added or removed from a server, rerun the agent discovery process.
Accessing GPU monitors
You can access GPU monitors in two ways:
- Via Server Monitor Dashboard: Open a server monitor and look for the GPU tab showing discovered GPUs. Click the GPU name to open its dedicated GPU monitor.
- Via Left Navigation Menu: Navigate to Server > GPU to view all GPU monitors across your account.
Performance metrics
The following metrics are organized by category. Metrics listed as Linux are currently supported by the Full-Stack Agent.
Overview metrics
Hardware identification and driver information for the GPU device.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| GPU Vendor | Text | The manufacturer of the GPU (for example, NVIDIA, AMD, and Apple). | Linux |
| GPU Model | Text | The model name of the GPU. | Linux |
| GPU Index | Text | The index of the GPU on the server (for example, 0 for the first GPU, 1 for the second). | Linux |
| GPU UUID | Text | A unique identifier assigned to the GPU device. | Linux |
| GPU Architecture | Text | The hardware generation of the GPU. Reported for NVIDIA GPUs only. | Linux |
| GPU Brand | Text | The product brand of the GPU. Reported for NVIDIA GPUs only. | Linux |
| Driver Version | Text | The version of the GPU driver installed on the system. | Linux |
| CUDA Version | Text | The CUDA version supported by the installed NVIDIA driver. | Linux |
| GPU Core Count | Count | The total number of processing cores available on the GPU. | Linux |
Utilization metrics
Real-time GPU activity and component utilization during the collection interval.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| GPU Utilization | % | The percentage of time the GPU was actively processing tasks during the collection interval. | Linux |
| Memory Bandwidth Utilization | % | The percentage of time the GPU memory controller was active. | Linux |
| Encoder Utilization | % | The percentage of the video encoding engine currently in use. | Linux |
| Decoder Utilization | % | The percentage of the video decoding engine currently in use. | Linux |
Memory metrics
GPU memory (VRAM) allocation and usage.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| Total Memory | MB | The total GPU memory (VRAM) available on the device. | Linux |
| Used Memory | MB | The amount of GPU memory currently in use by processes and the GPU driver. | Linux |
| Free Memory | MB | The amount of GPU memory currently available for allocation to new processes. | Linux |
| Memory Utilization | % | The percentage of total GPU memory currently in use. | Linux |
Temperature and power metrics
Thermal and power consumption characteristics of the GPU.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| GPU Temperature | °C | The current operating temperature of the GPU core. | Linux |
| Power Draw | W | The current power consumption of the GPU. | Linux |
| Power Limit | W | The maximum power the GPU is configured to consume. | Linux |
| Power Usage | % | The current power draw expressed as a percentage of the configured power limit. | Linux |
| Fan Speed | % | The current fan speed as a percentage of maximum speed. | Linux |
Clock speed metrics
GPU core and memory clock frequencies.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| Graphics Clock | MHz | The current operating clock speed of the GPU graphics engine. | Linux |
| Memory Clock | MHz | The current operating clock speed of the GPU memory subsystem. | Linux |
| Graphics Clock Usage | % | The current graphics clock speed as a percentage of the maximum supported speed. | Linux |
| Memory Clock Usage | % | The current memory clock speed as a percentage of the maximum supported speed. | Linux |
Throttling and performance state metrics
GPU performance state and thermal/power throttling status.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| Performance State | Text | The current power and performance level of the GPU. | Linux |
| Throttle Status | Text | Indicates whether the GPU is currently running at a reduced clock speed to protect hardware (Yes/No). | Linux |
| Throttle Reasons | Text | The active reasons why the GPU clock is being reduced. | Linux |
PCIe connectivity metrics
Data throughput and link configuration for the GPU's PCI Express (PCIe) connection to the system.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| PCIe TX Throughput | KB/s | The rate at which data is being sent from the GPU to the system over the PCIe bus. | Linux |
| PCIe RX Throughput | KB/s | The rate at which data is being received by the GPU from the system over the PCIe bus. | Linux |
| PCIe Generation | Text | The PCIe specification version currently active for the GPU connection to the system. | Linux |
| PCIe Link Width | Count | The number of PCIe lanes in use for the GPU connection. | Linux |
ECC memory metrics
Error-Correcting Code (ECC) memory error counts. These metrics are reported only on ECC-capable GPUs.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| ECC Mode | Text | Indicates whether ECC memory protection is currently enabled on the GPU (Enabled/Disabled). | Linux |
| Single-Bit ECC Errors | Count | The number of correctable single-bit memory errors detected since the last system reboot. | Linux |
| Double-Bit ECC Errors | Count | The number of uncorrectable double-bit memory errors detected since the last system reboot. | Linux |
| Aggregate Single-Bit ECC Errors | Count | The cumulative count of correctable single-bit ECC errors across the GPU's lifetime. | Linux |
| Aggregate Double-Bit ECC Errors | Count | The cumulative count of uncorrectable double-bit ECC errors across the GPU's lifetime. | Linux |
Memory page health metrics
Advanced memory reliability metrics tracking page retirement and row remapping.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| Retired Pages (Single-Bit) | Count | The number of GPU memory pages permanently taken out of service due to repeated correctable errors. | Linux |
| Retired Pages (Double-Bit) | Count | The number of GPU memory pages permanently taken out of service due to uncorrectable errors. | Linux |
| Retired Pages Pending | Text | Indicates whether retired memory pages are waiting to be finalized (Yes/No). | Linux |
| Remapped Rows (Correctable) | Count | The number of memory rows that have been remapped to spare rows following correctable errors. | Linux |
| Remapped Rows (Uncorrectable) | Count | The number of memory rows remapped following uncorrectable errors. | Linux |
| Row Remap | Text | Indicates whether a memory row remap is pending and the GPU requires a reset to complete it (Yes/No). | Linux |
| Row Remap Failure | Text | Indicates whether a row remap has failed (Yes/No). | Linux |
GPU mode metrics
GPU operational modes and feature status.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| Compute Mode | Text | Controls how the GPU is shared among processes for compute workloads. | Linux |
| Display Active | Text | Indicates whether a display output is currently connected and active on this GPU (Active/Inactive). | Linux |
| Persistence Mode | Text | Indicates whether the GPU driver is kept loaded between workloads to avoid reinitialization delays (Enabled/Disabled). | Linux |
| MIG Mode | Text | Indicates whether the GPU is partitioned into isolated Multi-Instance GPU (MIG) slices (Enabled/Disabled). | Linux |
GPU process metrics
Information about processes currently consuming GPU resources.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| Process ID | Text | The operating system process identifier for a process currently using the GPU. | Linux |
| Process Name | Text | The name of the process currently using the GPU. | Linux |
| Process Command Line | Text | The full command used to launch the process, including arguments and flags. | Linux |
| Executable Path | Text | The file system location of the process executable. | Linux |
| Process GPU Memory | MB | The amount of GPU memory being consumed by this specific process. | Linux |
| Process Type | Text | How the process is using the GPU. | Linux |
| Total Process Count | Count | The total number of processes currently using the GPU. | Linux |
| Compute Process Count | Count | The number of processes currently using the GPU for compute tasks such as AI and ML workloads. | Linux |
| Graphics Process Count | Count | The number of processes currently using the GPU for graphics rendering. | Linux |
vGPU Licensing
Get visibility into NVIDIA vGPU licensing metrics, and track license status, product details, and expiry information directly.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| vGPU License Status | Text | The licensing state of the NVIDIA vGPU software (Licensed/Unlicensed). | Linux |
| vGPU License Product | Text | The name of the licensed NVIDIA vGPU software product. | Linux |
| vGPU License Expiry | Text | The date and time when the current vGPU software license expires. | Linux |
NVLink connectivity metrics
High-speed peer-to-peer communication metrics between GPUs equipped with NVLink.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| Active NVLink Count | Count | The number of NVLink connections currently active on this GPU. | Linux |
| NVLink Version | Text | The NVLink generation supported by each link on this GPU. | Linux |
| NVLink TX Throughput | KB/s | The rate at which data is being sent to peer GPUs over NVLink. | Linux |
| NVLink RX Throughput | KB/s | The rate at which data is being received from peer GPUs over NVLink. | Linux |
| NVLink CRC Errors | Count | The total number of data integrity Cyclic Redundancy Check (CRC) errors detected across all NVLink connections. | Linux |
| NVLink Replay Errors | Count | The total number of packet retransmissions across all NVLink connections. | Linux |
| NVLink Recovery Errors | Count | The total number of times a NVLink connection had to retrain itself to restore connectivity. | Linux |
Note
In multi-GPU systems using NVLink, monitor NVLink error counts closely. Rising CRC or replay errors may indicate cable issues, improper seating, or cooling problems affecting link stability. Address issues before they impact your distributed training or inference workloads.
GPM deep profiling
Advanced performance profiling metrics that provide detailed insight into GPU execution units, memory activity, and compute pipeline utilization.
| Metric name | Unit | Description | Supported platform |
|---|---|---|---|
| SM Activity | % | The percentage of time at least one compute thread was active on the GPU's compute units (Streaming Multiprocessors). | Linux |
| SM Occupancy | % | The percentage of available compute thread slots that are occupied. | Linux |
| Tensor Core Activity | % | The percentage of time the GPU's Tensor Cores were actively processing. | Linux |
| Memory (DRAM) Activity | % | The percentage of GPU memory bandwidth currently being used. | Linux |
| FP16 Activity | % | The percentage of time the GPU's half-precision (FP16) compute pipelines were active. | Linux |
| FP32 Activity | % | The percentage of time the GPU's single-precision (FP32) compute pipelines were active. | Linux |
| FP64 Activity | % | The percentage of time the GPU's double-precision (FP64) compute pipelines were active. | Linux |
Related articles
-
On this page
- Supported GPU vendors
- Licensing
- Auto-discovery
- Accessing GPU monitors
- Performance metrics
- Overview metrics
- Utilization metrics
- Memory metrics
- Temperature and power metrics
- Clock speed metrics
- Throttling and performance state metrics
- PCIe connectivity metrics
- ECC memory metrics
- Memory page health metrics
- GPU mode metrics
- GPU process metrics
- vGPU Licensing
- NVLink connectivity metrics
- GPM deep profiling
- Related articles