> ## Documentation Index > Fetch the complete documentation index at: https://docs.zylon.ai/llms.txt > Use this file to discover all available pages before exploring further. # System Monitor > Monitor live host resources and model-serving load from Backoffice. ## Overview System Monitor gives Operators a live view of host-resource pressure and AI model-serving activity. Use it to spot CPU, memory, disk, GPU, or inference-queue pressure while the platform is running. The page is available from **Backoffice > Platform > System Monitor** when the [System Monitor feature](/en/operator-manual/configuration/features/feat-system-monitor) is enabled. System Monitor in Backoffice The exact layout can vary between Zylon versions. ## Available information Metrics update once per second. Charts keep the latest 30 samples, including recent samples replayed when the page opens. ### System metrics | Metric | What it shows | | - | - | | **System Load** | The highest normalized pressure across GPU, inference requests, CPU, and RAM. GPU reaches full pressure at 80%; running plus queued requests at 60; CPU and RAM contribute when usage exceeds 95%. | | **Active Inferences** | Chat inferences currently being processed across the platform. | | **Disk** | Used percentage and total capacity of the filesystem that stores Zylon data. | | **CPU** | Current host CPU usage and its recent history. | | **GPU** | Current overall GPU compute usage and its recent history. | | **RAM** | Current host memory usage, used and total GB, and recent history. | | **VRAM** | Current GPU-memory usage across all GPUs, used and total GB, and recent history. | ### Model metrics Select a loaded model to inspect: | Metric | What it shows | | - | - | | **Queue Inferences** | Requests waiting in Triton or in the model backend scheduler. | | **Running Inferences** | Requests currently running in the model backend. | | **Tokens/sec** | Output-token generation rate calculated between consecutive samples. | | **GPU Memory Usage** | VRAM allocated to the selected model, as a percentage of total VRAM. | | **GPU Compute** | GPU compute-core usage attributed to the selected model. | | **GPU Memory Bandwidth** | GPU memory-bus utilization attributed to the selected model; this is different from allocated VRAM. | ## Data sources and limitations * CPU and RAM come from Kubernetes node metrics when available, with host-level readings as a fallback. * Disk usage covers the filesystem containing `/var/zylon`. * GPU, VRAM, and model-serving data come from the Triton Prometheus metrics endpoint. * Per-model GPU metrics depend on process-level GPU metrics. Deployments that do not expose them show zero values. * The page is a live operational view, not persistent historical reporting. Reloading it requests the latest available samples again. * System Load is an instantaneous display signal. Request-overload protection additionally considers sustained resource usage and deployment thresholds. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.