Monitor inference health in layers: collect GPU telemetry with DCGM or an equivalent device monitor, scrape metrics from the inference server, and correlate both with request throughput, queue depth, latency, and failures. For NVIDIA Triton, the default Prometheus endpoint is http://localhost:8002/metrics; vLLM exposes LLM-specific request, token, and cache metrics. Build dashboards and alerts around service objectives, not GPU utilization alone.
What to monitor, and why GPU utilization is not enough
GPU utilization tells you whether a device is busy, but it does not tell you whether requests are waiting, taking too long, failing, or being served efficiently. Pair device-level signals with serving-layer measurements so you can distinguish resource pressure from queueing, execution delays, and collection problems. NVIDIA’s full-stack observability guidance emphasizes that a symptom in one layer can have a cause elsewhere in the stack.
- Device health: GPU utilization, memory use, power, and available health or error signals.
- Serving capacity: successful requests or tokens, running and waiting requests, pending work, and batching behavior.
- Latency: end-to-end timing plus queue and execution phases, using percentiles where supported.
- Failures: request failure rate and reason labels, correlated with server and backend logs.
- Telemetry health: whether exporters and metrics endpoints are reachable and being scraped.
Choose the collection layers
These components cover different parts of the system rather than serving as interchangeable monitoring products. The right combination depends on the GPU vendor, inference engine, deployment environment, and the monitoring stack already operated by your team.
| Layer or component | Role | Check before adopting |
|---|---|---|
| NVIDIA DCGM and DCGM-Exporter | GPU telemetry and health monitoring; NVIDIA describes DCGM-Exporter as its Kubernetes-oriented integration and lists Prometheus as an integration option. See NVIDIA DCGM. | GPU and driver support, available health signals, per-device visibility, deployment environment, and collector integration. |
| Triton metrics | Request outcomes, queue and compute timings, and GPU or CPU metrics where enabled. See Triton Inference Server Metrics. | Triton version, model labels, batching, metric update behavior, and failure-reason detail. |
| vLLM metrics | LLM request phases, token timing, queue state, KV-cache use, and preemptions. See vLLM Production Metrics. | Deployed version, metric lifecycle, histogram resolution, and label/cardinality cost. |
| AIPerf server-metric collection | Collection and troubleshooting of compatible serving metrics during benchmarking. See AIPerf Server Metrics Collection. | Whether the need is benchmark analysis or always-on operations, endpoint compatibility, collection interval, and output format. |
| Prometheus and Grafana | Collection/query and dashboards in the NVIDIA example monitoring stack; NVIDIA’s observability guidance describes a high-level triage dashboard followed by deeper, domain-specific investigation. | Retention, cardinality, alert integrations, existing expertise, and operational ownership. |
The implementation details below focus on NVIDIA GPUs, Triton, and vLLM. Metric names, availability, defaults, and deprecation behavior can change; verify them against the exact GPU deployment and serving release in use.
Recommended Free Tools
#1 Best Overall
- System Compatibility Note: This 2-slot card measures 271 x 112 x 39 mm and requires a single 12V-2x6-pin power connector. Please verify chassis and PSU compatibility before purchase.
- Dedicated Support: Please contact us directly through Amazon for any product questions or assistance you may require.
- Professional Intel Arc Pro B70 GPU: Built on the Intel Xe2-HPG architecture, it features 32 Xe cores and 256 XMX engines, designed to accelerate AI, rendering, and complex visualization workloads.
- Massive 32GB GDDR6 VRAM: Equipped with 32GB of high-speed GDDR6 memory on a 256-bit bus, running at 19 Gbps, which allows for handling large AI models and complex datasets locally.
- High-Performance Engine Clock: Delivers an engine clock of 2540 MHz, providing the compute power needed for demanding professional applications and AI inference.
Collect GPU and host telemetry
For NVIDIA data-center GPUs, DCGM provides GPU monitoring and health telemetry. Select the per-device utilization, memory, power, and health or error signals available for the deployed hardware and software versions. NVIDIA’s examples of useful infrastructure signals include GPU utilization, power use, XID errors, fabric error rates, and job queue wait; fabric and scheduler signals matter when the deployment architecture uses them.
Correlate device measurements with inference traffic and queueing. A busy GPU does not by itself establish that model execution is the source of user-visible latency; a request may be waiting in a queue, or degradation may originate in the host, fabric, or scheduling layer.
Rank #2
- PLEASE NOTE: Exporting an NVIDIA RTX Pro 6000 GPU outside the US requires strict adherence to the U.S. Export Administration Regulations (EAR) and issuance of an export license from the Bureau of Industry and Security (BIS). Compliance and Know Your Customer (KYC) screening may be required as a condition of order acceptance. [NVIDIA Blackwell Streaming Multiprocessor] The new SM features increased processing throughput, and new neural shaders that integrate neural networks inside of programmable shaders | DLSS 4: Multi Frame Generation ensures ultra-smooth frame pacing for lifelike simulations.
- [Double-Flow-Through Design] The RTX PRO 6000 Blackwell features a double-flow-through cooling design, optimizing efficiency and airflow to sustain peak performance under 600W power loads. | [5th Gen Tensor Cores] Deliver up to 3X the performance of the previous generation and support for FP4 precision for faster AI model processing times with reduced memory usage, enabling local fine-tuning of LLMs and generative AI | [4th Gen Ray Tracing Cores] Double the ray-triangle intersection rate of the previous generation to create photoreal, physically accurate scenes and immersive 3D designs with RTX Mega Geometry, which enables up to 100X more ray-traced triangles.
- [PCIe Gen 5] Support for PCIe Gen 5 provides double the bandwidth of PCIe Gen 4, improving data-transfer speeds from CPU memory and unlocking faster performance for data-intensive tasks like AI, data science, and 3D modeling. | [GDDR7 Memory] With 96 GB of GPU memory and 1.8 TB ps bandwidth, it can tackle massive 3D and AI projects, fine-tune AI models locally, explore large-scale VR environments, and drive larger multi-app workflows.
- [DisplayPort 2.1] Achieve unparalleled visual clarity and performance, driving high resolution displays at up to 8K at 240 Hz and 16K at 60 Hz. Increased bandwidth enables seamless multi-monitor setups while HDR and higher color depth support ensures superior color accuracy for precision work, such as video editing, 3D design, and live broadcasting.
- [Universal MIG] Divide a single RTX PRO 6000 Blackwell into multiple isolated instances, each with dedicated resources, allowing for concurrent execution of multiple workloads, optimized GPU utilization, and secure isolation of different applications or users. [WARRANTY] 3 YR Manufacturer's Warranty. Bulk OEM Packaging. Retail Packaging is NOT included.
Scrape metrics from the inference server
Triton Inference Server
Triton exposes Prometheus-compatible metrics for collection rather than pushing them to a remote server. Its default endpoint is http://localhost:8002/metrics, and the metrics configuration can change that endpoint. The documented groups include request counts, pending requests, latency components, GPU utilization and memory, and CPU utilization and memory. Triton’s GPU metric collection uses DCGM.
- GPU state:
nv_gpu_utilization,nv_gpu_memory_used_bytes, andnv_gpu_memory_total_bytes. - Request outcomes:
nv_inference_request_successandnv_inference_request_failure. Failure reasons includeREJECTED,CANCELED,BACKEND, andOTHER. For ensemble failures, reason-label detail has a documented limitation and may be reported asOTHER. - Waiting work:
nv_inference_pending_request_countcounts requests received but not yet executing in a backend model instance. - Latency phases:
nv_inference_request_duration_us,nv_inference_queue_duration_us,nv_inference_compute_input_duration_us,nv_inference_compute_infer_duration_us, andnv_inference_compute_output_duration_us.
Triton’s latency metrics in this group are cumulative counters, not individual-request latency readings. Use rate or delta calculations in the monitoring system, and use distributions or histogram data where available, to derive meaningful latency views. Triton documents average batch size for batch-capable models as inference count divided by execution count, which can help explain how request volume relates to model executions.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #3
- System Compatibility Note: 2-slot card, 271x112x39mm, single 8-pin power, 200W TDP. Verify chassis clearance and PSU capacity before purchase.
- Dedicated Support: Please contact us directly through Amazon for any product questions or assistance you may require.
- 24GB GDDR6 on 192-Bit Bus: Massive 24GB memory with 456 GB/s bandwidth – ideal for LLMs, AI inference, 3D rendering, and generative design.
- Intel Xe2-HPG Architecture: Built on Intel's next-gen architecture with 20 Xe cores and 160 XMX engines for AI acceleration (197 INT8 TOPS).
- PCIe 5.0 Support: PCI Express 5.0 x16 interface for maximum bandwidth with the latest workstation platforms.
Triton distinguishes per-request metrics from metrics updated at an interval. Changing the metrics polling interval affects the latter, not per-request measurements, so account for scrape cadence and metric update cadence when interpreting apparent delays.
vLLM
Use the metrics endpoint from the vLLM version actually deployed, and confirm names against that release. Documented core metrics include vllm:e2e_request_latency_seconds, vllm:request_queue_time_seconds, vllm:request_inference_time_seconds, vllm:request_prefill_time_seconds, vllm:request_decode_time_seconds, vllm:time_to_first_token_seconds, and vllm:inter_token_latency_seconds. Also track running and waiting requests, KV-cache use, and preemptions where exposed by the deployed version.
Rank #4
- 【High-Performance APU】The MS-S1 MAX features an AMD Ryzen AI Max+ 395 APU, integrating a Zen 5 architecture CPU (up to 5.1GHz, 16C/32T, 64M L3 Cache), an RDNA 3.5 GPU, and an NPU (50 TOPS). The total system output is 126 TOPS. It provides powerful parallel computing capabilities for demanding AI workflows. It is ideal for running local LLMs, multimodal models, and computationally intensive tasks
- 【128GB UMA Memory】Equipped with up to 128GB of LPDDR5x-8000MT/s unified memory, it enables the CPU and GPU to access a shared, high-bandwidth memory pool with extremely low latency. Ideal for large-scale AI inference, 3D workloads, and complex timelines in video editing. It eliminates traditional VRAM bottlenecks, ensuring smoother data transfer during high-intensity computations. The UMA design maximizes performance stability under high loads
- 【Flexible Expansion】The MS-S1 MAX features USB4 V2 (up to 80Gbps), dual 10GbE LAN, HDMI 2.1 (up to 8K60), a full-length PCIe x16 expansion slot, and dual M.2 slots supporting up to 16TB RAID 0/1. Wi-Fi 7 provides stronger signal coverage and a more stable wireless experience. The slide-out design facilitates upgrades and maintenance. It easily adapts to personal, studio, or rack-mount enterprise environments
- 【High-Efficiency Cooling System】Utilizing an aerospace-grade aluminum alloy chassis, copper base plate, six heat pipes, dual turbine fans, and advanced PCM thermal conductive material, it maintains stable cooling performance even under continuous load. This system supports 130W continuous power and 160W peak power operation, with a built-in 320W power supply. It boasts multiple global certifications including CCC, FCC, UL, CE, and UKCA, ensuring stable and reliable operation in various environments
- 【Cluster Design】Two MS-S1 MAX units can be configured as a dual-unit cluster to run a large 235B Q4 model locally, achieving an output speed of 10.87 tok/s. Supporting 2U rack deployment, multiple MS-S1 MAX units can be cascaded into a distributed cluster to create a high-efficiency AI computing center. A cluster of four MS-S1 MAX units successfully ran a DeepSeek-R1 671B Q4 large model. A reserved cluster power-on interface allows for unified start-up and shutdown
Choose histogram buckets to reflect the latency objectives you need to assess. vLLM warns that every configured boundary adds series for each metric and label combination, increasing storage, scrape size, and query costs. Keep boundary lists short and customize only the metric families you need to monitor.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Build dashboards and alerts around service symptoms
Organize panels so an operator can move from user impact to likely cause without treating any single metric as a diagnosis.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →- Service overview: show request success and failure rates, throughput, p50/p95/p99 latency where the metrics support them, and current service-objective or error-budget status.
- Latency breakdown: compare total latency with queue time and relevant execution phases. For LLM serving, include prefill, decode, time to first token, and inter-token latency when available.
- Capacity: chart pending, waiting, and running requests alongside GPU utilization and memory. For vLLM, include KV-cache usage and preemptions; for Triton, include batching or execution behavior relevant to the model.
- Platform health: expose GPU power and health/error events, adding host, fabric, or scheduler measurements when those layers are part of the deployment.
- Telemetry health: show target availability, scrape errors, expected series presence, and exporter or server process health.
Keep alerts few and actionable: tie them to service-level indicators or objectives and give each alert a defined remediation path. A missing series or unreachable endpoint is a monitoring failure to investigate, not evidence that the workload is healthy.
Quick Recap
Troubleshoot common inference symptoms
| Symptom | Compare | Next checks |
|---|---|---|
| Tail latency rises while median latency remains acceptable | vLLM waiting requests and latency distribution; Triton queue duration and pending-request count. | If waiting work is growing, investigate concurrency, scheduling, and serving capacity. AIPerf’s server-metric guidance also associates spikes in vLLM waiting requests with queue buildup. |
| OOM or memory-related crashes | GPU memory, vLLM KV-cache usage, and preemption count. | Cache use approaching 1.0 is a diagnostic reason to investigate memory pressure, not a universal threshold or proof of root cause. AIPerf’s vLLM troubleshooting example suggests considering a lower max_model_len or higher gpu_memory_utilization; validate those settings against the actual version and workload before changing them. |
| Throughput is low | Running versus waiting requests, successful-request rate, GPU utilization, and relevant host or network signals. | AIPerf distinguishes low running and low waiting request counts, which can indicate a client-side bottleneck, from high waiting counts, which point toward server-side queueing. Confirm with traffic generation and system measurements. |
| Failure counter rises | Triton failure-reason labels and corresponding backend or server logs. | Separate rejected requests, cancellations, backend execution errors, and other errors where labels permit. The counter describes outcomes; use logs and surrounding metrics to find the cause. |
| Metrics disappear | Configured endpoint response, scrape target, server state, network, and firewall. | Test the actual configured endpoint directly and verify that it returns Prometheus-formatted metrics. AIPerf documents endpoint and content-type checks for collection troubleshooting. |
| GPU utilization looks normal but service performance degrades | Request phases, queueing, GPU health, and node or fabric health. | Trace the issue across layers; a GPU utilization reading alone cannot rule out a host, fabric, or job-scheduling problem. |
Implementation checklist
- Confirm the GPU, driver, exporter, and serving-engine versions, then verify which documented metrics are available in the deployed combination.
- Scrape device and serving metrics into the existing monitoring system; confirm endpoints, target health, and expected series before building alerts.
- Use rate or delta calculations for cumulative counters and choose histogram resolution based on the service objectives and acceptable series cost.
- Correlate latency and failure changes with queue depth, throughput, memory pressure, and platform health rather than relying on utilization alone.
- Attach an owner and a concrete next action to every alert that pages an operator.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
