siltide — AI server management

AI server management tool that integrates monitoring of multiple types of accelerators such as GPU·NPU and visualizes utilization·temperature·power in real time.

Basic information

What is this program?

siltide is an AI server management tool that monitors GPUs and NPUs from different vendors such as NVIDIA, AMD, and Intel on a single screen. It collects each accelerator's utilization, memory usage, temperature, and power consumption in real time and provides them as graphs and alerts.

In AI training and inference servers, a bottleneck in a single accelerator often blocks the entire pipeline. siltide visually shows load imbalances across multiple accelerators and automatically detects and alerts on hotspots (overload and overheating).

What problems does it solve?

• ('In AI training server management', 'when utilization across multiple GPUs is uneven', 'you can compare each accelerator's util graphs to identify the bottleneck GPU.')
• ('In overheating and power management', 'when temperature and power approach thresholds', 'you can respond before throttling with real-time alerts.')
• ('In hardware compatibility checks', 'when you need to see which accelerators are installed in the server', 'it displays the GPU and NPU list and driver and CUDA/ROCm versions at once.')

Key features

Unified dashboard

Monitor NVIDIA, AMD, and Intel accelerators from a single dashboard

• Use case: In mixed hardware environments
• Result: Unified visualization without vendor-specific tools
• Limitations: Some newer accelerators may not be supported yet.

Real-time alerts

Sends alerts when util, temperature, or power thresholds are exceeded

• Use case: Preventing overheating during long training runs
• Result: Preventing throttling and downtime in advance
• Limitations: Requires configuring alert channels (Slack or webhook).

Load analysis

Automatically detects utilization deviation across accelerators

• Use case: Analyzing bottlenecks in training and inference pipelines
• Result: Contrasts low-load GPUs and overloaded GPUs graphically
• Limitations: Analysis accuracy depends on the sampling interval.

Resource snapshot

Saves accelerator state at a specific point in time as a snapshot

• Use case: Failure analysis and reproduction
• Result: You can compare and analyze snapshots later
• Limitations: Snapshot retention period depends on configuration.

Installation and usage

Linux (NVIDIA)

Install with pip in an environment where CUDA 12.x or later is installed.

pip3 install siltide
siltide --version

Linux (AMD ROCm)

Install in a ROCm 6.x environment.

pip3 install siltide[rocm]
ROCM_PATH=/opt/rocm siltide --detect

Linux (source)

Building from source requires Python 3.10+ and Go 1.21+.

git clone https://github.com/siltide-dev/siltide
cd siltide
make build
sudo make install

Real-world usage examples

Accelerator detection

Check the accelerators installed in the server and their driver versions.

$ siltide --detect

[detected] NVIDIA A100-SXM4-80GB  (CUDA 12.6, Driver 560.35)
[detected] NVIDIA A100-SXM4-80GB  (CUDA 12.6, Driver 560.35)
[detected] Intel Data Center GPU Max 1550  (oneAPI 2025.1)
[summary] 3 accelerators found

Real-time monitoring dashboard

Open the web dashboard to view real-time util, temperature, and power.

$ siltide --dashboard --port 9100
# browser http://localhost:9100
# GPU0: util 87% · 61°C · 380W
# GPU1: util 92% · 64°C · 402W
# NPU0: util 34% · 48°C · 120W

Threshold alert configuration

Set util and temperature thresholds to receive alerts when they are exceeded.

$ siltide --alert --max-temp 80 --max-util 95 --webhook https://hooks.example.com/alert
# [alert] GPU1 util 96% > 95% (webhook sent)

Advantages and limitations

Advantages

• Unified monitoring of NVIDIA, AMD, and Intel accelerators with a single tool
• Provides a real-time dashboard plus threshold alerts
• Supports training bottleneck analysis through automatic load deviation detection
• Lightweight agent architecture minimizes load on training servers

Considerations before use

• As an early version (0.9.x), support for some newer accelerators may be limited.
• NPU (Intel and AMD) support has fewer features compared to GPU support.
• Alert channels (Slack or webhook) must be configured separately.

Comparison with similar programs

Official resources and further learning

• Official website — https://siltide.io/
• Official documentation / knowledge base — https://siltide.io/docs/
• Official download — https://siltide.io/downloads
• Release notes — https://siltide.io/changelog
• Source / development resources — https://github.com/moezdil/siltide

Version and supported environment information was confirmed on the official site as of 2026-10-18. Installation commands may vary depending on your distribution and version, so please also check the official installation documentation.

Actual execution screen

The screen below is a synthetic example created to help understand the usage flow (not an actual execution result). The screen is a synthetic example and not an actual execution result.

!siltide accelerator detection and dashboard example: detected 2 NVIDIA A100s and 1 Intel GPU Max

Real production server

Source (HTML): https://games.aidebatehub.com/programs/apps/aiserver/siltide.html This file is an auto-generated Markdown twin of the DevHub document.