moezdil/siltide

siltide is a terminal monitor for AI accelerators from 15 vendors: GPUs, NPUs, XPUs, MLUs, DCUs, GCUs, and Apple silicon. It shows utilization, memory, processes, power, thermals, links, and health per device. History stays on disk. Pods and Slurm jobs sit next to processes. JSON and Prometheus cover fleets.

More on the site: every tab, every key, and screenshots.

Supported accelerators

The vendor list follows the device plugins in HAMi, plus Apple and Intel. Every vendor is auto-detected. --vendors nvidia,ascend limits the probe. NVIDIA (an H100 SXM, driver 570.211.01, every metric checked against nvidia-smi), Apple silicon (an M4 Pro), and AMD (a Radeon RX 9060 XT, RDNA4) run on real hardware today. On that AMD card (kernel 7.0.0-34) VRAM, utilization, temperature and power all read and match rocm-smi, including under load. Where amdgpu returns EBUSY for its SMU sensors, those values show N/A. The rest are built against each vendor's documented tool output, with a fixture behind every parser. A hardware report through an issue is the fastest way to move one from "should work" to confirmed.

Per-process is the column worth reading first: whether the vendor's tool names the processes holding a device, not just the device totals. It is what turns "this GPU is at 90%" into "this job is holding it". Where it says no, the vendor's own tool does not report it, so neither do we.

Vendor	Devices	Source	Per-process
NVIDIA	GPUs, MIG slices	NVML loaded with dlopen (no cgo), Xid events, NVLink, ECC, row remap, PCIe AER via sysfs	yes, with /proc enrichment
Apple	Apple silicon GPU	ioreg, IOReport (power, energy), SMC (temperature), AGX user clients (per-process GPU time)	yes
AMD	Instinct, Radeon	sysfs (/sys/class/drm) and DRM fdinfo	yes
Intel	Data Center GPU, Arc	sysfs and DRM fdinfo	yes
Huawei Ascend	NPUs	npu-smi info	yes
AWS	Inferentia, Trainium	neuron-ls, neuron-monitor	no
Cambricon	MLUs	cnmon	yes
Enflame	GCUs	efsmi	no
Hygon	DCUs	hy-smi (JSON)	no
Iluvatar CoreX	GPUs	ixsmi -q -x (XML)	no
Kunlunxin	XPUs	xpu_smi	no
MetaX	GPUs	mx-smi	no
Moore Threads	GPUs	mthreads-gmi	no
Biren	GPUs	brsmi	no
VastAI	VA series	PCI sysfs	no

Fixture sources: testdata/SOURCES. A value is measured or shown as N/A, never guessed. A tool whose output fails to parse shows its error in the Health tab.

Install

The short version, with every platform and the verification steps in docs/INSTALL.md:

# Download the installer, then run it. Two steps rather than curl | sh, so
# you can read the script first. It picks the build for this machine, checks
# it against the published checksums, and installs into /usr/local/bin
# (or ~/.local/bin if that needs a password).
curl -fsSLO https://raw.githubusercontent.com/moezdil/siltide/main/packaging/install/install.sh
sh install.sh

brew install moezdil/tap/siltide              # macOS and Linux
go install github.com/moezdil/siltide@latest  # from source
nix run github:moezdil/siltide -- --demo      # without installing anything

Using it

siltide                      # interactive terminal UI, vendors auto-detected
siltide --demo               # explore every view with a simulated fleet
siltide --once --json        # one snapshot on stdout
siltide --service --listen :9800   # headless, with the API and /metrics
siltide --mcp-stdio          # answer an agent over the Model Context Protocol
siltide --update             # move to the newest release, checksum verified

? in the interface lists every tab, key and filter. The rest is in docs/USAGE.md.

Documentation

docs/INSTALL.md	every platform, verifying a download, updating, uninstalling
docs/USAGE.md	the interface, the filter language, configuration, fleets, alerts, the API
docs/MCP.md	serving the snapshot to an agent
docs/PERFORMANCE.md	what it costs to run, and the scripts that measure it
docs/DEVELOPMENT.md	building, testing, adding a vendor
docs/PLUGINS.md	the plugin surface: a signed manifest that runs no code
examples/config.yaml	every configuration key, with its default

A systemd unit, a Kubernetes DaemonSet and a compose stack with Prometheus and Grafana are in deploy/.

Contributing

See CONTRIBUTING.md for how to add a vendor and what a change needs before it merges. Security reports: SECURITY.md. Roadmap: the issue list. Changes: CHANGELOG.md.

License

Apache-2.0. See LICENSE.