gxtop

A digital-rain style htop/nvtop for NVIDIA GB10 machines – ASUS Ascent GX10, NVIDIA DGX Spark and other Grace Blackwell desktops. It shows the system at a glance and answers the question nvtop can't: which AI models are loaded right now, and which tool loaded them?

_{Real capture from a GX10 (200×56 terminal). The hostname, IP address, user name and one container name were replaced with placeholders.}

• Single Python file, zero dependencies. It only uses the standard library and curses.
• Unified-memory aware. On GB10, nvidia-smi reports "Not Supported" for GPU memory. gxtop adds up the per-process GPU allocations and splits the shared CPU+GPU memory by engine.
• Model & engine detection for vLLM, Ollama, ComfyUI, llama.cpp, SGLang, TGI and TensorRT-LLM. Each engine is mapped to its Docker container or systemd service.
• Live digital rain. Density and speed follow GPU load and vLLM token throughput.

Example: the Loaded models panel.

╭─┤ LOADED MODELS · ENGINES ├──────────────────────────────────────────────────────────────────────┤ 3 active ├╮
│  ● vLLM      qwen3.6-35b                          GPU   47.7G  idle     docker vllm-main :8000 pid 151891    │
│    └ nvidia/Qwen3.6-35B-A3B-NVFP4 · ctx 131072 · kv fp8 · mem 0.4 · tp 1                                     │
│      ⚡  42.3 tok/s  ▁▂▅▇█▇▅▃  prefill 812 t/s  ▶ 2  ◌ 0  KV  3.1%  prefix-hit 77%  Σ 166.7k tok            │
│  ◉ ComfyUI   queue ▶ 1  ◌ 0                       GPU   36.8G  BUSY     systemd comfyui :8188 pid 1106632    │
│    ├ diffusion qwen_image_2.1_int8_convrot  6.8G                                                             │
│    ├ text-enc  qwen3vl_8b_int8_convrot  8.7G                                                                 │
│    └ vae       qwen_image_2.1_vae_bf16  644.2M                                                               │
│  ● Ollama    1 loaded · 2 installed               GPU  881.0M  idle     docker ollama :11434 pid 163390      │
│    └ embeddinggemma:latest  307.58M BF16  649.8M (100% GPU)  ctx 2048  unloads in 4m12s                      │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────╯

Panels

Panel	Shows
GPU	utilization with history graph, allocated GPU memory, temperature, power (incl. max seen), clocks, P-state, throttle reasons
Unified memory	one bar split by engine (vLLM / ComfyUI / Ollama / …), system, cache and free memory, plus swap and memory pressure (PSI)
CPU	per-core load and frequency, with performance (P) and efficiency (E) cores marked
System	SoC / GPU / NVMe temperatures, network and disk throughput, root filesystem usage
Loaded models	each engine with its models, origin (Docker container, systemd service or process owner), port, PID and live metrics
GPU processes	an nvtop-style process list with GPU memory, CPU %, RSS, engine and container/service
Events	live log of loaded/unloaded models, finished generations (tokens, tok/s), finished ComfyUI workflows, downloads, throttling and OOM-killer events

What gets detected

Engine	Detected via	Details shown
vLLM	vllm serve process + /metrics (no API key needed)	served name, model, context, KV dtype, tok/s, prefill t/s, running/waiting requests, KV-cache usage, prefix-cache hit rate
Ollama	ollama serve process + /api/ps, /api/tags	loaded models, size, quantization, GPU share, context, time until unload
ComfyUI	main.py process + /queue + model files the process has open	checkpoints by category (diffusion, text encoder, VAE, LoRA, …), queue, in-progress downloads (.part files)
llama.cpp	llama-server process + /health, /slots	model file, context, busy slots
SGLang, TGI, TensorRT-LLM	launcher process + health endpoint	model
anything else on the GPU	nvidia-smi compute apps	command line, open model files (if readable)

Ollama's internal llama-server runners are counted towards Ollama, not shown as separate engines.

Install

mkdir -p ~/.local/bin
curl -fsSL https://raw.githubusercontent.com/medpex/gxtop/main/gxtop.py -o ~/.local/bin/gxtop
chmod +x ~/.local/bin/gxtop
gxtop

Or clone it:

git clone https://github.com/medpex/gxtop.git
ln -s "$PWD/gxtop/gxtop.py" ~/.local/bin/gxtop

Make sure ~/.local/bin is in your PATH.

Requirements

• Linux with an NVIDIA GPU and nvidia-smi. gxtop is built for GB10, but the generic parts also run on other NVIDIA systems.
• Python 3 with curses. It is tested with Python 3.12 on Ubuntu (aarch64).
• A terminal with UTF-8 and 256 colors. The minimum size is 80×24, and the logo header appears from 46 rows.
• Optional: access to the docker CLI so container names can be shown. Without it, gxtop shows container IDs.

Usage

gxtop [-i INTERVAL] [--fps FPS] [--no-rain] [--no-intro] [--json]

Key	Action
q / Esc	quit
m	toggle digital rain
+ / -	change sample interval (0.5–10 s)
f	freeze values
?	help

gxtop --json takes one sample and prints everything as JSON. This is handy for scripting or for debugging the detection.

Permissions

gxtop runs fine as a normal user. Some details are only visible for your own processes, such as the model files a process has open. To list ComfyUI checkpoints, run gxtop as the same user as ComfyUI, or as root.

Privacy

gxtop only reads local data: /proc, /sys, nvidia-smi, docker ps and the HTTP endpoints of model servers running on the same machine. It sends nothing anywhere, collects no telemetry and writes no files.

License

MIT

Disclaimer

gxtop is an independent hobby project. It is not affiliated with, endorsed by or sponsored by NVIDIA, ASUS or any of the projects it monitors. NVIDIA, DGX, Grace Blackwell, ASUS, Ascent, vLLM, Ollama, ComfyUI, llama.cpp, SGLang, TGI and TensorRT-LLM are trademarks or names of their respective owners and are used here only to describe compatibility.