A fast, keyboard-first terminal monitor for GPU infrastructure.
gputop is to GPUs what btop is to hosts: a dense, low-overhead TUI that answers "what is happening with my GPUs right now?" and, with its built-in time machine, "what happened 15 minutes ago?"
It supports NVIDIA GPUs (NVML, Linux) and Apple silicon GPUs (M1 and later, macOS), detected automatically. No CUDA toolkit, DCGM, root or cgo required.
More screenshots
Download a binary from the latest release
(gputop-{linux,darwin}-{amd64,arm64}):
curl -fLo gputop https://github.com/riteshsonawane1372/gputop/releases/latest/download/gputop-linux-amd64
chmod +x gputop && sudo install gputop /usr/local/bin/Or with Go 1.25+: go install github.com/riteshsonawane1372/gputop/cmd/gputop@latest.
On macOS, clear the quarantine flag first: xattr -d com.apple.quarantine gputop.
Requirements: the NVIDIA driver on Linux (it provides libnvidia-ml.so.1),
or an Apple silicon Mac. In containers, use the NVIDIA Container Toolkit and
--pid=host so host PIDs resolve. If no GPU is found, gputop shows what it
checked and keeps monitoring the host.
gputop # interactive TUI
gputop --demo # simulated 8-GPU node: try it without hardware
gputop --once # one-shot text summary
gputop --once --json # one JSON snapshot (gputop --json streams them)
gputop --service # headless agent: HTTP API + Prometheus /metrics
gputop --remote gpu-node-01 # TUI for a remote agentPress ? for help. The essentials: ← → switch tabs, ↑ ↓ move, Enter
opens detail, / searches, f filters (gpu:0 user:alice ns:ml), s sorts,
h opens history, p pauses, q quits. The mouse works too: click tabs, rows
and column headers, scroll with the wheel. Every key is rebindable.
- Every GPU at a glance: utilization, VRAM, power, temperature, clocks,
throttle reasons, PCIe, NVLink, MIG, ECC and Xid errors. Unsupported values
show as
N/A, never as0. - Processes and workloads: which process, container, pod and workload owns each GPU and its VRAM.
- Inference serving: TTFT, inter-token latency, end-to-end and queue latency, tokens/s, queue depth and KV-cache usage from vLLM, SGLang, TGI and llama.cpp servers, discovered automatically from GPU processes.
- History: the last 30 minutes (configurable, persisted to disk). Scrub back in time with events on the timeline.
- Health and alerts: a 0–100 health score with itemized reasons, plus idle-but-allocated GPUs, stragglers and throttling.
- Kubernetes: k9s-style pod list, describe and logs for GPU pods.
- Fleet: a hardened agent (TLS, token auth, mTLS) with a JSON API and Prometheus metrics, and a remote TUI.
- Dashboard: the NVIDIA DCGM Grafana dashboard, in the terminal.
Tabs appear only when they have something to show: NVLink only with NVLink hardware, Inference only when a server is found, and so on.
gputop reads ~/.config/gputop/config.yaml. Every key is optional; start from
examples/config.yaml or gputop --print-config.
history:
retention: 2h
theme:
name: amber # green, amber, ice, mono, dusk or your own
inference:
endpoints: # servers that discovery can't see
- name: chat-api
url: http://10.0.0.5:8000/metrics| Configuration | every option and bindable key |
| Metrics & provenance | where each value comes from |
| Inference serving | TTFT, ITL, KV cache: supported servers and how values are computed |
| Derived metrics · Health score | formulas and limitations |
| Kubernetes | DaemonSet deployment and workload correlation |
| Service & remote | agent mode, security model, Prometheus |
| JSON output | the versioned snapshot schema |
| Architecture · Performance | internals and overhead |
| Themes | built-in and custom color themes |
gputop is pre-1.0. NVIDIA support is tested against NVML's struct layouts and a
fake libnvidia-ml on amd64 and arm64, but not yet across real GPU
generations; please share results from make test-nvidia. Apple silicon
support, the health score, derived metrics, Kubernetes correlation and service
mode are experimental. Planned: DCGM, AMD and Intel providers, and packages.
See the changelog.
make build # bin/gputop
make test # unit tests, no GPU needed (make race for -race)
make lint # golangci-lint
./bin/gputop --demo # work on the UI without hardwareSee CONTRIBUTING.md.
Apache License 2.0; see LICENSE. NVIDIA, NVML, CUDA and NVLink are trademarks of NVIDIA Corporation. gputop is not affiliated with NVIDIA.