Project · Open source · v2.70

sparkDash, every Spark in one window.

A real-time web dashboard for one or more NVIDIA DGX Sparks, and any Linux or Windows machine with an NVIDIA GPU, in a single browser window or as an app on your phone. It streams GPU, CPU, unified memory, storage, network and local LLM metrics, runs the benchmarks behind the speeds on my recipe pages, and lets you add, edit, reorder or remove machines without restarts.

sparkDash Overview in the dark theme: four DGX Sparks, one head and three workers, each with unified memory, GPU use, temperature and power, the model it serves and live decode tok/s
sparkDash on an iPhone: the Overview with Spark 1, the head, decoding live

What's new · v2.70 · October 22

Alerts that reach you, and a lab with a memory.

The big one

Alert notifications to ntfy, Discord, Slack or a webhook

Until now nothing told you about a problem unless a dashboard was open. Turn on Settings, Alert notifications and the server tells you when a health finding (GPU temperature, unified memory, Xid or OOM, RoCE) or an Overview exception (a unit offline, throttling, storage over 90%, an LLM or the tailnet unavailable) has held for its wait, and again when it has stayed clear as long. It uses the same rules as the cards, so a message always matches what the dashboard shows. Off by default, with every condition switchable and the channel addresses stored encrypted.

01

Metrics history

An opt-in setting keeps 48 hours at one-minute detail and 30 days at 15 minutes for every unit: GPU, memory headroom, CPU, network and each LLM's tok/s. A History panel on the unit page draws it with one shared crosshair, and gaps stay gaps when a unit was offline.

02

Cost compare

A new page prices your fleet's token counters as if they had run on any hosted model, with prefill at the input rate, completions at the output rate and cached prefill at the cache rate, so you also see what your prefix cache saves.

03

Decode benchmark: temperature and thinking

The decode benchmark now runs at temperature 0.7 by default instead of 0, and the Prose type can run with or without thinking. You can set the temperature (0 to 2) per run, and the result text and share card show both. Runs saved before 2.70 are labelled temp 0, thinking off.

04

NVIDIA Flare

Flare (Helios) is detected from its model card, with live decode and prefill tok/s, cached versus uncached prefill, slots, the queue and KV use, and all four benchmarks work on it.

05

Radiance and Strata

Both engines are recognised with live rates, instead of reading as another backend with 0 tok/s mid-request. Overview cards also say how many parallel slots an engine decodes at once.

06

Intel Arc GPUs

Discrete Intel Arc cards on the xe driver show up next to NVIDIA and AMD ones, locally and over SSH, with temperature, fan, clock, power and VRAM read straight from sysfs.

07

CPU temperature and power on the cards

A Spark's CPU runs hotter than its GPU under the same load, so the Overview cards now show CPU temp and CPU power bars with the same colours as the GPU ones.

08

Token totals, finer ranges

Last 24 hours, 6 hours and 1 hour join the range picker, and the Overview table lists one row per unit and model. Tool Eval Bench can compare a run on one Spark with a run from another.

Also fixed
  • The decode benchmark no longer changes its prompt as you add streams, which had made the 4-stream figure fall
  • vLLM prefill no longer counts prefix-cache hits as work, so a cached prompt stops reading as 16,000 tok/s
  • The cache hit rate is no longer a false 0% on engines that never report it
  • Quality bench history and compare cover every port on a unit

v2.50 and v2.40.1 · October 14 and 15

09

Tool Eval cards

A finished Tool Eval result becomes a share image with the score, rating, pass split and a tile per category, and two runs become one comparison image of the change between them.

10

GPU clock cap from the dashboard

Cap a Spark's graphics clock from Edit unit, optionally keeping it after a reboot. Without passwordless sudo it asks for the password once, or gives you a command to run on the Spark yourself.

11

A Temps page

GPU temperature across the fleet over the last day or week, one graph per unit, drawn from a minute-by-minute history kept for 7 days.

Also in these releases
  • Mac units come online with CPU, GPU and RAM through macmon
  • AMD GPUs report through amd-smi
  • Cards list every loaded vLLM model

Version 2.40 · October 11

LM Studio and Ollama, recognised for what they are.

The big one

LM Studio and Ollama, detected properly

sparkDash now finds LM Studio and Ollama through their own endpoints, so an LM Studio server is no longer labelled SGLang or vLLM, and a backend it has identified is never downgraded. Each LLM panel follows its port, not its place in the list, and the Add wizard and every LLM panel have LM Studio :1234, Ollama :11434 and TensorFold :8888 presets. From a contribution by @HassanEmam.

01

Windows PCs with an AMD iGPU

A Windows PC with an AMD iGPU (Strix Halo) now shows GPU use and memory from Windows' own GPU counters when there is no nvidia-smi.

02

Hide any Overview panel

Settings has a switch for Recent activity and for each tile: Fleet decode, Fleet power, Memory in use and Tokens today. The rest share the row, and hidden panels stay in the sidebar.

03

The Windows key gotcha, explained

For an administrator account, Windows reads keys from administrators_authorized_keys, not your own folder. The setup docs show the fix, and a refused key now says so.

v2.31 and v2.32 · October 10

04

RoCE / RDMA monitoring

Sparks with RDMA ports get a RoCE panel: link state and rate, traffic, errors and drops, PFC and QoS settings and the RDMA counters, read on the Spark itself with no root. Two new health alerts: a RoCE link went down, and RoCE is dropping packets.

05

A time zone setting

Pick a time zone in Settings and the Tokens, Energy and Activity charts follow it, so the 24-hour charts line up with your local time on any device you view them from.

06

LLMs reachable only over SSH

An engine on a port that only SSH reaches, like a remote Spark behind a NAT, is now monitored through an SSH forward, with a via SSH chip on its panel.

07

Honest speed readings

TensorFold prefill tok/s counts only the tokens it actually computed, so prefix-cache hits no longer read as 100k+, and the decode benchmark's live figure comes from its own streams instead of showing 0.

08

Calmer health alerts

A finding is held for 90 seconds, so a GPU sitting right at a threshold raises one event, not one a minute. /metrics also shows when each collector last returned real data.

Also fixed
  • The iPhone app's header clears the status bar, and its icon is the logo
  • The GPU history chart fills its panel on wide screens
  • Dropping a Spark after dragging it no longer opens it
  • Unified memory, CPU and Models share one row on the Spark page

Version 2.30 · October 10

Windows PCs, now in your fleet.

The big one

Windows PCs as units

Monitor a Windows PC with an NVIDIA GPU over its built-in OpenSSH Server, with no agent to install. Pick Windows PC with an NVIDIA GPU as the unit type, enter its LAN IP and your Windows user, and it joins the Overview next to your Sparks: GPU temperature, use, power, VRAM and processes, plus RAM, uptime, CPU load, disks, network and shutdown.

Why 2.30? It is 2.0.0 plus 0.01 for each of the 30 issues and pull requests closed since 2.0 shipped, one day earlier.

01

A phone app

sparkDash installs to your home screen as an app (PWA) and runs fullscreen with its own icon, over HTTPS such as Tailscale Serve. Live data always comes from the network.

02

Prometheus export

An opt-in /metrics endpoint, so Prometheus, VictoriaMetrics or Grafana can keep your fleet history beyond the dashboard's own.

03

Monthly energy history

Fleet energy keeps a permanent total per month, with cost, Wh per token and a year view of twelve monthly bars. It survives resets and restarts.

04

Every GPU on its own line

On hosts with several GPUs the history graph switches between Combined, All GPUs and a single card. Single-GPU hosts look as before.

05

Smarter benchmarks and controls

A model picker for multi-model servers in the decode benchmark, a per-Spark llmHost, a working poll-interval setting, and a Restart button after a fleet change.

Also fixed
  • Docker upgrades no longer end on "Frontend not built"
  • The SSH tunnel fallback for benchmarks and the Showcase works again
  • Model launchers no longer inherit sparkDash's own port
  • A token-protected remote setup no longer shows a blank page

Version 2.0 · October 9

A new sparkDash, rebuilt from the shell up.

01

A new shell

A sidebar, a command palette (Ctrl/⌘ K), four themes (White, Light, Dark, OLED), smooth page transitions and real links everywhere: right-click, new tab.

02

A rebuilt Overview

A card per Spark with the model row, head and worker relations, a Model/System memory bar, and a launcher to load a model from the card.

03

A Benchmarks section

Decode, Prefill, Quality and the new Tool Eval Bench, with trials, side-by-side compare and a one-click upgrade.

04

Pages of their own

The Prompt Showcase is now an in-app page with much smoother streaming, and Token totals, Fleet energy and Activity each get a page.

05

Run your models from it

Start and stop your own LLMs from the dashboard, watch live prefill tok/s during long prefills, and score models with GSM8K and MMLU in the Quality bench.

Everything since 1.0 is in the CHANGELOG.

A look inside

Monitor, benchmark, compare.

Screenshots from my own lab: four DGX Sparks serving GLM 5.3 Flash on TensorFold. Network addresses are hidden.

sparkDash Spark detail page: GPU at 95% with temperature, power and clock, a 30-minute history chart, unified memory with the processes holding it, CPU, and the LLM server panel with live decode tok/s
A Spark, in detailGPU, unified memory, CPU and the LLM server on one page: live tok/s, slots, context, and the models you can start or stop.
sparkDash Decode benchmark: results at 1, 2, 4 and 8 concurrent streams with aggregate and per-stream tok/s, and the run history
Decode benchmarkStreaming decode tok/s at rising concurrency, by type (Structured, Prose, Code, JSON), at the temperature you choose, with every run kept to compare.
sparkDash Tool Eval Bench page: run presets from a short run to the full standard suite, and the command it runs
Tool Eval BenchTool-calling quality across 69 scenarios, from a quick check to the full suite, powered by tool-eval-bench.
sparkDash Token totals page: 24.2M tokens in total, tokens per day by type, insights and a table by model
Token totalsWhat your Sparks generated and read, by model and by day, with the cache hit rate and the busiest day.
sparkDash Fleet energy page: current draw, energy in the last 24 hours, estimated cost, average power, and energy per hour by Spark
Fleet energyEstimated power and energy for the whole fleet: current draw, energy per day, cost and Wh per thousand tokens.
sparkDash Overview in the White theme: four DGX Sparks with memory, GPU, temperature, power, storage and the model each one serves
Four themesWhite, Light, Dark and OLED. This is the Overview in White.

On your phone

Your lab, in your pocket.

The same dashboard on a phone, with its own tab bar: the Overview, each Spark, the stats and the settings. Since 2.30 it installs to your home screen as an app and opens fullscreen. Installing needs HTTPS, for example Tailscale Serve; on an iPhone, use Share, then Add to Home Screen.

sparkDash on an iPhone: the Overview, Spark 1 decoding live
sparkDash on an iPhone: the Spark 1 detail page with GPU utilization and power

What you get

Built for the GB10, open to any NVIDIA GPU.

Multi-unit

Any number of units, each with a tabbed detail page, plus a shared Overview. Add, edit, reorder or remove them from the UI, with no restart.

Not only Sparks

Any Linux or Windows machine with an NVIDIA GPU (and Windows PCs with an AMD Strix Halo iGPU, AMD and Intel Arc cards, and Macs) is a first-class unit, with separate RAM and VRAM panels, and a block per card on multi-GPU hosts.

Live LLM probe

Auto-detects llama.cpp, vLLM, SGLang, LM Studio, Ollama, NVIDIA Flare, Radiance, Strata, ds4-server, EXL3, TensorFold, FreeToken and q27, with live decode and prefill tok/s and a daily peak.

Inference health

KV cache, run and wait queue, TTFT, E2E and ITL p95, preemptions, prefix cache and MTP acceptance, from the server's own metrics.

Unified memory and RoCE

The GB10's 128 GB pool, with the GPU/CPU split and bandwidth from nvidia-smi dmon, and the RoCE links between your Sparks.

Integrations

An opt-in Prometheus /metrics, ComfyUI jobs and queue, Hermes Agent updates over SSH, a Tailnet check that flags a unit that left its tailnet, and a Glance widget.

Power controls

A graceful shutdown per Spark or for all of them, and Wake-on-LAN for GPU hosts whose network card supports it.

Alerts and history

Opt-in notifications to ntfy, Discord, Slack or a webhook, and 30 days of per-unit metrics history, both off until you turn them on.

Secrets stay secret

SSH passwords are encrypted with AES-256-GCM and never appear in the config file or in API responses.

Quick start

Running in two commands.

1

Clone and start it with Docker

git clone https://github.com/MiaAI-Lab/sparkDash.git
cd sparkDash
docker compose up --build -d

Then open http://127.0.0.1:5555 on that machine. Add the machine it runs on as This host, and every other Spark or GPU box with its LAN IP and an SSH key.

2

Open it from another computer

ssh -N -L 5555:127.0.0.1:5555 user@sparkdash-host

sparkDash listens on loopback by default. For shared access use an authenticated TLS reverse proxy or Tailscale Serve, or set BIND_HOST=0.0.0.0 together with SPARKDASH_TOKEN: without a token, anyone who can reach the port can change settings and power your machines off.

The README has the rest: SSH keys in Docker, Windows PCs, the phone app over Tailscale Serve, Prometheus, every setting, the REST API and the security notes.

Credits

Built with the community.

Thank you to everyone who sent code, fixes and ideas: @MikeGibbsOnyx, @Lesilva, @vincenzopalazzo, @Acermax, @danielkuykendall23-boop, @0xdfi, @BHCC2025, @ayylemao, @SashaMIT, @krunkosaurus, @andrei-dotdna, @0xWhiteMage, @kesslerio, @nkavassalis, @Olyno, @willy92wins, @saitakarcesme, plus Nyx Voss (who wrote the macOS collectors with C. Michael Gibbs), Liao Shiwu and Philip Eriksson, and everyone who reported issues. In 2.30, thanks to @HassanEmam for the phone app and @oscarlius for the FreeToken tests. In 2.40, thanks to @HassanEmam again for LM Studio and Ollama detection and AMD iGPU support on Windows. The Glance widget was contributed by @linxichen.

Built on

Free, and open.

  • License: Apache 2.0 from version 2.0.0. Versions up to 1.9.0 stay MIT.
  • tool-eval-bench by SeraphimSerapis (MIT) powers the Tool Eval Bench.
  • GSM8K (OpenAI) and MMLU (Dan Hendrycks) supply the Quality bench's maths and knowledge questions.
  • The health findings were inspired by spark-doctor by joeynyc; no code from it is used.
  • React, Vite, Tailwind CSS, Express and ws; fonts Geist and Geist Mono by Vercel.

See your whole lab at a glance.

Free and open source, for one Spark or a whole fleet.