LLM Cluster Viz

A 3D film about one ChatGPT reply and one training run — except you can walk into it.

An open-source, interactive visualization that uses three zoom levels — datacenter → NVL72 rack → chip & algorithm — to answer one question:

When you send "hello" to a large language model, what actually happens on the hardware?

Open the live demo → auxten.com/llm-cluster-viz

---

What it tries to explain

You've probably heard these words: NVLink, InfiniBand, KV cache, MoE, pipeline bubble, continuous batching…

Each one is a long blog post on its own. Put together, they're a textbook. This site takes a different approach:

• No equations first — just a 3D datacenter you can orbit around
• No jargon first — just data flying between racks in front of your eyes
• No block diagrams first — just a trip inside the HBM of a single GPU

By the end you'll have a gut-level "oh, so that's what it looks like" feeling for things like:

• Why does training a frontier model take thousands of GPUs, and still fail every 3 hours?
• Why did NVIDIA make "one machine" (NVL72) as big as 72 GPUs?
• Why does industry split inference into two different kinds of machines (prefill vs. decode)?
• How did Llama 3 survive 419 interruptions across 54 days on 16,384 H100s and still finish training?

---

What you'll see

The tour is 4 acts, 10 chapters — a short documentary. Each chapter is 30 seconds to 2 minutes. Watch front to back, or jump around.

Act 0 · Prologue — a tour of the hardware Zoom from a wide shot of the datacenter all the way down to a single HBM stack.

• P1 · Datacenter overview — 4 NVL72 racks, 288 Blackwell GPUs
• P2 · GB200 Superchip — 1 Grace CPU + 2 B200 GPUs + 8 HBM3e stacks

Act 1 · Training — how a model actually gets trained One full training iteration, from input batch to weight update.

• T1 · The whole training iteration end to end
• T2 · Pipeline bubble — why does a naive pipeline idle 38% of the time?
• T3 · MoE all-to-all — 256 experts across 64 GPUs, and the shuffle that makes it work

Act 2 · Inference — what happens when you send a prompt The full lifecycle of one request, from browser to streaming tokens.

• I1 · Request lifecycle (router → prefill → KV transfer → decode → SSE)
• I2 · Prefill internals — why the whole prompt computes in one shot
• I3 · Decode & continuous batching — how thousands of users share one GPU
• I4 · KV cache budget — why long context is expensive

Act 3 · Reality — when the ideal meets the physical world

• R1 · Failures & recovery — hardware breaks every 3 hours, and training still gets >90% effective time

---

How to watch

Just open the site. No sign-up, no GPU needed, nothing to install.

• Mouse / touch — orbit, pan, zoom the 3D scene freely
• Keyboard — ← / → or j / k to step, Space to pause/resume
• Language switcher (top right) — English, Chinese (Simplified / Traditional), Japanese, Korean, Spanish, Portuguese, Italian, French. Nine languages.
• Sandbox mode — prefer to explore without a script? Click "Enter sandbox" in the bottom right
• Mobile — on small screens some chapters fall back to 2D diagrams to keep things readable

Every step has a short caption. Every number you see on screen (130 TB/s, 419 interruptions, >90% effective time…) comes from an NVIDIA whitepaper or a published paper — citations are always one click away.

---

Who this is for

• Engineers, PMs, students who want to actually understand how large models run
• People preparing for (or giving) LLM-infrastructure interviews
• Anyone who watched a Jensen keynote and wondered what that rack actually is
• Anyone who needs a demo-ready deck for an internal "LLM infra 101" talk

You don't need CUDA, PyTorch, or distributed training experience. If you can use a mouse, you're qualified.

---

Run it locally

If you want to tweak it, add a chapter, or run it offline:

git clone https://github.com/auxten/llm-cluster-viz.git
cd llm-cluster-viz
npm install
npm run dev

Open http://localhost:5173. For a production build, npm run build — output lands in dist/.

Tech stack: React 19 + TypeScript + Three.js (via @react-three/fiber) + Tailwind + Vite. Pure frontend, no backend required.

---

Inspiration

Dwarkesh Patel's interview with Reiner Pope — a brilliant verbal walkthrough of LLM inference on real hardware. This site is the visual companion to that conversation, extended to cover training, MoE, and failure recovery.

Key references:

• NVIDIA GB200 NVL72 whitepaper and Blackwell architecture docs
• Meta · The Llama 3 Herd of Models (interruption statistics)
• DeepSeek-V3 technical report (MoE design, EP domain sizing)
• vLLM · Efficient Memory Management for LLM Serving with PagedAttention
• Mooncake · disaggregated prefill/decode serving in production
• GEMINI · fast failure recovery via in-memory checkpointing
• ReCycle · elastic pipeline takeover

---

License

MIT