Skip to content

dingn42/EasierInfra

Repository files navigation

EasierInfra

An interactive, game-based way to learn GPU and LLM infrastructure. It contains 40 small browser games across 6 topics; each game teaches one concept through a short animation with synchronized code highlighting and a short quiz. The interface is bilingual (English and Chinese).

Live demo: https://easierinfra.vercel.app

Topics (40 games)

  • CUDA foundations — how the GPU runs your code Warp March · Coalesce Dock · Race & Lock · Latency Kitchen · Bank Run · Reduction Relay · Thread Grid · Warp Whisper · Wide Load Dock
  • Kernels & performance — making a kernel fast Roofline Ridge · Fusion Forge · Tile Smasher · Overlap Yard · Replay Roller (CUDA Graphs)
  • LLM internals — what a model does inside The Attention Spotlight · The Dice Workshop · FlashAttention · Spin the Clock (RoPE) · Expert Routing Floor (MoE) · Draft & Verify · Share the Cache (GQA)
  • LLM serving — serving a model at scale Continuous Batching · VRAM Landlord · Prefill vs Decode · Paged Stacks · Prefix Pantry · KV Cache Cellar · Prefill Slicer
  • Distributed training — scaling across many GPUs Ring Relay · Split the Model (TP/PP) · Recompute Ranch · Share the Load (ZeRO/FSDP) · Batch on Layaway · Tree vs Ring · Expert Airlift
  • Data & systems — data pipelines and memory Prefetch Pit Crew · Scale Keeper · FP8 Foundry · Hide the Wire (NVLink) · The Block Lender

Features

  • A brief mission description, a step-by-step animation, and a predict-then-reveal question per game.
  • Code that highlights line by line in sync with the animation.
  • Run / Step / Reset controls plus a game-specific interaction (slider, toggle, or drag).
  • A "in real CUDA" panel that maps each game back to the real concept.
  • A glossary with 130+ terms (short definition, full explanation, and an in-game mapping).
  • Bilingual UI (English / Chinese) and locally saved progress.

Numbers are scaled down for clarity (a warp is drawn as 4 lanes, not 32). The idea is the same at full scale.

Run locally

npm install
npm run dev      # http://localhost:5173
npm run build    # tsc --noEmit && vite build

License

MIT

许可证

MIT

About

Learn GPU & LLM infrastructure by playing

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages