Let’s see what happens after you send a prompt. Explore the machines it runs on, follow the work inside a GPU, and learn how one chip becomes part of an AI cluster.
No background needed. Explore at your own pace.
Choose a machine, drag it around, and select a part to see what it does. Try “Guide me” for a short tour.
Go back a layer, change your view, or follow one illustrative operation.
A whole system. Select an assembly, then step inside.
Pick a starting point. You can move between these whenever you like.
TokenQuest is independent of NVIDIA. The practice questions are written for this site.
Explore 8 production AI infrastructure concepts right here: PagedAttention, the Roofline model, Tensor vs. Pipeline Parallelism, Speculative Decoding, Disaggregated Serving, MIG partitioning, GPUDirect RDMA, and FlashAttention SRAM tiling.
Contiguous KV reservation pre-allocates the maximum context length per request, wasting memory on unused future slots. PagedAttention splits the KV cache into fixed-size blocks mapped through a block table so new requests can join as soon as blocks free up.
Preparing for an AI Infrastructure, ML Systems, or Kubernetes GPU Platform role? Try our live Whiteboard GPU Sizer, diagnose production DCGM & NCCL incidents, and review senior interview flashcards while we build out the full 4-track simulator.
Size GPU clusters for 8B to 405B+ models, balance TTFT vs. TPOT SLAs, and design PagedAttention, speculative decoding and disaggregated prefill/decode pipelines.
Architect production Kubernetes GPU clusters with NVIDIA GPU Operator, Dynamic Resource Allocation (DRA), Topology Manager, MIG partitioning and gang scheduling.
Design non-blocking rail-optimized InfiniBand and RoCEv2 fabrics, trace NCCL ring/tree collectives, and eliminate PCIe root-complex bottlenecks with GPUDirect RDMA and Storage.
Diagnose real-world cluster slowdowns using DCGM Exporter metrics, XID hardware fault codes, thermal clock throttling, straggler detection and NCCL watchdog traces.
Jump directly into any of the six interactive quest boards, calculate real 2×2 matrix tiles in the hardware compute lab, or run the 3-lane Token Dash arcade highway.
Follow one request with Mira. Watch each step, try it yourself and see what changes inside the GPU.