AI Infra Interviews
Work through the numbers behind GPU systems, from a single model to a training cluster. Learn to explain your decisions and handle the follow-up.
130 answers to read now · 260 with a free account · no card
A question to start with
Change the model or context. Watch the memory budget change.
KV/token = 2 × 80 layers × 8 KV heads × 128 × 2 B
Memory estimate only. Actual serving capacity depends on parallel layout, runtime overhead and latency targets.
GPU LAB / HANDS-ON GPU COMMAND EMULATOR
Run inference and training workloads. Watch memory fill, follow live GPU graphs, then investigate a broken link or a service that refuses to start. Practice NVIDIA commands in your browser, with guided missions from your first terminal session to cluster recovery.
3 free missions · 21 Premium missions · no hardware to rent
Change the workload. Read the machine.
Run a workload on GPU 0. Follow its activity and memory, pause the clock, then switch workloads to compare.
No account or GPU required. Starts only when you press Run.
Find your way around Linux, identify GPUs and separate activity from allocation.
02 / PREMIUMRead configuration, budget memory and verify the address your client uses.
03 / PREMIUMTrace device placement, inspect the fabric and recover a failed endpoint.
04 / PREMIUMCorrelate NVLink, InfiniBand and ECC evidence. Separate current device state from historical errors.
05 / PREMIUMRepair configuration, trade context for concurrency and verify placement without disrupting neighboring workloads.
06 / PREMIUMDiagnose several independent blockers, verify the complete local path and leave a useful handoff.
Browser-based GPU command emulation. CUDA execution and hardware performance testing require real GPUs.
Learn what the machine does, what the model stores, and why an answer can be slow. Then try a memory budget and choose your first lesson.
Start learning the basics →Five steps · free to read · no account or GPU needed
Build it step by step
Train a tiny model, compare fine-tuned adapters, or configure and tune a private model service. Follow the commands and inspect what each step produces.
Your first training run
Free to readBuild a 470,528-parameter causal language model on a CPU. Train it, inspect validation loss, save a checkpoint and resume the same run.
CPU · Python 3.12 · no GPU required
A trained byte-level model, a resumable checkpoint and sampled text.
CPU workflow executed
CPU selection checks executed · GPU run pending
Experiment client tested · GPU run pending
Local TLS and proxy executed
CPU and HTTP fixture executed
See how we teach
What happens between your prompt and the next token? Follow the loop in this diagram from our inference platform walkthrough.
Free to read · no account needed
Open the answer. Check the reasoning.
GPU & Accelerator Architecture
Hardware, Cabling & Cluster Build-Out
CUDA, Triton & Kernel Engineering
Distributed Training & Parallelism
A free account gives you 260 answers, bookmarks and saved progress.
Prepare for the work you want to do
Live CUDA or Triton, a kernel take-home, the roofline, FlashAttention internals.
NVIDIA CUDA teams, Fireworks, Together, lab performance teams02Parallelism strategy, collectives and fabrics, MFU, checkpointing, what breaks at 10k GPUs.
Frontier labs, xAI, Crusoe, hyperscaler training teams03The KV cache, batching, speculation, disaggregation, and the design round with TTFT on the board.
OpenAI, Anthropic, Baseten, Together, Fireworks, Red Hat04Kubernetes GPU scheduling, MIG, Slurm, DCGM and XID codes, the on-call reality.
CoreWeave, Modal, Nebius, Lambda, ML platform teamsBuild the foundations
Follow the lessons in order, then put the ideas to work.
Explore all 6 courses →15 lessons · 3 free to read
15 lessons · 5 free to read
16 lessons · 2 free to read
Go deeper when you’re ready
Work through the full bank, study the underlying concepts, and follow a course from the foundations into system design.
One payment. 6 months of access. No auto-renewal.
Full access
Start with a free account and decide after you’ve studied.
Focus your preparation
Explore reported interview formats and relevant topics. Each guide distinguishes documented details from limited public evidence. Loops vary by team; confirm yours with the recruiter.
Independent preparation. No affiliation with or endorsement by the companies named.
AI INFRA CAREERS · FREE
Explore GPU, training, inference and platform openings at companies we track. Filter by specialty and location, then prepare for that company.
Guides & handbooks
New to AI infrastructure? Start with the field guide, then follow a model from one GPU to a serving fleet. Keep the hardware and model guides nearby when you need to compare your options.
Start here / Included with Premium
Start with what lives in GPU memory and how a model produces an answer. The AI Infrastructure Field Guide builds from these basics to multiple accelerators, distributed training and production systems.
71 pages · 19 visual explanations · 269 glossary terms. Follow a beginner, intermediate or advanced reading route.
Samples are free to read. The complete PDF requires paid Premium or approved complimentary access. Referral Premium excludes paid guide downloads.
Inside the guide
Explore the sample pages →
Serving systems / Included with Premium
Follow a request through a serving system in The LLM Inference Systems Design Handbook. Work out how many users fit in GPU memory, see how batching changes latency, and learn when to add more GPUs. One running example connects the calculations.
60 pages · 17 chapters · 17 diagrams · 8 mock design rounds · 42 self-check questions.
Samples are free to read. The complete PDF requires paid Premium or approved complimentary access. Referral Premium excludes paid guide downloads.
Inside the guide
Read these samples as text →
Distributed serving / Included with Premium
Once you understand one serving engine, learn how a cluster routes requests and shares the work. Distributed Inference on Kubernetes follows those decisions through NVIDIA Dynamo and llm-d, with capacity calculations and failure scenarios.
100 pages · 25 chapters · 23 diagrams · 12 mock design rounds · 69 self-check questions.
Samples are free to read. The complete PDF requires paid Premium or approved complimentary access. Referral Premium excludes paid guide downloads.
Inside the guide
Explore the sample pages →
The free hardware reference
Place desktop cards, server GPUs and connected racks on one map. Learn what the names mean, then compare memory and work out what your model needs.
37 hardware profiles · 33 model references · 34 cloud configurations
Memory, roofline and training calculators →A few starting points
| Accelerator | Memory | TB/s |
|---|---|---|
| NVIDIA H100 80 GB SXM | 80 GB | 3.35 |
| NVIDIA H200 SXM | 141 GB | 4.8 |
| NVIDIA B200 HGX | 180 GB | 8 |
| Google TPU v6e (Trillium) | 32 GB | 1.638 |
Published memory bandwidth per device, not application throughput. GB and GiB retain the source units. Each profile links to its sources.
Take the reference with you
September 2026 · 47 pages · Comparisons, worked examples and linked sources. Free PDF with an account.
The free model reference
Compare 39 open-weight checkpoints. Start with plain explanations of parameters and quantization, then compare release timelines, architectures and documented hardware paths.
Included with paid Premium
51 illustrated pages: beginner explanations, architectural changes, lab timelines and deployment notes. Preview three pages below before you decide.
Model cards are free. The PDF download requires sign-in and paid Premium or approved complimentary access. Referral Premium excludes paid guide downloads.
Learn here. Help someone else start.
Know someone learning AI infrastructure? Share a useful resource on both LinkedIn and X and apply to become a Community Ambassador.
Explore the ambassador program →Recognition for helping others learn.
After team approval: a gold badge beside your comments and a one-time 48-hour Premium pass. Already Premium? Earn the badge and save the pass.
Keep exploring
