A working engineer's prep notes for both certification tiers. Five HTML decks, ten technical notes, four cheatsheets, five runnable exercises, and seventy-one mock-exam questions — all cross-referenced to a wider portfolio of NVIDIA and LLM topic repos.
NVIDIA offers two certifications for engineers working with generative AI and large language models. Both are scenario-based, multiple-choice, and online-proctored. This site collects everything I'm using to prepare.
Foundations — transformer architecture, prompt engineering, RAG, fine-tuning, and trustworthy AI. Aimed at engineers with working knowledge of LLMs.
Production depth — inference optimisation, distributed training, PEFT, NVIDIA stack (NeMo, Triton, TensorRT-LLM, NIM). Numerical reasoning expected.
Who is this for? Primarily me, so I can sit both exams. Secondarily anyone studying for the same. The repo is public because the structure and cross-references may be useful.
Three paths through the material depending on what you want.
12 weeks at ~6 hours/week. NCA in weeks 1–6, NCP in weeks 7–12. Heavy weeks (3, 8, 9, 11) flagged honestly.
Visual overview of the high-leverage topics. Each deck is a self-contained HTML page; no build, opens straight from a browser.
41 NCA questions, 30 NCP questions, plus three open-ended system-design scenarios. Every question has an answer-key rationale.
Self-contained HTML presentations — dark theme, no build step, speaker-notes toggle, scrolling sections rather than click-through slides. Each is a cert-focused tour of one major topic area.
Decoder-only end to end — tokens, embeddings, attention, FFN, KV cache, decoding strategies.
Open deck → DECK 02From embedding model selection to production evaluation — hybrid search, reranking, agentic patterns, GraphRAG, RAGAS.
Open deck → DECK 03The post-training stack — SFT, LoRA/QLoRA/DoRA, RLHF, DPO, Constitutional AI, VRAM trade-offs.
Open deck → DECK 04KV cache, paged attention, in-flight batching, FP8/FP4 quantisation, speculative decoding, framework matrix.
Open deck → DECK 05Driver to NIM — CUDA, NCCL, NeMo, NeMo RL, TensorRT-LLM, Triton, Base Command, AI Enterprise, decision matrix.
Open deck →Notes and exercises live as Markdown in the repo — click through to GitHub where they render properly.
Both cert tiers mapped row-by-row to notes, exercises, and existing portfolio repos.
NCA weeks 1–6, NCP weeks 7–12. Heavy weeks called out explicitly.
~21,000 words. One per NCA/NCP domain. Each ends with "Likely exam angles" and primary citations.
Single-page references — transformer maths, quantisation + KV cache, sampling, NVIDIA stack.
Runnable code — BPE tokeniser, attention from scratch, LoRA, Triton, TensorRT-LLM. Written, not yet hardware-verified.
41 NCA Q + 30 NCP Q (6 numerical) + 3 system-design scenarios + 12 STAR behavioural prompts.
Domain weightings sourced from NVIDIA Learn in April 2026. Verify again at exam time — weightings can shift between exam revisions.
Core ML and AI Knowledge — 30%
Software Development — 24%
Experimentation — 22%
Data Analysis and Visualisation — 14%
Trustworthy AI — 10%
Model Optimisation — 17%
GPU Acceleration — 14%
Prompt Engineering — 13%
Fine-Tuning — 13%
Data Prep — 9%
Model Deployment — 9%
Evaluation — 7%
Production — 7%
LLM Architecture — 6%
Safety & Compliance — 5%
The cert content does not exist in isolation. Each note links into the deeper portfolio repos where the topic gets full treatment.
NVIDIA stack: LLM_Hub_NVIDIA_GPUs indexes 37 GPU/architecture decks. RAG: LLM_Hub_RAG_Retrieval + RAG_01–07. Fine-tuning: LLM_Hub_Fine_Tuning + FT_01–05. Evals: LLM_Hub_Evaluations + LLM_Eval_01–05. All in the LLMs hub.