Company interview guide
NVIDIA Machine Learning Interview Guide (2026)
Team-run loops that mix algorithm coding, deep learning built from scratch and GPU-aware LLM performance.
NVIDIA's machine learning interview loop has 8 reported stages, from recruiter call to behavioral and team fit; typical length: 3-12 weeks.
Roles covered: Deep Learning Engineer, ML Engineer, Research Scientist, AI Engineer and Data Scientist. The question bank holds 28 questions reported from NVIDIA interviews, 6 of them free to solve in your browser.
- Interview process length
- 3-12 weeks
- Stages
- 8
- Reports linked to bank questions
- 34
- Bank report dates
- Jul 2021–Sep 2026
- Questions in bank
- 28 (6 free)
- Last reviewed
- October 2026
How the NVIDIA interview loop works
Most reports run like this: a short recruiter or HR screen (often not mentioned), a first technical call with the hiring manager or a team member that centres on your projects, one or more coding rounds in a shared editor, on HackerRank or on a whiteboard, then a panel of 45-60 minute interviews with team members, the lead and the manager. Reported panels range from three to ten interviews, most often four to seven. NVIDIA's hiring page says at least one onsite interview at an office is required before an offer, and that interns usually have phone interviews only. Research scientist loops are built around a one-hour job talk and research one-on-ones, sometimes with no coding. A few older reports describe take-homes on algorithms, CUDA or parallel programming. Gaps of days or weeks between rounds are common, and feedback can be shared across teams if you apply to several.
The NVIDIA interview process, stage by stage
01Recruiter call
Short call, about 30 min where reported; some candidates go straight to the hiring manager
A short conversation about your background, the team behind the requisition and logistics. Only a handful of reports describe it as a separate step; many candidates are contacted by a recruiter and then meet the hiring manager first. Because hiring is per team, this is where you learn which group you are talking to.
What it tests
- How your experience maps to the specific team and role
- Why NVIDIA, and which products or groups interest you
- Logistics: timeline, location, level, other open applications
How to prepare
- Ask how many rounds there will be and what each one covers; candidates often report not being told.
- Apply to a small number of roles you fit well; NVIDIA's hiring page suggests three to five and notes that feedback can be shared across teams.
- Mention competing deadlines early, since scheduling tends to be slow.
4 candidate reports (2023-2025) describe a recruiter or HR screen before the manager call; many others only mention being contacted. The hiring page does not name this step.
02Hiring-manager or team screen
45-60 min video or phone call with the hiring manager, a director or a team member; usually technical
The most consistently reported round. The interviewer walks through your resume, picks a project and keeps asking why and how, then often moves into deep learning questions that match the team, such as recent transformer changes, parallel training or GPU memory. Some managers add one or two coding questions, from an easy Python exercise to algorithm mediums.
What it tests
- Depth on your own projects: your contribution, design choices and results
- Mechanism behind every technique on your resume
- Deep learning breadth relevant to the team (transformers, training, optimization)
- Fit with the team's current work
How to prepare
- Prepare to explain one project down to tensor shapes, loss choice and what failed.
- Read the team's recent papers, blog posts or product pages and connect your work to them.
- Expect a coding question even when the round is described as a chat.
22 reports describe a first technical call, 11 with the hiring manager or a director, centred on a project deep dive (14 since 2024); NVIDIA's hiring page also says you meet the manager and team by phone or video.
03Coding screens and coding rounds
1-2 rounds of 45-75 min, live in a shared editor, on HackerRank or a whiteboard, often one of them in the panel; rarely a take-home
Classic data-structure and algorithm problems, mostly medium with the odd hard one, often two or three problems in an hour. Dynamic programming, intervals, binary search, sliding window and linked lists come up repeatedly, and some screens ask you to run tests. Systems-leaning and senior roles may add C++, multithreading, code review, unit testing or benchmarking; a few older reports had a take-home on algorithms, CUDA or parallel programming instead.
What it tests
- Solving two or three medium problems inside one hour
- Dynamic programming, intervals, binary search, sliding window, lists and trees
- Bug-free code that passes tests you run yourself
- Language fluency: Python details, and C++ or multithreading for systems teams
- Reviewing and testing someone else's code (senior roles)
How to prepare
- Practise at a pace of about 20 minutes per medium problem; reports mention three mediums in an hour.
- Confirm the language before the round; some interviewers require a specific one and are strict on syntax.
- Do not rely on being told the round is not algorithmic; one candidate was told that and then got only algorithm problems.
- Unapproved tools such as chat assistants lead to disqualification.
25 reports describe algorithm-style coding (10 since 2024) and 4 more only a short easy Python exercise; the hiring page mentions coding on HackerRank, a whiteboard or a laptop.
04ML coding from scratch
45-60 min in a shared editor or notebook, Python and numpy, no framework layers
Write standard deep learning building blocks yourself: a 2D convolution, multi-head attention, softmax, dropout or batch norm, sometimes with the backward pass of a small MLP. Research loops have also asked candidates to debug ML code in a notebook.
What it tests
- Implementing conv2d with stride, padding and channels correctly
- Multi-head self-attention with correct reshapes and masking
- Numerically stable softmax and cross-entropy
- Train versus inference behaviour of dropout and batch norm
- Forward and backward passes worked out by hand
How to prepare
- Write conv2d and attention from memory with shape comments on every line.
- Subtract the max before exponentiating and say why out loud.
- Know exactly which statistics batch norm uses at inference and how dropout scales activations.
Free reported questions for this stage
- Multi-Head Causal Self-Attention From ScratchML CodingHardSolve
- Numerically Stable Log-Sum-Exp and SoftmaxML CodingMediumSolve
8 reports ask for hand-written DL code (attention, conv2d, dropout, batch norm, softmax, tensor ops, notebook debugging), 6 since 2024; it sits in the panel or the manager call.
05Deep learning fundamentals and GPU performance
One or more 45-60 min panel conversations, often with a whiteboard
Rapid, deep questioning on how networks train and why they fail, then on how training and inference map onto GPUs. Expect follow-ups on optimizers and normalization, distributed training, numeric precision and, for hardware-close teams, why a CUDA kernel is slow.
What it tests
- Optimizers, normalization, regularization, activations and vanishing gradients
- Data, tensor, pipeline and sharded parallelism and their memory costs
- FP32, FP16, BF16 and lower formats: range, precision and failure modes
- Reasoning about memory bandwidth, coalescing and occupancy in GPU code
- Diagnosing data problems such as noisy labels and confusable classes
How to prepare
- For every optimizer or normalization method, be ready for three levels of why.
- Be able to compute the memory for weights, gradients and Adam state per GPU under each parallelism scheme.
- If the team writes kernels, review CUDA basics: memory hierarchy, coalesced access, warps and reductions.
25 reports describe DL or ML theory questions, the most consistent technical topic after coding; 9 of them add GPU, precision or distributed-training questions.
06ML / LLM system design
45-60 min discussion in the panel or a technical screen, often with back-of-envelope maths
Candidate reports describe open design prompts: explain a decoder-only LLM, tune it for a domain and control hallucinations in production, train an LLM for a set of tasks, retrieve over very large documents, design a benchmarking system or build a recommender end to end. For LLM inference and performance teams, public compiled question lists add serving maths: KV cache sizing, prefill versus decode, batching, quantization and splitting very large models across GPUs. Those specifics are not yet confirmed by individual reports.
What it tests
- Explaining a decoder-only LLM and how you would adapt and monitor it for a domain
- Designing a retrieval, benchmarking or recommender system end to end
- KV cache size and GPU memory budget for a given model and load (inference teams)
- Prefill versus decode, continuous batching and paged KV memory (inference teams)
- Quantization formats and how to split a very large model across GPUs
How to prepare
- Have a clear plan for adapting an LLM to a domain: data, fine-tuning without forgetting, evaluation and hallucination monitoring.
- If the team works on inference, memorise the KV cache formula and practise a roofline estimate of tokens per second for a 70B model.
- State your units and assumptions on every estimate so the interviewer can follow the arithmetic.
Free reported questions for this stage
- What the KV Cache Stores and Why It Speeds Up DecodingLLMs & GenAIMediumSolve
- GPU Memory Budget for Serving a 70B ModelLLMs & GenAIMediumSolve
- Roofline Estimate: 70B Decoding Speed on One GPULLMs & GenAIMediumSolve
6 reports mention a design question (LLM, retrieval, benchmarking, recommender); serving-maths topics rest on a compiled public question list, not first-hand reports.
07Research talk and research one-on-ones
Research scientist track: a 1-hour job talk, then 3-6 one-on-ones with team researchers
Research loops are built around your work. After the talk, each interviewer probes your methods, motivation and where you think the field is heading; some loops add basic ML questions, ML debugging or distributed training, and others have no coding at all.
What it tests
- Clarity and rigour of your research presentation
- Why you chose your problems and methods
- Views on future research directions and what you would do at NVIDIA
- Breadth across adjacent areas such as LLMs, diffusion or reinforcement learning
How to prepare
- Rehearse the talk to fit well inside an hour with time for questions.
- Prepare a concrete research agenda that fits the team's area.
- Reports note that a strong publication or internship record matters for these roles.
9 research-role reports describe rounds about your own research and future directions; 4 include a job talk of about an hour. One 2025 loop had no coding at all.
08Behavioral and team fit
Sometimes 1 round with the hiring manager or team lead, 45-60 min; often mixed into technical rounds
Some loops have a dedicated behavioral round with the hiring manager, in one report on NVIDIA's values; in others fit is judged throughout, a promised behavioral round never happens, and the culture interview turns technical. Expect questions on why NVIDIA, what you want to work on there, collaboration and how you handle hard problems.
What it tests
- A specific answer to why NVIDIA and why this team
- Collaboration, ownership and working across groups
- How you approach an unfamiliar technical problem
- Intellectual honesty about what you did and did not do
How to prepare
- Prepare three or four stories with clear personal contributions and measurable results.
- Know the company's products relevant to your team, such as its GPUs, CUDA libraries and inference software.
- Keep a technical answer ready for every behavioral story; follow-ups go deep.
11 reports describe behavioral or fit questions or a culture round, only 4 since 2024; one 2025 candidate was promised a behavioral round that never came.
Questions and themes reported most often
- Deep dive into a resume project with probing follow-ups (30 reports, hiring-manager or team screen)
- Timed algorithm problems: DP, intervals, binary search, lists, trees (25 reports, coding screens and coding rounds)
- Deep learning theory: activations, regularization, vanishing gradients, optimizers (12 reports, deep learning fundamentals and gpu performance)
- CUDA or GPU optimization questions, sometimes as a take-home (6 reports, deep learning fundamentals and gpu performance)
- How distributed and parallel training works in detail (5 reports, deep learning fundamentals and gpu performance)
- Research job talk followed by one-on-ones with the team (4 reports, research talk and research one-on-ones)
- Implement softmax, dropout, batch norm, cross-entropy or Adam by hand (3 reports, ml coding from scratch)
- Explain a decoder-only LLM, tune it for a domain and handle hallucinations (3 reports, ml / llm system design)
What NVIDIA asks about
- General coding28%
- DL & ML fundamentals24%
- Project deep dive & behavioral20%
- ML coding from scratch14%
- LLM, GPU & system design14%
NVIDIA interview prep checklist
- Solved three medium algorithm problems in one hour without help
- Wrote conv2d and multi-head attention from scratch with correct shapes
- Can explain optimizers, normalization and vanishing gradients three levels deep
- Computed the KV cache and memory budget for serving a 70B model
- Can compare data, tensor and pipeline parallelism and their memory costs
- Prepared a project deep dive including what failed and what you would change
- Have a specific answer to why NVIDIA and why this team
Tips
- Find out the team early and tailor prep: kernel and inference teams lean on GPU performance, product and research teams on modelling depth.
- Keep timed algorithm practice going all the way through; it appears in most loops regardless of title.
- Write conv2d, multi-head attention, stable softmax, dropout and batch norm in numpy until each takes under 15 minutes.
- Practise back-of-envelope maths for KV cache size, GPU memory per parallelism scheme and decode throughput.
- Know the trade-offs of FP16, BF16, FP8 and INT4 and how a model can break when precision drops.
- Prepare one project you can defend at any depth; the hiring manager will choose what to dig into.
- Ask each interviewer's focus at the start of the round; formats vary and are not always communicated.
- Plan for slow scheduling and follow up politely with your recruiter if a week passes without news.
See the rest of the NVIDIA loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the NVIDIA loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
NVIDIA interview FAQ
How many interview rounds does NVIDIA have for ML roles?
Candidates report 8 stages: Recruiter call, Hiring-manager or team screen, Coding screens and coding rounds, ML coding from scratch, Deep learning fundamentals and GPU performance, ML / LLM system design, Research talk and research one-on-ones and Behavioral and team fit. Most reports run like this: a short recruiter or HR screen (often not mentioned), a first technical call with the hiring manager or a team member that centres on your projects, one or more coding rounds in a shared editor, on HackerRank or on a whiteboard, then a panel of 45-60 minute interviews with team members, the lead and the manager. Reported panels range from three to ten interviews, most often four to seven. NVIDIA's hiring page says at least one onsite interview at an office is required before an offer, and that interns usually have phone interviews only. Research scientist loops are built around a one-hour job talk and research one-on-ones, sometimes with no coding. A few older reports describe take-homes on algorithms, CUDA or parallel programming. Gaps of days or weeks between rounds are common, and feedback can be shared across teams if you apply to several.
How long does the NVIDIA ML interview process take?
Reported timelines run 3-12 weeks.
What does NVIDIA ask in ML interviews?
By reported question mix: General coding 28%; DL & ML fundamentals 24%; Project deep dive & behavioral 20%; ML coding from scratch 14%; LLM, GPU & system design 14%. The most repeated themes: Deep dive into a resume project with probing follow-ups; Timed algorithm problems: DP, intervals, binary search, lists, trees; Deep learning theory: activations, regularization, vanishing gradients, optimizers.
How should I prepare for the NVIDIA ML interview?
Find out the team early and tailor prep: kernel and inference teams lean on GPU performance, product and research teams on modelling depth. Keep timed algorithm practice going all the way through; it appears in most loops regardless of title. Write conv2d, multi-head attention, stable softmax, dropout and batch norm in numpy until each takes under 15 minutes.
Are there free NVIDIA interview questions?
Yes. LeetML has 28 questions reported from NVIDIA interviews, and 6 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- IntelResume-driven technical conversations on ML fundamentals, light coding, and C++ and hardware awareness for performance teams.
- QualcommClassic algorithm coding, ML and deep learning fundamentals, and a resume deep dive, with C/C++ and hardware questions on on-device teams.
- GoogleClassic algorithm rounds plus ML design, ML domain depth and Googleyness, judged by a hiring committee.
- MetaFast classic coding, an AI-assisted coding project and product ML design for recommendations and integrity.
- MicrosoftProject deep dives in most rounds, quick ML and LLM fundamentals, standard coding, from-scratch ML code and GenAI design, in a loop shaped by the hiring team.