Delete Duplicate Folders in System (LeetCode 1948)

The key insight is to represent the entire folder hierarchy as a trie, then serialize each subtree into a canonical string and use a hash map to find duplicate subtrees. Any subtree whose serialized form appears more than once marks all its nodes for deletion, after which a second traversal reconstructs the surviving paths. Time complexity is O(N * L) where N is the number of paths and L is the average path length. A strong follow-up is: how would you handle case-insensitive file systems, and what changes if deletion must preserve the root? Comfort with trie construction and subtree hashing; ability to recognize that equality of folder structure (not just name) is the invariant. Tests whether the candidate can navigate a hard problem incrementally and articulate edge cases such as nested duplicates and single-child subtrees.

LRU Cache (LeetCode 146)

The canonical approach combines a doubly linked list with a hash map to achieve O(1) get and put. The list maintains recency order (head = most recent, tail = least recent); the map stores key-to-node pointers. On a get, move the accessed node to the head. On a put, insert at head and, if over capacity, evict from tail. A strong candidate names the invariant explicitly before coding. Follow-up: implement LFU cache (LeetCode 460), or extend to a distributed LRU across shards. Clean separation of the pointer manipulation from the map logic; no off-by-one errors on the sentinel dummy nodes. Evaluates design thinking and ability to maintain invariants under concurrent access, a likely follow-up for Nvidia's systems roles.

Number of Islands (LeetCode 200)

Standard BFS or DFS flood-fill: iterate the grid, and for each unvisited land cell trigger a traversal that marks all connected land cells as visited, incrementing the island count once per traversal. Time and space are both O(M*N). A strong candidate discusses the trade-off between recursive DFS (stack overflow risk on large grids) and iterative BFS, and mentions union-find as an alternative that supports incremental updates. Follow-up: what if the grid is distributed across machines? Correct boundary checks, visited-marking discipline, and awareness of the iterative vs recursive trade-off. Nvidia interviewers frequently push the distributed or streaming variant to test systems thinking beyond the core algorithm.

Design a Contact PhoneBook service: support storing contacts with name, multiple phone numbers, and address; support search by name

Clarify read vs write ratio (likely read-heavy) and expected scale. A strong design uses a relational store (contacts + phone_numbers tables) for structured data with a name-indexed inverted index or prefix trie for fast prefix search. For type-ahead, a Redis sorted set with prefix scoring works at scale. Discuss: pagination on search results, case-insensitive and phonetic matching (Soundex), and eventual consistency if the service is distributed. Draw the API surface (createContact, searchByName, updateContact) before diving into storage. Clear requirements gathering before jumping to a solution; explicit API and data model definition; awareness of search indexing vs storage separation. Tests structured thinking and the ability to scope a problem appropriately.

Design a system to handle petabytes of data coming from multiple machines and reduce bottlenecks in data ingestion and processing

This is a distributed data pipeline design. Start with the ingestion layer: a message bus (Kafka) partitioned by machine/source to parallelize ingest. Then a processing tier: stateless stream processors (Flink or Spark Streaming) that fan out to compute nodes. Storage tier: columnar storage (Parquet on object store or a data lake) for cold analytics, a time-series store (InfluxDB or Druid) for hot queries. Key bottlenecks to address: head-of-line blocking in the message bus (replication factor, partition count), network bandwidth saturation (compression, batching), and storage hotspots (consistent hashing for data locality). Hardware awareness is expected at Nvidia: discuss RDMA, NVLink topology, and GPU-direct storage for accelerated pipelines. Whether the candidate identifies the bottleneck layers (network, I/O, compute) before proposing solutions; hardware-aware trade-offs appropriate for Nvidia's infrastructure context. Senior candidates should connect the design to Nvidia's actual product stack.

Design a chatbot service that provides users with a variety of information (reported by multiple Nvidia candidates on Glassdoor and via interviewing.io)

Define the chatbot's scope: retrieval-based vs generative, domain-specific vs general. For a production system at Nvidia scale: a frontend API gateway routes queries to an NLU/intent classification layer; matched intents fan out to specialized knowledge-base retrievers (RAG with a vector store); a response synthesizer (LLM) merges results and formats the answer; conversation state is persisted in a session store (Redis). Discuss: latency budget (streaming tokens vs batch), caching repeated queries, abuse/safety filtering, and how to version the underlying models without downtime. Follow-up: how do you evaluate chatbot quality at scale? Ability to decompose a vague open-ended prompt into concrete components; evidence of familiarity with modern NLP infrastructure. At Nvidia specifically, examiners look for awareness of where GPU acceleration enters the picture such as inference serving and embedding generation.

Why do you want to work at Nvidia? (reported across multiple Glassdoor and IGotAnOffer candidate write-ups as a near-universal opener)

A strong answer is specific to Nvidia's technical position: connect the candidate's work (e.g., GPU programming, AI infrastructure, autonomous systems) to a concrete Nvidia product line or research area. Name a specific technology (CUDA, Triton, NIM, the Grace Hopper architecture) rather than giving a generic AI-hype answer. Conclude by connecting Nvidia's Innovation and Excellence values to the candidate's own demonstrated drive. Interviewers probe with: what do you know about our specific team's work? Genuine technical curiosity about Nvidia's products versus generic enthusiasm; specificity signals preparation and intellectual honesty, one of Nvidia's named core values.

Tell me about the most technically complex project you have worked on. (reported verbatim on Glassdoor and by multiple IGotAnOffer sources)

Use STAR but weight heavily on the Technical and Action components. Describe the system architecture briefly, then focus on the hardest constraint or failure mode you encountered and the specific decisions you made to resolve it. Quantify the outcome (latency improvement, scale, cost reduction). Be prepared for deep follow-up: Why did you design it that way? What would you change now? How did you balance performance and maintainability? Nvidia interviewers probe to distinguish candidates who built something from candidates who merely observed it. Depth of ownership and willingness to acknowledge trade-offs or mistakes (Intellectual Honesty value); ability to articulate system-level thinking and quantify impact (Excellence and Determination).

Tell me about a time you had a conflict with a teammate or disagreed with your team's technical direction. How did you handle it and what was the outcome? (reported on Glassdoor and by Exponent Nvidia guide)

Choose a real technical disagreement (not a personality clash) where you had a defensible position backed by data. Describe the disagreement clearly, then explain how you gathered evidence (benchmark, prototype, customer data) to make the case objectively. If you were overruled, describe what you learned. If you prevailed, describe how you brought the team along without friction. Nvidia explicitly values Intellectual Honesty and One Team, so the ideal answer demonstrates that you can advocate for a position while genuinely updating on new evidence. Maturity in handling disagreement without politics; evidence of data-driven reasoning; ability to distinguish disagreements about facts from disagreements about values. Probes directly for Intellectual Honesty and One Team alignment.

Nvidia Interview Questions

Nvidia's SWE interview loop typically runs 6-8 weeks and consists of a recruiter screen, an online assessment (OA) with 2 coding problems plus aptitude MCQs, a phone/technical screen, and a full onsite of 4-5 rounds covering coding, system design (for senior roles), a domain or team-specific round, and a hiring manager behavioral round. The process is notably decentralized: there is no internal question bank, so individual interviewers set their own questions, and rounds vary by team. Most coding questions are LeetCode medium difficulty; roles that require CUDA work expect C++. The onsite may include a code-review or debugging round in addition to or instead of standard algorithm problems, and senior roles consistently include a 1-hour system design round.

18 real questions across 3 rounds

Reported via Questions sourced from: GeeksforGeeks Nvidia SDE Internship On-Campus 2024 experience (geeksforgeeks.org/interview-experiences/nvidia-interview-experience-for-sde-internship-on-campus-2024/) for Delete Duplicate Folders; GeeksforGeeks Nvidia System Software Engineer experience (geeksforgeeks.org/nvidia-interview-experience-for-system-software-engineer/) for the XOR odd-count question; Medium Nvidia SDE-2 interview experience by Hardik Dadhich (medium.com/@dadhichhardik26/nvidia-sde-2-interview-experience-offer) for the server log parsing question; interviewing.io 2024 Nvidia interview page (interviewing.io/nvidia-interview-questions) for the beam search question; CodeJeet Nvidia LeetCode tag (codejeet.com/company/nvidia) and krishnadey30/LeetCode-Questions-CompanyWise nvidia_alltime.csv for LRU Cache, Number of Islands, and Search in Rotated Sorted Array frequency data.

6 more coding questions Nvidia actually asks, each with how to approach it and exactly what the interviewer is evaluating. Unlock the full set across every round.

Continue with Google

Or run a Nvidia mock with the AI interviewer. It asks questions like these, follows up on your answers, and tells you exactly what to fix. The only score that means anything is the one on your real answers.

Continue with Google

Keep prepping

All company interviews