NVIDIA Interview Questions — Page 15

NVIDIA·Product Manager·Hiring Manager Screen

Jun 2026

PM screen at Nvidia that leaned pretty hard into data and AI infrastructure knowledge. Two questions, both circling around LLM training data, which I wasn't fully prepped for coming from a more traditional product background.

What do you know about data annotation and how data is collected for training large language models?
Walk me through something you've done that demonstrates experience with data collection or shaping a company's data strategy.

“I knew enough to not embarrass myself but not enough to impress anyone.”

View Post

NVIDIA·Software Engineer·Onsite - System Design / Architecture

Jun 2026

NVIDIA software engineer interview that went deep on ML systems knowledge, specifically around optimization techniques and distributed training strategies. Pretty technical throughout, felt more like a systems design conversation than a standard coding screen.

Walk through the main AI optimization techniques used in training and inference. Cover the goals behind each, how they work in practice, and the trade-offs involved. Things like quantization, pruning, knowledge distillation, kernel fusion, and memory or throughput optimizations.
Compare tensor parallelism and pipeline parallelism. How does each one actually work, what does the communication pattern look like, when would you choose one over the other, and where do the performance bottlenecks typically show up?

“This is a big question and I tried to structure it by separating training-time from inference-time concerns, which helped.”

View Post

NVIDIA·Software Engineer·Technical Phone Screen

Jul 2026

Got a low-level C++ memory question at NVIDIA that looked simple on the surface but required knowing struct padding rules cold. Not a lot of context in what I have, but it was the kind of thing where if you haven't thought about alignment before, you're just guessing.

Given a class with members `char buf[BUFF_SIZE]` (where BUFF_SIZE is 1), `size_t length`, and `char *ptr`, what is the result of `sizeof` on that class? Explain using struct alignment and padding.

“This one tripped me up because I jumped straight to adding up the raw sizes: 1 byte for the char array, 4 bytes for size_t, 4 bytes for the pointer on 32-bit.”

View Post

NVIDIA·Software Engineer·Technical Phone Screen

Jul 2026

Nvidia SWE interview with a classic in-place array sorting problem. Nothing too wild but the constraints matter more than people expect.

Given an array of n elements each representing one of three colors (red, white, or blue), sort the array in-place so all same-color elements are grouped together in the order red, white, then blue.

“The naive approach works but they clearly wanted something better.”

View Post

NVIDIA·Software Engineer·Technical Phone Screen

Jul 2026

NVIDIA software engineer interview with a pretty meaty coding problem around JSON traversal. The question had a lot of moving parts and I don't think I covered everything cleanly, but it was an interesting problem to work through.

Given a deeply nested JSON document (roughly five levels deep), write code to traverse it and extract specific fields. Your solution needs to handle missing keys, arrays mixed with objects, and unknown nesting depth. Then compare a recursive approach to an iterative one, discuss time and space complexity, and talk through how you'd handle errors and validate the schema.

“This is the kind of question that looks manageable until you actually start writing code.”

View Post

NVIDIA·Software Engineer·Onsite - System Design / Architecture

Jul 2026

NVIDIA system design round for a software engineer role, heavily focused on compiler internals. The question was brutally specific and I spent most of the time trying to remember what SSA even stood for before getting into the actual design.

Design a shader compiler that takes GLSL or HLSL as input and outputs SPIR-V or native ISA. Walk through the full pipeline: lexing, parsing to AST, conversion to SSA-based IR, optimization passes like constant folding, CSE, and dead-code elimination, then register allocation and final code generation. Also describe how you'd test such a system using differential and property-based testing.

“This was one question but it's basically five questions stapled together.”

View Post

NVIDIA·Data Scientist·Onsite - System Design / Architecture

Jul 2026

NVIDIA data scientist interview that went deep into system design territory, specifically around building a production-grade REST API for an image inference service. Way more backend-heavy than I expected for a DS role, and it covered a lot of ground in one question.

Design a REST API for an image inference service that handles large image uploads, returns class probabilities and Grad-CAM heatmaps, supports async processing with job IDs and webhooks, batching, pagination, rate limiting, OAuth2/JWT auth, versioning, retries, circuit breaking, and a well-defined error taxonomy. Also address input validation, content-type enforcement, secure storage, and backward compatibility across model upgrades and rollbacks.

“This was basically a full system design session crammed into one prompt.”

View Post

NVIDIA·Software Engineer·Onsite - System Design / Architecture

Jul 2026

NVIDIA system design round focused almost entirely on GPU internals. One big question that sprawled into several follow-ups about memory, kernel design, and performance modeling. Felt more like a grad seminar than a typical SWE interview.

Walk me through the GPU memory hierarchy. How do registers, shared memory, L1/L2 cache, and HBM differ in terms of latency, bandwidth, capacity, and how much control a programmer has over them?
How does the GPU memory hierarchy motivate designs like tiled GEMM? Walk through why naive matrix multiplication is inefficient and what tiling actually buys you.
FlashAttention avoids materializing the full attention matrix in HBM. Why does that matter, and how does the memory hierarchy drive that design choice?
What is arithmetic intensity and how do you use a roofline model to decide whether a kernel is compute-bound or memory-bound?

“This is the kind of question where you think you know it until you're actually talking and realize you're fuzzy on the numbers.”

View Post

NVIDIA·Software Engineer·Technical Phone Screen

Jul 2026

NVIDIA technical phone screen for a software engineer role. The focus was entirely on tree reconstruction, which sounds straightforward until they start asking about complexity tradeoffs and you realize they actually want a full analysis.

Given two arrays representing the inorder and postorder traversals of a binary tree with unique values, reconstruct the original tree and return its root. Walk through your recursive approach, define your base cases, and analyze the time and space complexity. Be prepared to discuss how array slicing affects complexity versus using an index-range approach with a value-to-index map.

“I started with the naive slice version because it was easier to explain, and they immediately asked me to quantify the cost.”

View Post

NVIDIA·Software Engineer·Technical Phone Screen

Jul 2026

NVIDIA systems interview, one meaty low-level question about implementing a memory allocator inside a fixed-size array with no external data structures allowed. The kind of problem that sounds manageable until you're actually sitting there trying to figure out where to stash your metadata.

You have an array of 32 elements, each holding 64 bits of data. Implement allocate(size) and free(handle) to manage memory blocks within this array. You cannot use any external data structures; all bookkeeping must live inside the array itself.

“My first instinct was to reach for a separate free list and I literally started describing one before remembering the constraint.”

View Post