Autocoder

NVIDIA Interview Questions and Answers for India Software Roles

Last updated 2026-11-19

NVIDIA software interviews tend to test whether you understand why a GPU is fast and where it is slow. You do not always need CUDA experience, but you are expected to reason about parallelism, memory access patterns and the cost of synchronisation, on top of solid C++ and data structures. This page gives twelve practice questions with worked answers.

They are written for the topics the role is commonly reported to cover, not taken from any interview. Read how to crack the NVIDIA interview for the process and NVIDIA coding round questions for problems to code.

Typical rolesNew college graduate software engineer, systems software, driver and compiler, deep learning software
Core topicsC++, GPU architecture, parallel algorithms, memory hierarchy, DSA
Stages (NVIDIA careers page)Phone interviews, then virtual or in-person interviews, then an onsite

Rounds as NVIDIA describes them

NVIDIA's careers page says full-time candidates have phone interviews followed by virtual or in-person interviews, that an in-person onsite interview is required before an offer, and that interns typically have phone interviews only. Interviews are 30 to 60 minutes each, and technical candidates may get a coding exercise on HackerRank, on a whiteboard or on a laptop NVIDIA provides.

Beyond that, the content is up to the team. Prepare for coding with explanation aloud, plus questions drawn from the skills listed in the job description.

Topics to revise

Hiring status

There is no NVIDIA hiring page on this site, because NVIDIA has not published an India fresher hiring number or calendar that we can quote. One report says NVIDIA signed a long lease for a large Bengaluru office as it expands its India base, but the report gives no hiring figure.

Check the company's own careers site at nvidia.com for current openings, and use the process facts above, which come from NVIDIA's careers page.

NVIDIA Interview Questions and Answers for India Software Roles: questions with worked answers

  1. 1

    What is a warp, and what is warp divergence?

    easy

    On NVIDIA GPUs a warp is a group of 32 threads that execute the same instruction together (the SIMT model). If threads in a warp take different sides of a branch, the hardware runs each path in turn with the other threads masked off, so the warp takes the time of both paths. Divergence matters only within a warp, not across warps. To reduce it, make branch conditions depend on warp-aligned values, sort or compact the data, or use arithmetic instead of branches.

  2. 2

    Describe the GPU memory hierarchy and the relative cost of each level.

    easy

    Registers are per thread and fastest. Shared memory is on-chip, per block and programmer managed, with latency close to registers when conflict-free. L1 and L2 caches sit between the cores and global memory. Global memory is large, off-chip and has the highest latency, hundreds of cycles. Constant memory is cached and good when all threads read the same address. Local memory is thread-private but physically in global memory, used for register spills and large arrays. Good kernels reuse data from registers and shared memory to avoid global traffic.

  3. 3

    What is memory coalescing and why does it matter?

    medium

    When the threads of a warp access consecutive addresses, the hardware combines the requests into a few wide memory transactions. If they access scattered addresses, or addresses with a large stride, one warp can generate many transactions and the effective bandwidth collapses. Prefer a structure-of-arrays layout so that thread i reads element i, and make the fastest-varying thread index map to the contiguous dimension.

  4. 4

    What is a shared-memory bank conflict and how does padding fix it?

    medium

    Shared memory is split into 32 banks, and consecutive 4-byte words fall in consecutive banks. If several threads of a warp access different words in the same bank, the accesses are serialised. In a 32 by 32 float tile, a warp reading a column hits one bank 32 times. Declaring the tile as [32][33] shifts each row by one bank so column accesses are spread across banks. All threads reading the same word is a broadcast and causes no conflict.

  5. 5

    Write a CUDA kernel to add two vectors and show how to launch it.

    easy

    Each thread computes one global index from its block and thread indices and guards against running past the end. Launch enough blocks to cover n with a ceiling division. The guard matters because n is rarely a multiple of the block size.

    cpp
    __global__ void add(const float *a, const float *b, float *c, int n) {
        int i = blockIdx.x * blockDim.x + threadIdx.x;
        if (i < n) c[i] = a[i] + b[i];
    }
    
    // host side
    int threads = 256;
    int blocks = (n + threads - 1) / threads;
    add<<<blocks, threads>>>(d_a, d_b, d_c, n);
  6. 6

    Explain parallel reduction. What is its work and step complexity?

    medium

    A reduction combines n values with an associative operator such as sum or max. In a tree reduction, each step adds pairs, halving the active elements: n/2, n/4, and so on. Total work is O(n) additions and the number of steps is O(log n). On a GPU each block reduces its tile in shared memory, with a synchronisation between steps, then partial results are combined by a second kernel or by atomics. Using sequential addressing (thread i adds element i + stride) avoids divergence and bank conflicts that interleaved addressing causes.

    c
    #include <stdio.h>
    int main(void) {
        int a[8] = {3, 1, 4, 1, 5, 9, 2, 6};
        for (int stride = 4; stride > 0; stride /= 2)
            for (int i = 0; i < stride; i++) a[i] += a[i + stride];
        printf("%d\n", a[0]); /* 31 */
        return 0;
    }
  7. 7

    A program is 90 percent parallelisable. What is the speedup on 10 processors, and the limit with infinite processors?

    easy

    Amdahl's law: speedup = 1 / ((1 - p) + p / n). With p = 0.9 and n = 10, speedup = 1 / (0.1 + 0.09) = 5.26. With unlimited processors it tends to 1 / 0.1 = 10. The serial 10 percent caps the gain, which is why GPU kernels try to minimise host-side serial work and transfers over PCIe.

  8. 8

    What limits occupancy, and is higher occupancy always better?

    medium

    Occupancy is the ratio of active warps on a streaming multiprocessor to the maximum it supports. It is limited by registers per thread, shared memory per block, block size and the cap on resident blocks. More active warps help hide memory latency, but not always: a kernel that uses many registers to keep data in fast storage can run faster at lower occupancy. Measure with a profiler instead of assuming.

  9. 9

    How would you build a histogram on a GPU, and what is the hazard?

    medium

    If every thread increments a global bin directly, two threads can update the same bin at once and updates are lost, so use atomicAdd. Heavy contention on popular bins is slow, so each block builds a private histogram in shared memory with atomics and then merges it into the global one with one atomic per bin. The trade-off is extra shared memory for fewer slow global atomics.

  10. 10

    What is the difference between unique_ptr and shared_ptr in C++?

    medium

    unique_ptr owns an object exclusively, has no overhead beyond the raw pointer, and can only be moved. shared_ptr uses a reference count held in a control block, copies share ownership, and the object is deleted when the last owner goes away. Prefer unique_ptr by default, use shared_ptr only for genuinely shared ownership, and beware of cycles, which need weak_ptr to break. Both implement RAII so cleanup happens even if an exception is thrown.

  11. 11

    How does tiling speed up matrix multiplication?

    medium

    A naive kernel reads each element of A and B from global memory N times. In the tiled version, a block loads a TILE by TILE sub-block of each matrix into shared memory once, synchronises, and every thread uses those values TILE times from fast memory. Global reads drop by a factor of TILE, and the access pattern becomes coalesced. The cost is shared memory, which limits TILE, and barriers between phases.

  12. 12

    HR: Why NVIDIA, and tell us about a project where you made something faster.

    easy

    Answer with a measured example: for instance, you profiled a program, found a hot loop, changed the data layout or added threading, and the run time fell from one figure to another. Say what tool you used and what you would try next. Connect it to NVIDIA by naming the kind of problem you want to work on, such as inference performance or drivers, rather than repeating that the company is well known. See how to crack the technical interview.

Frequently asked questions

Do I need CUDA experience to interview at NVIDIA?

It depends on the team. For some roles the job description asks for CUDA or GPU experience, and for others strong C++ and algorithms are the focus. Read the posting and prepare the skills it lists.

What does NVIDIA say about its interview rounds?

Its careers page says candidates have phone interviews, then virtual or in-person interviews of 30 to 60 minutes, and that an onsite in-person interview is required before an offer. Interns typically have phone interviews only.

Are these NVIDIA interview questions real ones?

No. They are practice questions in the style of topics the role is commonly reported to cover, and nothing here records a real interview.

Where can I see current NVIDIA openings in India?

Use NVIDIA's own careers site. See also how to crack the NVIDIA interview for the process.

Sources

Related guides

More in Company Interview Questions

See all in this section