Recent successful acquisition of NVIDIA New Grad/Intern Offer. The interview process for NVIDIA is highly dependent on the team, with significant differences between different groups. Some candidates go through a two-month OA plus telephone screening and virtual onsite (VO) process before being offered an internship at NVIDIA's SDE department. Others are directly recommended by their companies without going through the OA stage to enter the NVIDIA SDE VO process. Below, I break down the entire offer process from resume submission to the final offer, hoping to help aspiring candidates at NVIDIA avoid unnecessary detours.
Resume Submission Stage
- ChannelApplying through WeChat (Websites), internal referrals, and Career Fairs (campus recruitment fairs) are all options.
- SuggestionStrongly recommended to use an internal referral for higher success rates and faster processing compared to general advertising.
- Most overseas recruitment (OA) processes are typically required before internal referrals and career fairs. Many groups can skip the OA and proceed directly to interviews.
2. Online Assessment
- Not all groups have OA; some (especially internal referrals) skip it directly.
- Groups with OA use HackerRank; the quantity and difficulty vary by group.
- RemindBefore submitting for a position, it is important to clarify with the recruiter whether their group has an Open Access (OA) system to avoid unnecessary preparation.
Phone Screen / Technical Screening
Many groups also have this round for interns and recent graduates.
- DurationThere is a significant variance; some groups manage in 60 minutes while others take nearly 90.
- StructureIntroduction of NVIDIA and the position for the first 10 minutes, followed by 10 minutes of technical questions about projects and resumes. The remaining time was primarily for Technical Questions.
- Features:
- The questions were highly random; some people focused on the Infrastructure team while asking almost nothing about the OS and instead delved into detailed specifics of C++.
- Many interviewers would scrutinize resumes deeply, digging into the details of every project you have worked on.
- Preparation RecommendationsEnsure that every experience on your resume can clearly articulate the details and trade-offs involved. Be prepared with common issues in C++, focusing particularly on topics such as multithreading and memory management.
VO (Video Interview)
Some groups may skip the phone round and proceed directly to VO.
- ContentIntroduction and resume mining followed by technical questions.
- Some teams hardly write code but will make you hand-write concurrency-related content such as SPSC/MPSC (Single Producer Single Consumer/Multi Producer Multi Consumer).
- KeyThere are significant differences in style across different groups; it is best to consult with seniors in the group before interviews to understand their preferences.
Onsite
NVIDIA's Onsite is different from traditional large firms in that:
- At times, both phone screens and VO technical are already very hardcore.
- During the Onsite phase, some meetings have become somewhat lighter (with more behavioral focus and in-depth project exploration).
- The intensity and difficulty of each round depend heavily on the specific group. It is recommended to maintain high-level preparation throughout, and not to slack off simply because earlier rounds went smoothly.
Overall timeline: From receiving an invitation to a job offer, it typically takes around 1.5 to 2 months (variations may occur based on group differences).
NVIDIA SDE VO First Round Interview Experience Share
BQ
- Please provide experience you have in projects related to GPU or parallel computing.
- When optimizing a CUDA kernel, what aspects are typically addressed?
- Describe a time when you solved a performance bottleneck in an embedded environment with limited resources.
This round of the BQ interview focuses on candidates’ genuine understanding of GPU architecture and parallel programming. I mainly addressed these points, including Memory Coalescing, Occupancy, Bank Conflict, and Shared Memory, from a project I had worked on with CUDA. The interviewer seemed to pay close attention.
Coding part
First Question: Calculating Angle Between Clock Hands
给定时间字符串（如 “3:45″），计算时针和分针之间的最小夹角。
IdeasCalculate the absolute difference in angles between the hour and minute hands from 12 o'clock, then take the minimum value with 360 degrees, taking into account floating-point precision and boundary conditions.
Second Question: Sequential Printing with Multithreading
Design three threads to sequentially print numbers in order (A→1, B→2, C→3, A→4, ...).
IdeasImplement thread synchronization using a Semaphore. Each thread has its own semaphore, allowing only the first thread to proceed initially. Upon completing each print statement, the semaphore of the next thread is released, while a mutex protects the shared counter.
Follow-up：
- How would an algorithm be adjusted to support millisecond-level precision if clocks require it?
- How do you ensure that the system does not hang if a thread terminates abnormally?
The first round focuses on foundational aspects, but the follow-up will extend towards engineering robustness.
NVIDIA SDE VO Second Round Interview Experience Share
BQ
- Describe a time you handled a large-scale data or high-concurrency scenario in a project.
- How do you facilitate reaching consensus when there is a technical solution disagreement within your team?
- Experiencing an instance where I optimized overall system performance at a granular level.
System Design (Core Focus)
TopicDesigning a High-Throughput GPU Inference Service System
Discussion needs to be undertaken from the following dimensions:
- Overall Architecture Design (Frontend + Queue + Scheduler + Worker Pool)
- GPU Resource Scheduling and Dynamic Batch Processing (Dynamic Batching + Continuous Batching)
- Memory Management (Memory Pool + Tensor Allocator)
- Multi-card Multi-GPU Expansion (NCCL Communication)
- Balancing Throughput and Latency (pp. 99 Latency Control)
- Fault tolerance and monitoring mechanisms
The interviewers asked in great detail, covering aspects such as the hot/cold model loading strategy, degradation mechanisms when GPU memory is insufficient, and the use of CUDA Graph. I was well-prepared and spent nearly forty minutes discussing these topics.
Coding/Algorithm
TopicOptimization of Parallel Compute Task Scheduling Given a batch of GPU tasks (including computation volume, memory requirements, and priority), design a scheduling algorithm to minimize the overall completion time (Makespan).
IdeasFirst, I will present the Greedy + Priority Queue approach, and then discuss how to optimize it by combining Critical Path and List Scheduling when tasks have dependencies (DAG).
Follow-up：
- How to Monitor and Optimize Utilization of an Entire GPU Cluster in Production Environments?
- How should a system be extended to support mixed precision inference including FP8, FP16, and BF16 in the future?
Write at the end
That concludes my full NVIDA SDE VO interview experience share from the past two rounds. If you find yourself needing to enhance your GPU/CUDA project experience, system design skills, or the Voice of Experience (VO)环节, Programhelp's services might be worth considering. The team comprises alumni with backgrounds from Oxford, Princeton, Peking University, and senior professionals from companies like Amazon, Google, and Alibaba. Alums will directly communicate with you to assist with resume polishing, preparation for technical interviews, and overall interview readiness.Interview Coaching Assistance such as OA assistance.
Students in need can directly contact Programhelp for further discussion. The mentor will personally assess and provide advice.
Thanks for reading and I wish you all can get the NVIDIA Offer as soon as possible!
