AMD · OA → Phone screen → Onsite
In processAMD New Grad OA / Interviews by Role through Oct 2026: 11 Experiences (Part 1 :)
Note: This is the most complete collection of interview information I could find in this round of searching. If friends need more information in the future, I may start another round of research. Good luck with your interviews! — 11-02-2026 (jane)
In this series
- Part 1 (current): Software Engineer - GPU systems; C++ Software Engineer - Graphics; Software Development Engineer - ROCm/kernel; Applied Research Scientist; SDK Developer — ROCm Support; SDE2 - MTech campus hiring; Software Development Engineer - GPU functions; Graphics Profiler Software Engineer.
- Part 2: Firmware Engineer; Hardware role — exact title not reported; Full-stack Software Engineer - DevOps team.
Coverage: 8 of 11 AMD New Grad accounts in 8 separate role sections.
Other AMD hiring levels
- Experienced Full-Time: 16 accounts in 3 parts.
- Hiring Level Not Reported: 9 accounts.
- Intern / Co-op: 18 accounts in 2 parts.
Historical Xilinx reports
New Grad at Xilinx covers 1 account under its original employer, with separate counts.
Roles in this article
- Software Engineer - GPU systems — 1 account.
- C++ Software Engineer - Graphics — 1 account.
- Software Development Engineer - ROCm/kernel — 1 account.
- Applied Research Scientist — 1 account.
- SDK Developer — ROCm Support — 1 account.
- SDE2 - MTech campus hiring — 1 account.
- Software Development Engineer - GPU functions — 1 account.
- Graphics Profiler Software Engineer — 1 account.
Software Engineer - GPU systems
1 account in this role.
What came up
Reported topics: coding tasks; OS, memory, architecture or concurrency; GPU, graphics, ML or numerical topics.
Questions we found
Algorithms / Data Structures
| Question | Frequency | All last seen dates |
|---|---|---|
| Allocator with alloc/free and free-block coalescing | 1 | Reports: 2026-02-12 (report); Additional same-company practice history: 2026-09-06 |
| Producer-consumer ring buffer with wraparound | 1 | Reports: 2026-02-12 (report); Additional same-company practice history: 2026-02-10 |
| Parse a binary GPU firmware header with endianness handling | 1 | Reports: 2026-02-12 (report) |
System Design
| Question | Frequency | All last seen dates |
|---|---|---|
| GPU context-switching software stack | 1 | Reports: 2026-02-12 (report) |
Project Coding
| Question | Frequency | All last seen dates |
|---|---|---|
| Debug a GPU hang caused by missed fence signals | 1 | Reports: 2026-02-12 (report) |
Behavioral (BQ)
| Question | Frequency | All last seen dates |
|---|---|---|
| Discuss long projects, hard-to-test code and interest in low-level work | 1 | 2026-02-12 (report) |
01: New graduate: allocator, ring buffer and a GPU hang
Report date: February 12, 2026 · Software Engineer, new graduate · Result still pending
Senior-engineer phone screen — about one hour
This application went straight to a technical screen, without a preceding OA or recruiter call in the account. Memory questions covered virtual and physical addresses, page tables and the TLB.
The implementation task was an allocator with alloc, free and coalescing of adjacent free blocks. The candidate described using a doubly linked list and boundary tags. The linked exercise defines a finite arena, unit addresses and first-fit behavior; those details are practice assumptions.
Onsite C/C++ coding
They implemented a producer-consumer ring buffer with wraparound handling, then parsed a binary GPU firmware header with endianness handling. The account does not include the header’s actual layout. The linked exercise uses an authored 11-byte GPUF header, rather than an AMD firmware specification.
Practice: Growing Circular Queue covers FIFO wraparound using an authored growing-capacity contract; it does not grade thread synchronization.
Onsite design
The task was to design the software stack for GPU context switching between applications: saved and restored state, command buffers, priorities and preemption.
Practice: GPU Context-Switching Software Stack, with explicitly authored hardware and workload assumptions.
Onsite debugging
A driver reportedly caused a GPU hang after about ten minutes under a workload. The candidate traced it to a command-submission race that missed fence signals under high load. The buggy source code was not shared.
Practice: Repair a Command-Submission Fence Race. This is a disclosed CPU-only reconstruction with Python and Java workspaces, not the original driver or a ten-minute GPU reproduction.
Behavioral — about 30 minutes
Questions covered long-running projects, testing code that is hard to test, and the candidate's interest in low-level development.
C++ Software Engineer - Graphics
1 account in this role.
What came up
Reported topics: coding tasks; projects or resume.
02: A graduate graphics screen with guided reasoning
Report date: October 31, 2025 · Graduate C++ Software Engineer, graphics division
The completed first screen lasted about an hour. It covered the resume and small logic/code questions, with the interviewer guiding the candidate's reasoning. No exact exercise was shared.
A two-hour concurrency-focused second round was still upcoming. It is not a completed stage in this account.
Software Development Engineer - ROCm/kernel
1 account in this role.
What came up
Reported topics: coding tasks; OS, memory, architecture or concurrency; GPU, graphics, ML or numerical topics; projects or resume.
Questions we found
Behavioral (BQ)
| Question | Frequency | All last seen dates |
|---|---|---|
| English introduction, three strengths and why AMD should hire you | 1 | June, likely 2025 (inferred year; interview) |
03: Taiwan ROCm: three collaborative-editor tasks and cache layouts
Interview: June, likely 2025; year inferred from the job-search timeline · Software Development Engineer, ROCm/kernel · Taiwan · No further response · New graduate job search; still a student at this interview
The author was still completing their degree when this June interview took place; the same job-search account continues into their first post-graduation applications.
D+57 — HR phone call
The candidate gave an English introduction and discussed relevant experience, three strengths and why the company should hire them.
D+66 — technical interview
Scheduled for 90 minutes, this ran for about two hours. Three coding tasks were worked on in a collaborative editor. The author did not disclose the operations, but described being allowed to change types/signatures, repair deliberately wrong types, write tests and discuss complexity.
Questions also covered cache/register/RAM relationships, row-major versus column-major layouts and cache misses. Their projects and assembly experience received follow-ups.
D+70 — manager
This was scheduled for an hour but lasted a little over 20 minutes. It covered OS and computer-architecture coursework, architecture questions and a resume deep dive.
The June year is inferred from the surrounding job-search sequence; it is not an explicitly supplied interview year.
Applied Research Scientist
1 account in this role.
What came up
Research applicability appeared in several rounds. The most detailed recollection concerns attention arithmetic and a throughput discussion.
Questions we found
ML Modeling / Research Case
| Question | Frequency | All last seen dates |
|---|---|---|
| Explain how the candidate’s research could train, improve or apply to an O1-style reasoning model | 1 report | Reports: 2024-11-01 (report); Oct–Dec 2024 (header interview period) |
ML / LLM Technical Questions
| Question | Frequency | All last seen dates |
|---|---|---|
| Discuss attention changes that could reduce arithmetic, and define what is included in the count | 1 report | 2024-11-01 (report); Oct–Dec 2024 (header interview period) |
| Debug low training or serving throughput without assuming a trivial code bug | 1 report | 2024-11-01 (report); Oct–Dec 2024 (header interview period) |
04: A research new grad: six onsite rounds and an attention-compute disagreement
Interview period: Oct–Dec 2024 (header) · Report: Nov 1, 2024 · Applied Research Scientist · Reinforcement-learning PhD new grad · Rejected
This PhD new graduate worked on reinforcement learning and RLxF and had previous research internships. They applied to a new generative-AI team. A research-focused technical screen led to six onsite rounds.
The onsite sequence
- A research presentation.
- A coding-focused round, with discussion of other papers. The coding question itself was not disclosed.
- An open brainstorming or case-study round, again touching the candidate's other work.
- ML/LLM knowledge questions and paper discussion.
- A hiring-manager conversation in which the manager did most of the talking.
- Behavioral questions and ML breadth.
Several interviewers asked how the candidate's work could train or improve a reasoning model. They discussed Monte Carlo tree search, process reward models and possible uses of their own research. The candidate reported little follow-up discussion. The linked case makes this open research discussion into a practice exercise; its dataset, budgets and numerical decision rules were not interview givens.
Attention arithmetic
One interviewer asked about improvements to classical attention that could save compute or arithmetic operations. The candidate considered grouped-query attention and disagreed with the interviewer about the arithmetic being counted. In later author replies, they separated the cost of the K/V projection layers from the attention operation and said the wording had been ambiguous. This account records that disagreement rather than treating either participant's broad claim as a settled answer.
Throughput
The candidate was also asked about model-parallelism choices for low training or serving throughput. A later author clarification described the prompt as a customer-throughput debugging question that ruled out a trivial code bug. The candidate suggested reducing tensor parallelism when it was not needed because all-reduce and all-gather could be costly. They felt the interviewer was unconvinced.
The application ended in rejection. The candidate suspected a particular round had gone badly, but received no specific feedback confirming why. January and February 2025 author replies clarify this same application; they do not create another interview sighting.
SDK Developer — ROCm Support
1 account in this role.
What came up
The support role combined ROCm/CUDA experience, teamwork and a short C++ memcpy task.
Questions we found
Algorithms / Data Structures
| Question | Frequency | All last seen dates |
|---|---|---|
| Implement memcpy in C++ | 1 report | Reports: 2024-09-12 (report; author confirms a 2025 new-grad opening); Additional same-company practice history: 2025-01-06 |
Behavioral (BQ)
| Question | Frequency | All last seen dates |
|---|---|---|
| Introduce yourself | 1 report | 2024-09-12 (report; author confirms a 2025 new-grad opening) |
| What do you know about AMD, and why do you want to join? | 1 report | 2024-09-12 (report; author confirms a 2025 new-grad opening) |
| Describe a conflict within a team | 1 report | 2024-09-12 (report; author confirms a 2025 new-grad opening) |
05: ROCm SDK support: CUDA thesis questions, team conflict and memcpy
Report: Sep 12, 2024 · SDK Developer, ROCm support · Toronto, Canada · New-grad opening · Rejected
The role supported ROCm SDK work and GitHub issues. HR introduced the team, then asked for an introduction, what the candidate knew about AMD and why they wanted to join.
Because the thesis involved CUDA, the interviewer asked for an explanation and one or two follow-ups. A behavioral question covered conflict within a team.
The last task was a simple C++ memcpy implementation. The candidate was not comfortable enough with C++ to complete it and thought that might explain the rejection; that was their assessment, not confirmed interviewer feedback.
In a September 26 reply, the author confirmed that this was a 2025 new-grad opening. The structured header contains an older 2023 period, so the September 2024 report date is retained without claiming a precise interview date.
SDE2 - MTech campus hiring
1 account in this role.
What came up
Reported topics: coding tasks; OS, memory, architecture or concurrency; GPU, graphics, ML or numerical topics; projects or resume.
06: MTech SDE2: an OA, GPU-course projects and tree/graph coding
Post displayed as 2 years old on October 30, 2026; exact date unavailable · SDE2 campus applicant · IIT Madras context · Offer
OA — 60 minutes
A matrix task, a greedy task and MCQs. The candidate did not give the operations or individual MCQs.
First technical round — 60 minutes
The discussion covered the MTech thesis, GPU-programming course projects, networks, computer organization, OS, compiler design, Linux and basic DSA.
Second technical round — 60 minutes
The account names a tree question, tree DP and a graph question, without enough detail to identify them. The candidate described clarifying the tasks, arguing correctness and then coding.
An HR round followed, but its questions were not shared.
Software Development Engineer - GPU functions
1 account in this role.
What came up
Reported topics: coding tasks; OS, memory, architecture or concurrency; GPU, graphics, ML or numerical topics.
Questions we found
Algorithms / Data Structures
| Question | Frequency | All last seen dates |
|---|---|---|
| Matrix multiplication, multithreading, then row/grid/K splitting | 1 | Reports: 2021-04-07 (interview); 2021-04-06 (interview); Additional same-company practice history: 2025-10-05 (report date), 2024-03-31 |
| Reverse the bits of an unsigned 32-bit integer | 1 | Reports: 2021-04-06 (interview); Additional same-company practice history: 2020-05-20 (report date) |
| Remove duplicates from a sorted linked list | 1 | Reports: 2021-04-06 (interview) |
| Maximum depth of a BST | 1 | Reports: 2021-04-06 (interview) |
| Convolution (shape/kernel/padding not shared) | 1 | Reports: 2021-04-06 (interview) |
C/C++ / OS Questions
| Question | Frequency | All last seen dates |
|---|---|---|
| Races, locks, pipelines and bubbles | 1 | 2021-04-07 (interview) |
07: Taipei GPU/SDE: matrix multiplication became a parallelization discussion
Interviews: April 6–7, 2021 · Software Development Engineer, GPU functions · Taipei · Offer on April 9 · New graduate, master’s completed January 2021
The candidate had completed their master’s degree in January 2021 and was looking for their first full-time role after internships.
April 6 — onsite, two interviewers
The candidate was asked about MMU, caches and pages, and said they did not know those answers.
Coding covered reversing the bits of an unsigned 32-bit integer, removing duplicates from a sorted linked list, finding the maximum depth of a BST, multiplying matrices and a multithreaded follow-up. Convolution also appeared, without a specified shape, kernel or padding rule. The convolution practice contract explicitly chooses two dimensions, stride one, no padding and a flipped kernel.
April 7 — online
For M×K and K×N matrix multiplication with a large K, the discussion compared block-row splitting, grid splitting and splitting along K.
Other questions covered races, locks, pipelines and bubbles, FFT, loop unrolling, the TLB and channel conflicts.
Graphics Profiler Software Engineer
1 account in this role.
What came up
Reported topics: OS, memory, architecture or concurrency; GPU, graphics, ML or numerical topics.
08: Graphics profiler: C++, 3D math and a missed threading question
Report date: April 6, 2021 · Graphics profiler Software Engineer · Withdrew after another offer · New graduate hiring; master’s graduation expected June 2021
The author was completing a master’s degree and expected to graduate in June 2021; these interviews were part of that graduate job search.
The candidate described C++ and 3D-math questions, and said they could not answer a basic multithreading question. The exact prompts were not shared.
The manager was willing to move forward later, but the candidate had taken another offer. The wider job search was in the United States; this account does not establish the exact AMD office.
In this series
- Part 1 (current)
- Part 2
Comments
0No comments yet — be the first to share what you know.