AMD · Onsite
In processAMD Experienced Full-Time Interviews by Role through Oct 2026: 16 Experiences (Part 3 :)
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: Senior Software System Designer; GPU Software Engineer; Senior Software Engineer - uProf; Compiler / Performance; Mid-level Software Engineer.
- Part 2: SDE2; Senior Software Engineer - AI; MTS Software Development Engineer - GPU; Role not reported — post-offer start delay; Software System Designer 2 — Test Automation; Verification — PCIe; Role not reported — C and memory panel.
- Part 3 (current): GPU team — exact role not reported; MTS Verification; Member of Technical Staff - Performance.
Coverage: 3 of 16 AMD Experienced Full-Time accounts in 3 separate role sections.
Other AMD hiring levels
- Hiring Level Not Reported: 9 accounts.
- Intern / Co-op: 18 accounts in 2 parts.
- New Grad: 11 accounts in 2 parts.
Roles in this article
- GPU team — exact role not reported — 1 account.
- MTS Verification — 1 account.
- Member of Technical Staff - Performance — 1 account.
GPU team — exact role not reported
1 account in this role.
What came up
FIFO behavior was examined twice. C++ ownership, cache policy and a disputed matrix working-set explanation filled the other technical sessions.
Questions we found
Algorithms / Systems Modeling
| Question | Frequency | All last seen dates |
|---|---|---|
| Compute IPC through two backpressured FIFO stages | 1 report | Reports: 2019-10-28 (report); Oct–Dec 2019 (header interview period) |
| Implement a shared-ownership pointer class | 1 report | Reports: 2019-10-28 (report); Oct–Dec 2019 (header interview period) |
| Implement an LRU cache, with a hardware follow-up | 1 report | Reports: 2019-10-28 (report); Oct–Dec 2019 (header interview period); Additional same-company practice history: 2026-10-03 (report date) |
| Model a FIFO | 1 report | Reports: 2019-10-28 (report); Oct–Dec 2019 (header interview period) |
Cache / Architecture Reasoning
| Question | Frequency | All last seen dates |
|---|---|---|
| Discuss the cache working set for two N×N matrices with four-byte elements; the interviewer later excluded output storage | 1 report | 2019-10-28 (report); Oct–Dec 2019 (header interview period) |
14: A GPU-team onsite: FIFO stalls, shared ownership and a hardware LRU discussion
Interview period: Oct–Dec 2019 (header) · Report: Oct 28, 2019 · GPU team; exact title not reported · Explicitly not a new-grad role
The candidate said this GPU-team role required experience and was not a new-grad opening. The onsite had five sessions, one of them lunch.
Session 1 — FIFO latency and throughput
The interviewer asked for a C++ model involving modules A and B. A had a FIFO of capacity 10 and B had one slot. A could receive one instruction per cycle and B could consume one. Full buffers and a stalled consumer could prevent progress. The requested API had to compute the IPC of both modules. The FIFO-pipeline exercise covers that backpressure core.
Session 2 — C++ details and ownership
C++ questions led to implementing a shared-pointer class. The report did not retain a fuller interface or ownership contract.
Session 3 — lunch
This was an informal conversation.
Session 4 — caches, LRU and GPU concepts
Questions covered cache indexing, an LRU cache in C++, a Verilog implementation, GPU concepts and warps. After explaining a doubly linked list for the C++ version, the candidate proposed counters for the hardware version. In later author replies, they acknowledged knowing the pLRU tree approach but said they had not thought to use it in the interview. The interviewer accepted the counter explanation and moved on; the report does not establish which implementation was expected.
A separate matrix-multiplication discussion became a disagreement about what the cache had to hold. For two N×N matrices with four-byte elements, the interviewer eventually gave N × 2 × 4 bytes and said output storage could be ignored. The candidate had been considering input lines and intermediate results and felt the working-set assumption had not been made clear. This is the interviewer's reported answer under that discussion, not a general matrix-storage requirement.
Session 5 — another FIFO task
The last session asked the candidate to model a FIFO. The candidate found the questions open-ended and the feedback hard to read: interviewers answered clarifying questions but gave little indication of whether an answer was right.
MTS Verification
1 account in this role.
What came up
Four interviews moved from UVM and programming to verification plans and conversations with designers.
Questions we found
Verification / UVM Coding
| Question | Frequency | All last seen dates |
|---|---|---|
| Write a UVM package class with constraints; the class purpose was not supplied | 1 report | 2019-07-04 (report); Oct–Dec 2017 (header interview period) |
Algorithms / Data Structures
| Question | Frequency | All last seen dates |
|---|---|---|
| A sorting exercise; exact inputs were not retained | 1 report | Reports: 2019-07-04 (report); Oct–Dec 2017 (header interview period) |
Verification Design Questions
| Question | Frequency | All last seen dates |
|---|---|---|
| Explain a testplan | 1 report | 2019-07-04 (report); Oct–Dec 2017 (header interview period) |
Behavioral (BQ)
| Question | Frequency | All last seen dates |
| How do you communicate with a designer? | 1 report | 2019-07-04 (report); Oct–Dec 2017 (header interview period) |
15: MTS verification: a constrained class, programming and two testplan discussions
Interview period: Oct–Dec 2017 (header) · Report: Jul 4, 2019 · MTS Verification · Experienced-hire context · Offer / pass
The author described four interviews and lunch with the manager. The header places the interview in late 2017, while the account was published in July 2019.
The first senior technical interviewer asked for a UVM package class involving constraints and protocol-related work. The report does not identify the class's purpose or the constraints. It cannot establish a particular packet format or a complete UVM task.
The second interviewer asked programming questions involving sorting and string processing. The sorting practice link supplies a bounded exercise; the original input contract was not recorded. No specific string operation was disclosed.
The manager asked about testplan design and the candidate's previous projects. A design manager then discussed testplan design and communication with designers. In an author reply, a testplan was described as a document laying out the design's features and how they would be verified.
The candidate described a relaxed atmosphere and a positive outcome. A later question about export licensing received an answer about that candidate's circumstances; the report does not establish a current policy for other applicants.
Member of Technical Staff - Performance
1 account in this role.
What came up
Reported topics: coding tasks; projects or prior work; OS, memory, architecture or concurrency.
16: Performance-team MTS: seven coding tasks and six puzzles
Account available before July 22, 2024; exact interview year unknown · Member of Technical Staff, Performance team · About 1 year 11 months of prior Qualcomm experience
All three rounds were online.
- Round 1: two coding questions, described as array and matrix questions; three puzzles; Linux basics.
- Round 2: five coding questions across arrays, stacks, graphs, submatrices and DP; one puzzle.
- Round 3: two more puzzles and discussion of the candidate's Qualcomm work.
Neither the seven coding operations nor the six puzzle premises were preserved. Other employers' code samples elsewhere in the same repository do not identify these AMD tasks.
Comments
0No comments yet — be the first to share what you know.