AMD Interview Questions and Answers for India Software Roles
Last updated 2026-11-19
AMD designs CPUs, GPUs and adaptive processors, so software interviews tend to reach down into the hardware: how coherence keeps caches consistent, why a loop with a power-of-two stride is slow, what out-of-order execution does, and how well you know modern C++. This page offers twelve practice questions with worked answers for that kind of role.
They are written for the topics such roles are commonly reported to cover, not copied from any interview. For problems to code, see AMD coding round questions.
| Typical roles | Software engineer, firmware, driver and compiler, performance and validation |
|---|---|
| Core topics | Computer architecture, C and C++, OS, parallel programming, DSA |
| Rounds (commonly reported) | Online or written screening, technical interviews, HR |
Rounds as commonly reported
Prep write-ups describe a screening stage, then two or three technical interviews and an HR or manager conversation. AMD does not publish a fixed format, and it varies between a driver team, a compiler team and a validation team, so rely on your own invitation for the real structure.
Interviewers for hardware-adjacent software often ask you to reason from first principles: given a cache size and a stride, will this loop hit or miss? Practise explaining your reasoning in numbers.
Topics to revise
- Caches: associativity, conflict misses, coherence protocols, false sharing.
- Virtual memory: page tables, TLB, huge pages, NUMA.
- Core microarchitecture: pipelining, out-of-order execution, renaming, branch prediction.
- Parallelism: SIMD, threads, memory ordering, the roofline view of performance.
- C++: RAII, move semantics, templates, virtual dispatch. See C++ OOPS interview questions.
Hiring status
This site holds no sourced hiring statements for AMD, so it quotes no drive dates, openings or packages. Treat any post that gives them without a source link with caution.
Check the company's own careers site for current openings and application instructions.
AMD Interview Questions and Answers for India Software Roles: questions with worked answers
1 Describe the MESI cache coherence protocol.
mediumEach cache line in a private cache is Modified (dirty, only copy), Exclusive (clean, only copy), Shared (clean, may be in other caches) or Invalid. A read miss loads the line as Exclusive if no other cache has it, otherwise Shared. A write to a Shared line broadcasts an invalidation so other copies become Invalid and the writer moves to Modified. A write to an Exclusive line moves to Modified silently. Snooping or a directory implements this. The cost of a write to a widely shared line is why false sharing hurts.
2 A 32 KB, 8-way set-associative cache has 64-byte lines. Why is a loop with a 4096-byte stride slow?
mediumThe number of sets is 32768 / (64 x 8) = 64. The set index is address bits 6 to 11, so addresses that differ by a multiple of 4096 bytes (64 sets x 64 bytes) map to the same set. A loop touching more than 8 such addresses keeps evicting its own data, so almost every access is a conflict miss even though the total data is small. Fixes: pad the row length so the stride is not a power of two, tile the loop, or change the layout.
3 What does a TLB do, and what happens on a TLB miss?
mediumThe translation lookaside buffer caches recent virtual-to-physical page translations so that most memory accesses do not need to walk the page table. On a miss, the hardware (or the OS on some architectures) walks the multi-level page table, typically four memory reads on x86-64, then fills the TLB. If the page is not present, a page fault lets the OS fetch it. Huge pages (2 MB or 1 GB) cover more memory per entry and cut TLB misses for large working sets.
4 What is out-of-order execution and how does register renaming help?
hardAn out-of-order core fetches instructions in order but executes them as their operands become ready, then retires them in program order using a reorder buffer, which keeps exceptions and visible state precise. Register renaming maps the few architectural registers onto many physical registers, which removes write-after-read and write-after-write false dependences, leaving only true data dependences. Together with branch prediction this lets the core overlap cache-miss latency with independent work. A long dependent chain, such as pointer chasing, still limits it.
5 Array of structures or structure of arrays: which is better for SIMD and why?
mediumIf a loop reads only the x field of many particles, an array of structs interleaves x with unused fields, wasting cache bandwidth and preventing contiguous vector loads. A structure of arrays keeps all the x values together, so a vector load fills a register with useful data and the loop can use SIMD at full width. AoS is better when most code touches whole objects. Hybrid layouts (small blocks of SoA) are common in high-performance code.
6 What is NUMA, and how can it affect a multi-threaded program?
mediumIn a Non-Uniform Memory Access system, each processor socket or chiplet group has local memory, and access to another node's memory is slower. A thread that allocates memory on one node and then runs on another pays the penalty on every miss. Linux by default places a page on the node of the thread that first touches it, so initialise data in parallel with the threads that will use it, and pin threads or use affinity to keep them near their data.
7 Explain move semantics in C++ and show a move constructor.
mediumA move constructor takes an rvalue reference and steals the resources of a temporary instead of copying them, leaving the source in a valid but empty state. That avoids deep copies when returning or inserting large objects. Mark it noexcept, because standard containers use the move only if it cannot throw. Follow the rule of five: if you define a destructor, copy operations or a move operation, think about all five.
cpp#include <cstddef> #include <utility> class Buf { int *p_; std::size_t n_; public: explicit Buf(std::size_t n) : p_(new int[n]()), n_(n) {} Buf(Buf &&o) noexcept : p_(o.p_), n_(o.n_) { o.p_ = nullptr; o.n_ = 0; } Buf &operator=(Buf &&o) noexcept { if (this != &o) { delete[] p_; p_ = o.p_; n_ = o.n_; o.p_ = nullptr; o.n_ = 0; } return *this; } Buf(const Buf &) = delete; Buf &operator=(const Buf &) = delete; ~Buf() { delete[] p_; } std::size_t size() const { return n_; } }; int main() { Buf a(100); Buf b(std::move(a)); // a is now empty return (a.size() == 0 && b.size() == 100) ? 0 : 1; }8 How does release and acquire ordering let one thread publish data to another?
hardThe writer stores the data with ordinary writes and then sets a flag with memory_order_release. The reader loads the flag with memory_order_acquire. If the reader sees the flag set, it is guaranteed to see all writes that happened before the release store. Relaxed ordering gives no such guarantee, and on weakly ordered CPUs the reader could see the flag but stale data. Sequentially consistent ordering is the default and the simplest, but can cost more fences.
cpp#include <atomic> #include <thread> #include <cassert> int data = 0; std::atomic<bool> ready{false}; int main() { std::thread w([]{ data = 42; ready.store(true, std::memory_order_release); }); std::thread r([]{ while (!ready.load(std::memory_order_acquire)) {} assert(data == 42); }); w.join(); r.join(); }9 What does restrict mean in C, and why can it speed up a loop?
mediumA pointer qualified with restrict is a promise from the programmer that, during its lifetime, the memory it points to is accessed only through that pointer or values derived from it. The compiler can then assume that stores through one pointer do not change data read through another, so it can keep values in registers and vectorise. If the promise is false, for example the arrays overlap, the behaviour is undefined. Compare with memcpy, which has restrict parameters, and memmove, which does not.
10 What is the roofline view of performance? Classify a float vector addition.
mediumAttainable performance is the lower of the peak compute rate and memory bandwidth times arithmetic intensity (flops per byte moved). For c[i] = a[i] + b[i] with floats, each element needs one add and moves 12 bytes (two 4-byte reads and one 4-byte write), so the intensity is about 0.083 flops per byte. That is far below any modern machine balance, so the loop is memory-bound and adding compute units will not help; improving locality or fusing loops will.
11 Differentiate a spinlock from a mutex and say when each is right.
mediumA spinlock loops on an atomic flag until it can take the lock, wasting cycles but avoiding a context switch, so it suits very short critical sections held by threads running on different cores and code that cannot sleep. A mutex puts the waiting thread to sleep, saving CPU time at the cost of a system call and a wake-up delay, so it suits longer sections. Mention test-and-test-and-set and backoff to reduce cache-line bouncing.
12 HR: Why AMD, and what area of the stack interests you?
easyBe specific. For example: you enjoy performance work, you profiled a program and saw how the cache layout changed its speed, and you want to work on software that sits next to a CPU or GPU, such as drivers, compilers or libraries. Name the project, the measurement and the lesson. Avoid comparing the company with its competitors in your answer.
13 HR: How do you learn a new codebase quickly?
mediumGive a method: read the build and test setup first, run a test under a debugger, trace one request or one function end to end, read recent commits for the area, and ask a targeted question after you have tried to answer it yourself. Give a real example, such as an open-source project you contributed to. See how to crack the technical interview.
Frequently asked questions
Are AMD software interviews hardware-heavy?
Commonly reported for driver, firmware, compiler and performance roles: caches, virtual memory and parallelism come up. Application software roles may focus more on C++ and data structures.
How should I prepare for C++ questions?
Revise RAII, smart pointers, move semantics, templates, virtual functions and the memory model. Write short programs to check your understanding, not just read about them.
Are these real AMD questions?
No. They are practice questions in the style of topics such roles commonly cover. AMD does not publish its interview questions.
How do I find current AMD openings?
Use the company's own careers site. This site has no sourced hiring facts for AMD, so it quotes none.
Related guides
- AMD Coding Round Questions with Solutions
- Intel Interview Questions and Answers for India Software Roles
- NVIDIA Interview Questions and Answers for India Software Roles
- Qualcomm Interview Questions and Answers for India Software Roles
- Computer Architecture Interview Questions with Numericals
- C++ OOPS Interview Questions
More in Company Interview Questions
- American Express coding round questions with solutions
- American Express interview questions with answers (engineering)
- Analog Devices Coding Round Questions with Solutions
- Analog Devices Interview Questions and Answers for India Embedded and DSP Roles
- Apple Coding Round Questions with Solutions in Python and Java
- Apple Interview Questions and Answers for SDE Roles
- Arcesium Coding Round Questions With Solutions
- Arcesium Interview Questions and Answers
- Atlassian Coding Round Questions with Solutions in Python and Java
- Atlassian Interview Questions and Answers for SDE Roles
- Barclays coding round questions with solutions
- Barclays interview questions with answers (technology graduates)