A recruiter first found me through a professional network and referred me to IMC.
After two phone screenings with a hardware engineer and a machine-learning algorithm engineer, I was invited to the Chicago office for an onsite.
The onsite had five rounds, all on the same day, and I met ten interviewers across those rounds.
First round: Hardware architecture. I was first given a small program written in their internally developed language and asked to run it in their own simulated compilation and execution environment, then list the instruction trace. The engineer then asked various questions based on the hardware specifications, latency, and throughput, such as how to improve it from the algorithm or cache side. The questions were not difficult. They progressed gradually, with follow-ups going deeper step by step. I was also asked about the roofline model.
Second round: Behavioral interview. The interviewers were a tech lead and an HR lead. When I later reviewed the interview with AI, it thought that my main red flag appeared in this round. The format was fairly casual, more like a conversation that went wherever it went. They asked several questions, including:
What are the three words people around you use most often to describe you?
If you were mentoring a new employee, what would you tell them to pay attention to? I answered directly that the core members of my previous team were all PhDs and there were no new graduates, so I felt there had not really been a situation where I mentored a new employee. Also, because my current company's management style is relatively top-down, I asked the interviewers how they would handle several similar scenarios, such as whether a manager would intervene in technical details. The AI review thought I should not have spoken negatively about my current employer.
At lunch, a manager and a young entry-level engineer ate with me. I mainly learned about their usual work culture, environment, and return-to-office policy. The current policy is roughly four days in the office each week. IMC does not have internal levels like the Senior, Staff, and Principal divisions at traditional tech companies. The goal is to let everyone discuss things equally and judge ideas purely by their technical merits.
Third round: Another architecture interview. It felt like the main task was for me to explain projects I had worked on before, with the interviewers stopping me to ask questions whenever they had doubts. One possible mismatch was that they pursue zero latency and therefore use very small models, while my previous work mainly involved optimizing and deploying LLMs. They still seemed very interested and asked a great many questions.
Fourth round: Machine learning and algorithms. Two young engineers working more on the ML algorithm side conducted this interview. I was given a piece of training code written in PyTorch and asked to use a profiling tool to dump traces, analyze which parts of the training process were inefficient, and find the bottleneck. To be honest, my previous work had mainly been inference acceleration, so this may have been my worst round. I was not very familiar with training or with PyTorch profiling tools. With their step-by-step guidance, though, I eventually found the problem.
Fifth round: System design and design problem. Two engineers also conducted this round. It did not involve specific low-level details. I was mainly given two machines with different specifications, throughput, and latency, along with several throughput and latency curves, and asked when each machine would be faster and how to allocate the workload. By the afternoon my mind was getting sluggish, and I answered one or two questions rather slowly, but overall I did not think there were any major problems.
Two business days later, I received a template rejection email. I tried to ask for the reason, but they did not reply.
I asked AI to review the process. It thought that, besides my criticism of my current employer during the behavioral round, I had been somewhat demanding while arranging a hotel with the recruiting coordinator. The hotel I chose was initially rejected, but after I explained the reason, the coordinator changed it back. Only after those issues did the AI rank my missing knowledge about ML training.
Another thing I noticed was that the younger interviewers emphasized their elite-school backgrounds, such as CMU and Columbia, but none had a PhD. I felt the atmosphere was still very different from tech.
None of the questions went into very deep technical detail. There was no SystemVerilog, no LeetCode, and not even detailed microarchitecture. Everything was fairly abstract. They may care more about problem-solving methodology and habits of thought.
Discussion
Loading comments…