Went through the NVIDIA data scientist interview loop for a role on the cloud analytics team (supporting DGX Cloud adoption metrics). Going to break down the SQL, case, and stats components because that's what I searched for and couldn't find.
Structure of my loop: Recruiter screen (30 min) Technical phone screen with DS manager (45 min: SQL + one stat concept) Virtual onsite: 5 rounds total
Phone screen SQL: One medium SQL question. Window functions. Something like 'find the second-highest revenue customer per region per month.' Window functions and CTEs are table stakes; know them cold. The DS manager also asked me to explain the difference between a correlated subquery and a join in terms of performance. That's the kind of thing that filters people out fast.
Onsite breakdown:
SQL round (60 min): Two SQL problems. First was moderate (rolling averages, standard window function). Second was harder: a recursive CTE problem about hierarchical data. I'd practiced recursive CTEs but not enough. Got through it but not cleanly. Know your recursive CTEs.
Stats / ML round (60 min): Started with probability basics (classic conditional probability question), then moved into experiment design. 'We want to run an A/B test to see if a new onboarding flow increases DGX Cloud trial activation. What do you measure, how do you size the test, what are your power and significance thresholds.' I spent most of the round on this one. They wanted me to walk through sample size calculation out loud. Brush up on power analysis math, not just the concept.
Case / product sense round (60 min): This was the closest thing to a PM interview. 'User activation for DGX Cloud has dropped 15% over the past month. How do you diagnose it.' Standard product analytics framework but they wanted quantitative specifics at each step: what metrics, what queries, what the expected baseline is.
Behavioral + HM (45 min): Standard stories. Focus on data-driven impact. Have at least one story where your analysis changed a product decision and you can give the before/after numbers.
Overall DS difficulty at NVIDIA: harder than mid-market tech, comparable to a rigorous Google or Meta DS loop. They genuinely care about statistics depth, not just SQL fluency. The recursive CTE is probably not universal but know it anyway.