FP8/BF16 Mixed Precision Training & Numerical Overflow

Tech Deep DiveOpenAILast reported April 2026Low Frequency
Reported
1× across candidate reports
First seen
April 2026
Last reported
April 2026
Reported outcome
unknown

Problem Overview

Explain how mixed precision training works when combining FP8 and BF16 formats. Specifically, discuss: (1) why these formats are used together in training large neural networks, (2) what numerical overflow risks arise from FP8's very limited dynamic range…

  • 4 candidate-reported follow-ups — the exact probes interviewers asked, with the trigger for each
  • The rest of the problem statement — full requirements, constraints, and edge cases
  • Approach and trade-offs — what passing candidates did, and the mistakes that sink people
Unlock the full OpenAI catalog
Full problem statements, candidate-reported follow-ups, and walkthroughs — for every OpenAI question.
Unlock with Pro
Already a member? Sign in
Verified Source
Every question is reconstructed from multiple independent candidate reports. Verbatim follow-ups, not invented ones.
Codex Fact-Checked
Technical claims, formulas, and scale numbers are reviewed against primary sources.
Interviewer Follow-ups
The exact follow-ups reported by candidates, with the trigger that prompts each one — plus the mistakes that sink people.
Is this helpful?