You want to reduce GPU memory use during training with little accuracy loss. Which technique helps?
Why B is the answer
Mixed precision training
Mixed precision uses lower precision where it is safe, which saves memory and often speeds up training.
In detail
- Mixed precision uses smaller number formats where it is safe.
- It often speeds up training too.
Use case
An illustrative scenario, not a real incident.
A training run hits GPU memory limits. Mixed precision uses lower precision where it is safe, and the run fits.
See it in different ways
stateDiagram-v2 [*] --> Watch Watch --> Healthy: both losses fall Watch --> Overfitting: validation loss rises Watch --> Unstable: loss swings wildly Overfitting --> Regularize Unstable --> LowerLearningRate Regularize --> Watch LowerLearningRate --> Watch
flowchart LR S0["Identify safe operations"] --> S1["Use lower precision"] --> S2["Train"]
Key point: Mixed precision saves memory.
Why the other options are wrong
- A. Using higher precision floats
- Higher precision uses more memory.
- C. Adding more layers
- More layers use more memory.
- D. Doubling the batch size
- A larger batch uses more memory.