FP8 and Low-Precision Formats Quiz

Answers and explanations appear as you take the quiz.

1. How many bits does an FP8 number use compared to a standard FP32 number?
   • 8 bits vs 16 bits (a half)
   • 8 bits vs 64 bits (an eighth)
   • 8 bits vs 32 bits (a quarter)
   • 8 bits vs 8 bits (the same)
2. What do the 'E4M3' and 'E5M2' labels describe about an FP8 format?
   • The error rate and model size
   • The number of exponent bits versus mantissa bits
   • The clock speed and memory channel
   • The training epochs and model layers
3. Why do FP8 pipelines apply scaling factors to tensors?
   • To encrypt the weights for security
   • To increase the number of model layers
   • To convert the model to a different language
   • To keep values inside FP8's narrow representable range and avoid overflow/underflow
4. Which trade-off is the main downside of using FP8?
   • Reduced numerical precision (rounding errors)
   • Higher memory usage
   • Slower matrix multiplication
   • Inability to run on any GPU
5. Which NVIDIA GPU generation first added dedicated hardware support for FP8 matrix math?
   • Pascal (P100)
   • Hopper (H100)
   • Kepler (K80)
   • Maxwell (M40)
6. What is the trend in precision formats moving beyond FP8?
   • A return to 64-bit double precision for inference
   • Eliminating floating point entirely in favor of text
   • Using only FP32 for all model weights
   • Even lower-bit micro-scaling formats like MXFP4 and NVFP4
7. A tensor’s values need more dynamic range. Which common FP8 encoding allocates more exponent bits?
   • E5M2
   • E4M3
   • Both allocate zero exponent bits
   • Neither encoding uses an exponent
8. Why can an FP8 result differ from the same operation using FP32 inputs?
   • Eight-bit input storage is automatically exact for every real value
   • FP8 removes all mathematical operations
   • The lower-precision input representation can introduce rounding error
   • Higher-precision accumulation restores every lost input detail