A. FP8.
B. INT32.
C. INT8.
D. INTFB8.
2. FP16 is a numeric representation that uses 16 bits to represent floating-point numbers. What is this numeric format also known as?
A. Half-precision floating-point representation.
B. Single-precision floating-point representation.
C. Double-precision floating-point representation.
D. OneQuarter-precision floating-point representation.
3. Which of the following is a numeric representation for floating-point numbers created by Google to attend to machine learning and artificial intelligence workloads?
A. GP16.
B. GFP16.
C. FP16.
D. BFP16.
4. NVIDIA created the TF32 data representation. Which of the following number of bits does it use to represent floating-point numbers?
A. 32 bits.
B. 19 bits.
C. 16 bits.
D. 20 bits.
5. What is the default numeric representation that’s used by PyTorch to execute the operations for the training process?
A. FP32.
B. FP8.
C. FP64.
D. INT32.
A. Mixed precision tries to adopt lower-precision formats during the training process’ execution.
B. Mixed precision tries to adopt higher-precision formats during the training process’ execution.
C. Mixed precision avoids the usage of lower-precision formats during the training process’ execution.
D. Mixed precision avoids the usage of higher-precision formats during the training process’ execution.
A. Simple usage and reduction of performance improvement.
B. Simple usage and reduction of power consumption.
C. Complex usage and avoidance of errors involving numeric representation.
D. Simple usage and avoidance of errors involving numeric representation.
8. Besides the lack of lower-precision operations, which of the following options is another reason to not use a pure lower-precision approach in the training process?
A. Low performance improvement.
B. High energy consumption.
C. Loss of information on the gradient.
D. High usage of main memory.