As I understand, the A100 GPU performs significantly better with half-precision (FP16) computations compared to single-precision (FP32). I would like to know:
In your training process, did you use single precision, mixed precision, or half precision?
Have you experimented with different precision settings and their impact on training effectiveness? If so, what differences did you observe?
Thank you again for your work. I look forward to your response.
As I understand, the A100 GPU performs significantly better with half-precision (FP16) computations compared to single-precision (FP32). I would like to know:
In your training process, did you use single precision, mixed precision, or half precision?
Have you experimented with different precision settings and their impact on training effectiveness? If so, what differences did you observe?
Thank you again for your work. I look forward to your response.