Asked about What the hardware is doing
What is mixed precision training?
An answer that works
Using 16-bit for most operations and 32-bit where precision matters, with loss scaling to stop small gradients underflowing. It roughly halves memory and can double throughput at essentially no quality cost.
Reading an answer is not being able to give one. The section that teaches this is What the hardware is doing, 5 hours of GPUs & accelerators.
Learn this properlyAlso asked about GPUs & accelerators
Why are GPUs used for deep learning?Your GPU shows 30 percent utilisation. What is likely wrong?Is LLM inference compute-bound or memory-bound?What does kernel fusion do?How do you tell whether a GPU workload is limited by compute or memory bandwidth?You need to serve a 7B model cheaply at moderate traffic. How do you choose hardware?