Asked about What the hardware is doing
Your GPU shows 30 percent utilisation. What is likely wrong?
An answer that works
The GPU is waiting, almost always for data. Raise data loader workers, enable pinned memory, and check for per-item preprocessing that should be precomputed or moved onto the device.
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?What is mixed precision training?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?