Hi, your observation about batch size and training speed could be influenced by several factors related to MATLAB’s deep learning framework and GPU memory management. While increasing batch size generally improves GPU utilization, it also increases memory demands. The 4060Ti, despite having more VRAM, may have different memory bandwidth and optimization behaviour affecting efficiency. MATLAB's built-in deep learning functions may also handle memory allocation differently depending on the GPU architecture.
Additionally, factors like mixed-precision training, data pipeline efficiency, and MATLAB's automatic batch size adjustments could be at play. You can try monitoring GPU usage using gpuDevice and adjusting the ExecutionEnvironment and DispatchInBackground options. Testing with datastore optimizations might also provide insights.
You can refer to these MATLAB documentations for better understanding GPU Computing in MATLAB
Run MATLAB Functions on a GPU
and Datastore
