Why does using the Metal plugin on my M2 Max with TensorFlow reduce processing times but lower accuracy, and how can I fix it?

Question

I have an Apple computer with an M2 Max chip. I've downloaded TensorFlow and the Metal plugin to utilize the GPU on Jupyter Notebook. The message 'Plugin optimizer for device_type GPU is enabled' appears. Indeed, the processing times are significantly lower (down to 20 seconds from 1 minute), but I've noticed that the accuracy scores are considerably lower as well (approximately 10% less for the same number of epochs) compared to the results obtained when running the same code using the CPU. Is this decrease in accuracy a common issue when using the Metal plugin on an M2 Max chip? How can I resolve this?

I attempted to update the versions and use different platforms for Conda installation, expecting to improve the accuracy of my model when using the Metal plugin with TensorFlow on my M2 Max chip. However, despite these efforts, the model's accuracy still significantly decreases.