[question] desktop PC with non ECC RAM and lightGBM

[luigimacedonia]

I am (starting to be) aware of the risks of using non ECC memory. I read that basically you cannot trust anything that was generated by a computer without error checking memory. I also read that DDR5 performs a basic check inside the chip that guarantees that it will be correct while the data is still inside the chip (no warranties after it is transferred to the CPU)

Now I would like to know the scope of corruption I could get while training and predicting for instance a classifier with hundreds of features and several millions of samples in the following scenarios:

1. DDR4, 'device_type': 'cpu'
2. DDR4, 'device_type': 'gpu'
3. DDR5, 'device_type': 'cpu'
4. DDR5, 'device_type': 'gpu'

As a hobbyist, I am using a desktop PC. I am thinking 2 options: to move from Intel to AMD Ryzen Series 9000 to -at least- get unbuffered ECC support or go full for a Threadripper, which is a cutie but it is 3X more expensive. That makes me wonder:

5. does a faster processor improves the performance of training a model while using 'device_type': 'gpu' ?

thanks

(I didn't know where to post this question. It was deleted from stackoverflow because it doesn't contain a code question. please let me know if it is making noise here and should be deleted)