Usually, I do this sort of thing somewhat manually, building up an algorithm (mostly classical, with a little ML as a treat) that can deal with the problem.
I'd start by detecting common typos. Typos are similar to un-typo'd data, so I'd do a frequency analysis on the textual representations of manufacturer name and model name, and a Levenshtein distance calculation, then synonymise the obvious synonyms (looking things up when I wasn't sure). The key idea is that you have access to more information than just this dataset: Tony and Tomy are different manufacturers, but Sony and Somy aren't (even though somy is in the dictionary and tomy isn't).
Once the manufacturer and model fields are mostly typo-free (after typo replacement – don't modify the original file, if you can help it!), you can start looking at dimensions and colour. Sort by manufacturer, and start de-duping entries. Once you get a feel for the process you're doing (e.g. under what circumstances do you check whether there's a 102mm Phillips screwthread?), you can start automating bits of it. There will always be special-cases, but your job is to get the data processed correctly, not to get the computer to process the data.
Accidentally aliasing two different products is much worse than leaving the same product described twice, so err on the side of “these are different”. (Keep in mind that manufacturers of some things, e.g. SD cards, often pretend two different products are the same – so you can't always win!) Remember, humans exist: bothering them a few million times is a problem, but bothering them a few hundred would be okay.
When new data comes in, I'd run all the code I used to come up with my system, and see if the output was notably different. If it was, I'd get the computer to let me know.
I'd also add some way for users to flag duplicates. Many humans make light work.