I'm very surprised that Nvidia chips don't have memory size locked in hardware.
Due to chip shortages, most likely, we got the smaller amount of RAM on production models. This leads to weird situations where the 3060 has more RAM (12GB) than the 3070 (8GB) and 3080 (10GB). nVidia must use RAM in multiples of the bus width, so the only other option was 16GB for the 3070, which is what this person enabled by swapping chips.
nVidia doesn’t care about modders swapping memory chips because it’s not a threat to their sales. I say this as someone who has all of the hot air rework equipment and experience to do this swap: It’s not going to become a common mod. It’s likely not even feasible for someone to offer as a service due to the risk involved in shipping expensive GPUs two ways and then reworking the PCB with hard to source chips that can’t necessarily be guaranteed to work at the exact speeds expected by the drivers.
It’s also possible that this mod will be rendered useless if a future driver update introduces different timing parameters, such as when nVidia releases official 16GB 3070 models.
That's too bad. I prefer to buy a top-end graphics card once every 3 or 4 generations. I was up for a refresh with the 20-series, but the value was pretty bad so I waited. Now the 30-series looks great except for one glaring issue (aside from availability): VRAM capacity. With the imminent release of the direct storage API for Windows, I expect the 30-series cards will show their age much faster than previous generations thanks to their relatively small VRAM capacity compared to Radeon and game consoles.
Maxwell cards could also be upgraded from 4-6GB to 8GB. This has actually been a cool mod for a long time, earliest I can remember are the ATi Radeon 9x00 and nVidia GeForce 4 Ti cards, people used to double the RAM for fun and profit.
But beyond that, what would be the reason of spending effort locking this down even more?
Famously, many discrete GeForce cards don't play well with virtualized OSes, and there have been some hacky ways to bypass that. This is done because only Quadro cards are blessed by Nvidia to run in "enterprise" environments. Besides this soft lock, there's usually no architectural limitation preventing GeForce cards from working with virtualized OSes if the hypervisor can provide passthrough.
All that to say, it wouldn't be the most shocking news in the world if it turned out they prevent this kind of VRAM modding.
A big fraction of Hackaday's content is this kind of fun modding that might not make immediate economic sense.
[1] https://www.hardwaretimes.com/doubling-graphics-memory-will-...