i just dont get why they give the 16gb models to the 60 series cards but then the base 70 series is stuck with 12gb. Just manufacture a 16gb 5070 you frauds.
When Nvidia creates an architecture, it produces subdivisions that have different uses of the wafer (there are always some that don't come out perfect, but it is still possible to use parts of them).
In the case of the 5000 series, we have:
GB-202 - This is basically the entire chip, 24k CUDA Cores, has a 512-bit memory interface.
GB-203 - This is basically half the chip, 10k CUDA Cores, has a 256-bit memory interface.
GB-205 - Another cut almost in half, 6k CUDA Cores, has a 192-bit interface.
GB-206 - Here basically 1/6 of the chip, with 4.6k CUDA Cores, has a 128-bit interface.
GB-207 - The rest of the rest, with 2.5K CUDA Cores, has a 128-bit interface.
Memories are now supplied in chips with 1 GB or 2 GB of data per chip. Each chip communicates with a 32-bit memory controller.
The 5060 Ti has a 128-bit memory bus, since it uses GB-206: 128/32 = 4, that is, it would be possible to use 4 GB of data if you use the 1 GB chips or 8 GB if you use the 2 GB chips. To get the 16 GB, companies use a technique called clamshell, in which they place more chips on the PCB, dividing the 32-bit communication into 2 16-bit channels. That's why you gain in VRAM size, but the bus speed remains the same.
The 5070 has 192 bits, since it uses GB-205: 192/32 = 6, you would use 6 GB with 1 GB chips or 12 GB with 2 GB chips. If you make a clamshell, you could have 24 GB, but this GPU is faster and therefore generates more heat. Packing more chips would make them into a mini-oven or at least would greatly increase power consumption and raise the price of the board. That's why we only see these techniques in low-end chips, since they consume less power, are cheaper, etc.
The problem is that Nvidia uses the worst chips for the strongest cards. The 3060 used 192-bit buses. The 2060 Super used 256-bit!
There are 3GB GDDR7 chips, which in the future will allow a 5070 Super to have 18GB of VRAM (if they don't change the current bus). Nvidia didn't use them now because they are much more expensive.