There are many memory manufacturers that Nvidia uses today for its computational parts and GPUs and the memory will come from one or more of them. This is not a big surprise, but it is good to clear this up, as Nvidia doesn’t want to produce its own memory when other companies are quite good at that.
The memory will come on the package together with the Parker chip and we can expect the usual amount of stacked memory that will be appropriate for a 2016 GPU or computational part. In other words we will see a multiple gigabytes of memory placed in the package.
Stacked 3D RAM will be the only memory on the board and we were told that there is no need for additional memory in the form of BGA chips that we see on today’s graphics cards. All the memory comes as 3D stacked RAM and it's powered by NV Link, so it should be able to "talk" to the GPU at very high speeds.
The exact figure is unclear, but it should be 2.5 times faster than the memory we see today that will get us to 800GB to 1000GB per second, quite a step up from current 288GB/s that Nvida has with its pricey GTX Titan card today.