Quote from: QuintLeo on February 21, 2018, 09:05:03 PM

Quote from: Chromexnet on February 21, 2018, 12:20:02 AM

Quote from: QuintLeo on February 20, 2018, 08:41:34 PM

Most HBM usage DOES have the RAM on it's own die.
Reference the AMD "Fury" line, and their current "Vega" line.
I'm pretty sure that's also the case for the Titan V and the Tesla "volta generation" on the Nvidia side.

It's still a bus, it's just a very WIDE bus (4096 bit IIRC?).

Sorry, I ment to say HBM is located on the same silicon. The physical connection between HBM and DDR is quite different so it's not meaningful to compare them. It's like comparing a conveyor belt to a truck - both transport goods from one place to another but that's where the similarities end. Designing a wide bus is not difficult if the physical connection distance is short; Playstation 2's GPU had a 2560 bit bus to its' 4 MB EDRAM memory over 15 years ago. On the other hand with DDR the physical connection is distance is much longer so none of the big silicon giants, Intel, AMD or Nvidia have gone for a very wide DDR bus in their consumer products.

To date, HBM used in video cards is NOT on the same silicon, much less in the same package, as the GPU, though that might change when the Bristol Ridge APUs get released.
It does tend to be clustered a lot closer TO the GPU though, than ram on a motherboard.

HBM and GPU die are connected together via silicon interposer so technically it's all one big silicon. The interposer, HBM and GPU die are manufactured separately before they are integrated together so I'll give you that. That doesn't change the point though; it's much easier to route lots of data lanes when they don't need to though PCB, pins and sockets (like is the case with DDR).