Since we first started getting leaks about AMD’s new graphics cards the big thing we were hearing about was High Bandwidth Memory (HBM). This will be the type of memory that will be used on AMD’s Flagship graphics card in the Radeon 300 series and NVIDIA’s Pascal GPUs in 2016. The whole idea behind HBM is to stack the memory modules to save space and also integrate them into the GPU package itself. This way the communication distance between the each of the modules and the GPU is reduced, which it turn increases bandwidth and power draw.
HBM will help AMD save valuable space on their graphics cards. If you think of a large graphics card like the R9 290X with 8GB of VRAM, you have the GPU in the center and the memory chips surrounding it, which takes up quite a lot of space. Then on top of that you need all of the voltage regulation components etc, so you need a larger card to accommodate all of this.
So what AMD wants to do is not only decrease the size of the RAM footprint, they want to get it as close to the GPU as possible. 1GB of HBM is made up of a single die, compared to four for GDDR5. The HBM footprint is also much smaller at 5mm x 7mm compared to 24mm x 28mm of GDDR5. The HBM memory is also stacked, which saves quite a lot of space.
With the reduced distance between the GPU and the memory you will see a massive increase in bandwidth, it also uses much less power because there are fewer circuits.
Looking at performance compared to GDDR5, HBM actually operates at a lower clock speed, but if you look at the bandwidth figure it is almost 4x better! This is with a lower clock speed and also less voltage.
AMD’s implementation of HBM (called HBM v1) will be limited to 4GB (4 1GB stacks), but they did say more stacks could be added. Their main focus with HBM will be on graphics cards, but eventually we could see it implemented on APUs and CPUs. With NVIDIA already saying they will be using HBM on their Pascal graphics cards it looks like eventually GDDR5 will be going away. What do you think of HBM?