High Bandwidth Memory (HBM) debuted with AMD’s Fury series graphics cards and was quite the success, although it had some limitations. The two biggest were that it was limited to 4GB and only had a maximum bandwidth of 512 GB/s. This is going to change with HMB2 which will introduce capacities up to 32 GB and speeds of 1 TB/s. This will be more than powerful enough for next-generation graphics cards coming from both NVIDIA and AMD.
So looking at the specifications we can see that HMB2 will more or less double the bandwidth offered by HMB1. This is very impressive considering that HMB1 is already 4 times faster than GDDR5. Power consumption will be reduced by 8% over HMB1, which means that HMB2 will have a power reduction of 48% over GDDR5.
HMB1 could be stacked in a 4-Hi stack (4 layers) whereas HMB2 will be able to go up to 8-Hi (8 layers). What we saw on the Fury X was a combination of four 4-Hi stacks, each of which contributed 1 GB to the cards total 4 GB. HBM2 will offer 4GB on a single stack, so if we recreated the Fury X with HBM2 it would have 16GB running at 1 TB/s, which is very impressive.
Both AMD and NVIDIA are expected to use HBM2 on their upcoming graphics cards in 2016.