Up to 24GB and 819GB / s per stack

SK Hynix announced that it has completed development of its HBM3 memories, “the fastest DRAM memories on the planet.” As revealed by the company, they have increases in bandwidth of 78% and capacity of 50% compared to HBM2E, being the ideal option for the next generation of cards oriented to High Performance Computing and Artificial Intelligence.

On the bandwidth side, these new HBM3 memories have speeds of up to 819GB / s per stack, up from 460GB / s for HBM2E, marking a huge performance leap. To compare it with some current implementation, Nvidia A100 80GB GPUs have HBM2E of 409.6GB / s and using 5 of these stacks has a bandwidth of 2TB / s, so with this speed of 819GB / s per stack 4TB / s bandwidth can be achieved, doubling the bandwidth to allow the most out of next-generation architectures.

The other big improvement is the capacity. In addition to being offered in 16GB per stack versions like HBM2E, HBM3 will also be offered in 24GB per stack versions, so a GPU with 5 24GB stacks could have 120GB of VRAM without issue, and if all 6 stacks are activated (Nvidia usually disable one) you could reach up to 144GB of VRAM, allowing you to work with huge amounts of data at high speed.

With the design completed, the next step ahead for SK Hynix is ​​to begin mass production. It has not yet revealed details about when it will start producing these memories, much less about when it will start selling them, although considering that we will not see new products from AMD or Nvidia until the end of 2022, it is not necessary to be too rushed.

What do you think about these SK Hynix HBM3 memories? Would you like to see a product for consumers that uses them or do you think that because of their cost it is better if they continue to be used only in products for AI?

Source: Videocardz