AMD has announced a new card in the Radeon Vega product line — The Radeon Instinct Vega. The Radeon Instinct Vega GPU has a 7nm die, a first for a GPU. AMD originally intended to release 12nm GPUs, but they skipped over that and went straight to 7nm instead, doubling the density of the Vega GPU die (it used to be 14nm).
The new Instinct Vega is intended for commercial uses such as the booming machine learning and A.I. cloud service industries. Hence the 32GB of HBM2 memory, which would be a bit much for a consumer GPU anyway. The 32GB of HBM2 memory puts it on par with the updated NVidia Tesla V100, which will have 32GB as well.
The new Instinct Vega GPUs pack a punch for their relatively small footprint and are more power efficient, which is good news for multi-GPU setups. Datacenters can easily contain hundreds (or more) of computers, and possibly even more GPUs, and therefore consume colossal amounts of power.
To Put GPU Power Efficiency Into Perspective: A datacenter with thirty-thousand 250W GPUs in it would require up to 7,500 kW for the GPUs alone.
Further Reading: GPU Power Consumption
The AMD Radeon Instinct Vega is slated for release in the second half of 2018.