The card will have double amount of HBM 2 memory making it possible to make a 32GB card with a faster memory. It would be a two times 16GB card as each Vega 10 chip can address 16GB HBM 2. This would be a record amount of superfast memory, plus the fact that the computational power can double. At least in theory.
Nvidia has GP102 based Quadro card and a deep learning, but expensive, P100 based card with 16GB HBM 2, but putting two Vega 10 chips together, could jeopardize Nvidia’s performance leadership.
Our sources believe that a dual Vega 10 card might get announced even this year and it will be a good match for Professional market as well as for deep neural net learning tasks.
Vega 10 will be replaced with Vega 20, another revisited chip, probably using improved Vega / Polaris technology. It is hard to give a final judgement if there will be a big difference compared to Polaris, but we would imagine that the architecture fundamentals will remain the same. Since AMD / RTG benefits from a memory speed up, Vega 10 should give a good boost compared to Polaris 10 / Radeon RX480 cards and if you imagine two of these with a lot of HBM 2 memory, things should get really interesting.
Realistically a dual card will manage to ship in volume in early 2017, but the fact that AMD’s might demonstrate the technology before the end of the year sounds impressive.