Not built for gaming, the Radeon Vega Frontier Edition packs 16GB of HBM2 memory and should deliver 13 TFLOPs of FP32 compute performance, "just" 1.5x more than the Fury X. The FP16 compute performance peaks at 25 TFLOPs, which is a significant 3x boost compared to the Fury X and it comes with 16GB of HBM2, which is four times more.
According to AMD's Radeon Pro site, the Vega Frontier Edition has a pixel fillrate of around 90 Gpixels/s and has a memory bandwidth of around 480GB/s. The same site also suggests that there will be a water cooled version, which is a bit different than the air-cooled one shown by Raja during the presentation.
During Raja's presentation, the Radeon Vega Frontier Edition is a special graphics card with a fancy new shroud, LED logos all around and needs two 8-pin PCIe power connectors, which means it draws over 300W of power.
AMD Radeon Vega Frontier Edition is meant machine learning, advanced visualization, game design, 360 VR workflows and plenty other professional oriented tasks.
Raja was quite keen to note that Radeon Vega Frontier Edition is far from being paper-launched and should be in hands of keen scientists, engineers, and designers by mid-June. Raja also compared the graphics card against the Nvidia P100, showing that it is significantly faster in Deepbench.
Unfortunately for AMD, Nvidia is soon going to have a 21B transistor Volta deep learning card, which is a completely different beast. AMD just wanted to show that it wants to play a role in this market too.