We know that Intel has discrete graphics cards in the works, but not much has been known about their Xe architecture until now. Leaked Intel documents have given us quite a lot of new information about the Intel Xe architecture. These documents show just how Intel plans to release their graphics cards and attack both NVIDIA and AMD on the consumer front and enterprise front.
The first thing is that Intel Xe GPUs will use “tile” modules, which Intel does not call “chiplets”, although Intel did reveal back in November that Xe cards would use a multi-die system (MCM) by way of Foveros 3D stacking.
Looking at the leaked document above we can see that Intel is planning 1, 2, and 4-tile cards. Intel did show their DG1 SDV (software development vehicle) at CES this year and it is expected that it will have 96 EUs (execution units), but it has been rumored that each tile may feature up to 128 EUs. This means that the DG1 is not a full tile design, but rather a cut down version. Intel has divided these cards by number of tiles and TDP. So 1 tile cards will be 75W (DG1) and 150W, a 2-tile card will be 300W, and the maxed out 4-tile card will be 400-500W.
Given this information we can speculate that the 1-tile 75W card will be the entry-level card, the 1-tile 150W card will be the mid-range card, and the 2-tile 300W card will be the high-end consumer card. This product placement will directly complete with NVIDIA and AMD at those levels.
Looking at the input voltage of the 4-tile card (48W) we can tell that this card will be for datacenter and enterprise, not the consumer market. Intel is calling this card Arctic Sound. It is also interesting that (according to the documentation) this card specifically will make use of HBM2e memory, which is the latest evolution of HBM memory, which both SK Hynix and Samsung will launch in 2020. The documentation also shows that the card will be PCI-Express 4.0.
Via Digital Trends