AMD has already told us how they will create scalable GPUs. There will be a core GPU of say 32 GCN CUs, 128 TMUs, 64 ROPs, 4 ACEs and all of the uncore stuff like encode/decoders, display ports, HDMI, PCIe, etc. In each of the attached HBM2 the logic die will contain an additional 16 CUs, 2 ACEs, 64 TMUs, 32 ROPs. And there will be HBM2 stacks without logic for cases where one just needs more memory. So a single base die and a HBM2 logic die allows GPUs having from 32 to 96 CUs with memories to match. The smallest with 1 HBM2 stack and no GPU logic would take care of entry level VR around what a 290 gets. With 1 stack including the GPU logic die, you get 48 CUs and go above a 290X/390X. 2 HBM2wL stacks gets better than FuryX performance, 3 HBM2wL stacks probably takes out the top Pascal with plenty to spare, and 4 HBM2wL is king dGPU. A larger core die with 64 CUs could attach 8 HBM2wL stacks and have 192 CUs total, 20 ACEs, 768 TMUs, 384 ROPs, 64GB of VRAM and handle 6-10 VR headsets simultaneously.
Pete |
