More information about NVIDIA’s upcoming “Pascal” GPU architecture was revealed at GTC Japan. This new GPU architecture is designed to offer nearly double the performance per watt over NVIDIA’s current “Maxwell” architecture. This will be made possible by implementing stacked high-bandwidth memory 2 (HBM2). The flagship “Pascal” graphics card will feature four 4-gigabyte HMB2 stacks totaling 16 GB of memory. This will make the memory bandwidth 1 TB/s. The chip itself will actually have the ability to support up to 32 GB of memory. NVIDIA’s enterprise models (Quadro, Telsa) will most likely max out the memory.
NVIDIA’s Pascal chips will be built on the 16 nm FinFET process. NVIDIA is also working on a new interconnect called NVLink, which will change the way they have been building dual-GPU graphics cards. Current daul-GPU cards are more or less two graphics cards on a single PCB with PCIe bandwidth from the slot being shared by a bridge chip and an internal SLI bridge connecting the two GPUs. Now with NVLink these two GPUs will be interconnected with an 80 GB/s bi-directional data path which will let each GPU directly address the memory controlled by the other. This new change will improve the memory management in games that take advantage of newer APIs like DirectX 12 and Vulkan. Expect to see the first NVIDIA Pascal products in the first half of 2016.