Discrete GPU and HBM2 Memory
Benchlife showed us a new roadmap that confirms Intel's preparation for the new processor series. The new CPUs will feature a totally new processor design using 14nm process, a dedicated PCI-Express link to a separate discrete GPU paired with faster HBM2 memory. At Intel's Technology and Manufacturing Day they already said that they'd be spending more time in creating a more heterogeneous architecture. Thanks to the new layout of the CPU, Intel will be able to 'mix and match' a variety of CPU components and dies based on diverse process nodes on one chip.
Doing this allows Intel to reduce the cost of a monolithic design and switch to a smarter and more efficient design(s) for the future. Future CPUs will use embedded, multi-die interconnect bridges to allow high density and bandwidth die-to-die interconnects which are not only denser, but also brings huge advantages for a lower cost.
Until now, Benchlife spotted two chips that'd be part of the new series featuring a BGA design. The package sizes seem to be quite a bit bigger in comparison to current processor designs, but we know the reason.
Intel Kaby Lake-G series SKUs listed by Benchlife.
Both are quad core designs with a GT2 graphics chip. One has a TDP of 100W, while the other is more power friendly with a 60W TDP. No on-package cache reported because of the use of HBM2 memory, this explains the high TDP rates. The seperate GPU will connect to the main CPU die with a PCI-e X8 Generation 3.0 interface. HBM2 memory is utilized as main on-board VRAM for the discrete GPU. There's no clue if intel is prepping with an in-house GPU or part of the alleged licensing deal between AMD and Intel allowing Intel to use AMD's graphics IP in their CPUs.
The chips won't have integrated PCH as it'll be located on the board. Kaby Lake-G processors are aiming at Two-Chip platforms including a main chip plus PCH. While we don't know much about the new series, it tends to get more interesting as time goes by. We'll keep you informed!