RKL is 14 nm product with 12th gen iGPU.
To be an iGPU all that is required is that is inside the same package and share the main system RAM for working space. That doesn't necessarily require being on the same die. In fact, there are Intel roadmaps out there with Rocket Lake 14nm and 10nm iGPUs.
If you actually look at the chart for some characteristics of a CPU ( like AVX) instead of just hand waving at the Gen 12 GPU stuff you'll see the point. AVX 256 bits. Guess what? Sunny Cove and up are all 512
Intel can put a 14nm CPU die in the same package as a 10nm GPU die and hook them together similar to how AMD hooks 7nm CPU dies to 14nm I/O does. Multiple dies in a package don't all have to be fabbed at the same process. For example, Intel could take a subset of the cooper lake cores (which don't have a iGPU) couple about 10 of those to a 10nm Xe-LP die for the GPU, put them in a page with a shared eDRAM sizable cache and ta-da done. ( maybe weave in some tweaks to boost overclocking and max turbo. )
It has been backported to 14 nm process.
So how do you backport and loose 256 bits of functionality? You don't. Gen 12 is at least of as big of a transistor budget bloat as the Sunny Cove (and up) designs. Trying to port that back to 14nm just gets a much bigger die which isn't going to help intel much.
If that is the case what makes you all believe that 10 nm server parts are really coming in 2021?
That Intel is using advance packaging ( EMIB and/or Foveros ) ... yep that's why. Right there in your embedded chart if you were paying attention. That is going to ship in 2020; not 2021.
Intel is going to take two 10-14 core sized 10nm dies and do a Xeon SP ( and likely W series ) product. Given that they were doing 10 , 18 , and 28 core size does for 14nm going to something around 14 cores is actually less. So it really shouldn't be a major problem for 10nm+ . The volume may be down ( since take longer to "bake" (more steps/passes )) , but there are no big show stoppers here.
Intel's profit margins have a good chance of going down a decent amount, but they should be able to ship server product in 2020 to folks who aren't myopic about highest core count. ( have A.I inference workloads. have live VM images need to move between Intel servers, more even mix of single and multthreaded workloads , etc. )
