There are many way to play that games
I'll try based on an "off the radar possibility", a NV only system.
2013 granted
5 cores (cortex A9) as in tegra 3+ (and an IGP) with 2GB of low power DDR2 or DDR3
8 cores (cortex A9)
4GB of high bandwidth RAM
I was just reacting in another thread about a Nvidia based system (not because of odds because of boredom) but it could matched those numbers and the philosophy described in early paper about Durango pretty well.
The tegra 3+ (or more recent) runs the OS, apps, and kinect it's connected to its own 2 GB of RAM.
Then we have a second SOC that include a mid range Nv GPU and 8 cortex A9.
This SOC is connected to 4GB fast RAM (either DDR4 or GDDR5).
I chose A9 versus A15 or something custom NV may or may not have because they are really tiny and don't consume much.
If I look at AMD offering (and up coming offering) I would be surprised to see SMT making it in a CPU any time soon. There is no known implementation of SMT on RAM (for now at least).
So to have 8 cores you indeed need tiny cores. More than that you need really low power cores.
Actually I'm not sure even up coming Jaguar cores would be low power enough if you plan to have 8 of them at decent clock speed.
Nvidia said that they are planing for GPUs integrating CPU cores (in Maxwell or another project?) It could make sense. Anyway they know how to put a SOC together.
Overall it would not be that much silicon and power. Tegra 3 is tiny ~80sq.mm @40nm and cheap. THey could use ore recent iteration of tegra though I don't expect them to be big/costlier.
Something like half the Nvidia actual high end GPU + 8 A9 and there L2 would be definitely sane (in size and power consumption) using 28nm lithography.
Anyway off the radar though lol.