It will likely be a single chip solution with a single Ryzen CCX and loaded up with GPU Compute units... Much like how AMD builds current APU's regardless of segment... A single CCX is relatively small, so it shouldn't blow-out chip sizes of the entire chip when combined with the GPU... Plus it allows Microsoft and Sony to take full advantage of smaller geometry sizes in the future when they become available at the fab.

It will then branch out to a chipset with memory controllers and so on.

GDDR6 is definitely in... Whether it will have extra memory controllers for DDR4 support remains to be seen, maybe in the Playstation 5's case.

It is pretty much confirmed at this point that next-gen is having 8-ryzen cores.
Leaks have pegged base clocks of 1.6Ghz and boosting to 3.2ghz.

Cache wise... Zen+ APU's have 512kb of L2 cache and 1MB of L3 per core whilst the non-APU derivatives have 512kb of L2 per core, 2MB of L3 per core.

Zen 2 retains the 512kb per core, but doubles the L3 per core yet again to 4MB.

So total cache for an 8-core APU is likely to be 4MB of L2, 16MB of L3 if AMD keeps the APU L3 cache half of the non-APU chips and retains the historical L2 sizes.

They might introduce an L4 cache or blow out L3 sizes to assist the GPU, they might not as well. - Not enough information on this front yet sadly.

It won't be using the same GPU architecture... A derivative sure, but it's already been confirmed that the console GPU's will have deviated from the PC's GPU's.

It is likely to be the same situation the Xbox 360 GPU found itself in... It was based on the Radeon x1800, but took design principles from the Radeon HD 2900 and was something fairly unique.

Next-gen is based on Navi, but they have implemented more forward technologies like Ray Tracing, hence why it won't be directly comparable to any AMD GPU on the market today. (Which makes flops comparisons even more useless.)

Consoles can dissipate 300W~ of heat easily enough, PC coolers have gotten advanced and cheaper, consoles tend to rely on such developments.

On the RAM, GDDR6 is by far the most likely, but not the only option. HBM2, while probably too expensive, would also be possible, especially if one of the two wants to make it smaller than usual. And GDDR5X would technically also be an option.

But my question is also on the amount of lanes. If they come with odd amounts of RAM like 20 or 24 GiB and they're using GDDR6, then I'd hope they would also give the memory 320 or 384 lanes to connect them all at full speed. But somehow, I doubt it, especially if they come with 20 GiB.

The cache size will sure be interesting. They could increase it for efficiency, add an external large L4 cache otot, or make it smaller to get smaller, and thus cheaper chips with better yields. Wait and see, I guess.

PC's have still one advantage over consoles for heat dissipation: They're normally thicker, allowing for bigger exhaust fans. 80mm fans are pushing it for consoles unless you put them on the top/bottom (comparable to the side of a PC tower), which is generally not done, while even small, thin PC builds allow for at least some 120mm fan to be installed. So not sure if those 300W, outside of very short peaks, hold up. They might technically be able to dissipate 300W of heat, but you'd better play with earphones then as the smaller fans would be running at close to 100% then to achieve that.

For comparison's sake, the PSU of the PS4 was rated for 250W maximum and that was overkill, at no point the PS4 pulled over 180W even at stress tests. As a result the one for the Slim is only rated for 165W anymore despite the consumption reduction only being about 30% according to Sony. Even the One X PSU is only rated for 240W, and draws less than 180W under full load, so there's no way that comes anywhere near those 300W of heat to dissipate. The consoles are just not meant to do that.

I'd imagine performance closer to an ~RTX 2080 since console manufacturers will have a somewhat higher transistor budget to spend on from using TSMC's 7nm EUV logic node rather than 7nm DUV ... 

I'm personally more interested in what new hardware features they'll bring to consoles compared to first generation RDNA ... (hopefully it's more than just ray tracing)

Possible, the Navi and Zen 2 CPU chiplet would add up to about 330 mm², which to my knowledge is slightly less than the Pro and the X have (350-360 mm² for those) - though I have no idea what parts of the IO chip might be needed. Still, 7nm EUV should push the total below 300mm² if those 15-20% size reduction are to be believed, so there's definitely some more space in the budget.

The question is, what will it get filled up with? RTX 2080 performance seems certainly in the realm of possible, but let's not forget that the raytracing bloats the RTX chips a whole lot bigger than a GTX with same performance would be. I think ~2 billion tensor cores like in the RTX 2080 won't cut it if they want to make raytracing really worthwhile on consoles. I hope AMD comes up with a solution that doesn't need specific circuitry just for raytracing.

I don't think AMD are all that interested in tensor cores and that's not going to help with ray tracing. Those things are mostly used for machine learning frameworks like PyTorch or Tensorflow so Nvidia tried to get a use out of them in gaming by making a library for AA ... 

As for not needing to implement some fixed function circuitry, I don't think that's quite possible considering how optimized Nvidia's hardware implementation is. You would definitely at the very least need some sort of hardware accelerated BVH generation because building acceleration structures for global scene representations is very slow since not even Nvidia can afford to keep rebuilding the BVH every frame so game developers just do BVH refitting on dynamic objects which can degrade the performance of ray intersection tests. Already, game developers need to do hacks even with fixed function BVH generation! 

Even DXR takes advantage of fixed function ray intersection tests so Microsoft/Nvidia obviously sees this as a performance win right there. There'd have to be a compelling case to make the hardware more flexible than just for the current implementation ... (only thing I can think up of are ray traversal shaders and that's pretty much it as far as the immediate future is concerned so AMD are pretty much thinking of heading into a similar route as Nvidia did here) 

Ray tracing isn't THAT bad of a performance hit as long as developers remain reasonable their ray budget. Tricks like using a simplified BVH for many dynamic objects or updating the GI/AO at quarter frame rate/quarter resolution and reflections/shadows at half frame rate/half resolution can make things very manageable. Even taking in those compromises will make for a vastly superior visual quality than what state of the art rasterization can do ... (obviously ultra settings are going to be an issue so maybe developers should consider putting a ceiling on their ray budget) 

I hope AMD doubles down on ray tracing since it'll be well supported in the future. I wish consoles will reach 20B rays/s which is double that of the RTX 2080 Ti and enhance it even further implementing "ray reordering" too for even more performance advantages ...