My approach is to consider the problem unsolvable and will wait until there's enough information to go on, which in this case there will never be.
Such an approach isn't suitable for any kind of prediction/design. One has to take the data which is available, lay it all on the table, weigh the pros and cons of different options.
In the months since talking about this, some of the knowledgeable members have introduced some interesting alternate hardware.
My take on such an approach is this: If the hardware is fast enough to emulate a foreign architecture in realtime, then it would seem a suitable upgrade to dump any legacy architecture as the new one has proven to thoroughly outclass the alternate.
Similar to how there are emulators out for many older consoles which can run on modern PCs and some even run on a smart phone.
Obviously, ARM doesn't have a 6800 or z80 in it, yet the ARM used in most android devices is so fast, it can emulate the hardware without breaking a sweat.
When someone can introduce an architecture which can emulate Xenos/Cell without issue (and within the die budget) then it will have proven to be a suitable replacement and at that point I don't care if it's x86, MIPs, Alpha, Cray, or ARM.
As long as it fits the budget and isn't taking away die space from the GPU.
All I can talk about are generalisms. If you won't accept them without these impossible datapoints then that's that. Especially when you've decided BC is just a matter of picking a CPU with a compatible ISA.
Not just a compatible ISA, it would also have to have VMX128 (or better), the cache locking feature (or suitable replacement), at least 6 HW threads, and the ability to chew through each thread at least as fast as Xenon so that games aren't stuttering along.
Cell is a different animal, as your Cell emulation thread found, the SPE is pretty much impossible to emulate on a mm vs mm basis. So, Sony can:
1) Completely ignore the desires of Devs that have complained about Cell and literally just ramp it up (6-16SPE)
2) Modify SPE's with larger local stores, OOOE, or more registers.
3) Shoot themselves in the foot and drop SPE's and deal with the repercussions of destroying the concept of a platform that Sony had built up since 1994.
I personally think Sony should keep the SPE's on board and find a use for them. Even if Devs NEVER have to directly access the SPE's in the design, Sony should still include them and find a way to easily utilize them.
- Sony could institute a mandatory hardware AA method which utilizes the minimum 6 SPEs.
- Sound obviously fits in the SPE mold.
- A Physics Middleware kit which runs on the SPE's would also be a great venture for Sony to handle for their devs.
With Sony instituting the minimum 6 SPE's needed for compatibility, they can use the rest of the die budget for Better/More PPE's. Bottom line, Cell is an interesting and potentially dominant performer. To dump it would destroy BC, would negate all the code bases and technique lessons which devs have developed over the past 6 years, would be a waste of all the Cell R&D investment up to this point, and a waste of the IP.
All of this, for what alternative?
If it's something which is powerful enough to emulate Cell (not likely), then great. If not, seems a bit backwards innit?
Why doesn't the GPU get a mention?
The XENON is indeed a tricky proposition with all of the oddball functionality which was thrown in, MEMEXPORT, the tessellator, not to mention the EDRAM/logic. However, AMD/ATI is handling the new GPU, so I'm assuming they know their baby better than anyone and will not have an issue replication the functionality with GCN's more flexible and programmable nature.
GCN/Tesla should have Zero issue emulating the machine code of RSX. Sony may just have to pony up for the licensing fee to Nvidia if the go with GCN.
Now you've shifted the goalposts a little. If you agree that Cell is better than Xenon, and PS4 with Cell2+GPU will be better than XB3 with Xenon2+same GPU, then there is a cost for scaling XB360; it'll lose performance.
No, it actually isn't losing performance. Xbox never had Cell.
I'm not convinced MS would be willing to license the tech from Sony and vice versa. Nor am I convinced it is CLEARLY the better CPU. Harder to dev for, but higher theoretical throughput. I've explained my vision for how to handle Cell in ps4 above.
It could be the better CPU, but many of the performance advantages which Cell has can be lessened via VMX type execution units.
It's up to Sony/MS to take advantage of the CPU architectures they have and improve/modify them where they feel necessary. But I'm not convinced either architecture is so inefficient that it would be better to start over from scratch rather than modify/extend what is already there.