fun again:
XBox 360:
4 core PWRficient CPU @ 2.4 GHz with 4MB L2-Cache
1GB ; 256bit UMA GDDR3 @ 800MHz (=51,2GB/sec)
64 ALU AMD GPU (complete GPU with ROPs, northbridge for the CPU etc..) @ 600 MHz (=384 GFlops)
no eDRAM-chip. The 51,2 GB/sec should be good enough for 1280x720 HD
HDD for all versions
quiet and efficient cooling system
Downside:
the 1GB 256bit GDDR3/5 would still be expensive during the lifetime of the console due to the 8 memory chips needed.
I was thinking of something really close in
previous post . Such a set-up could have been troublesome to shrink past the 55nm node though may the generation had a "normal" length it would not have been an issue.
Taking that in account my "last" word on the matter, still the same easy a posteriori line of thinking, knowing where software and hardware was headed +8 years after the design, I would go with something like that:
1 APU + Smart eDRAM
APU: TSMC bulk 90nm process
4 PPC 750 @800MHz
4 Xenos SIMD array + 24 Tex units @ 400MHz
A sound processor.
Smart eDRAM: NEC 90nm process
8 ROPs, 8MB, @ 400MHz
Memory set-up:
128 bit bus to 1 GB of DDR2 800MHz
Same 32GB/s link between the main die and the smart eDRAM.
Power consumption and costs:
IBM has PPC 750 at 16mm^2 on their process, Broadway is 19mm^2. Bulk is denser.
Xenon is ~180mm^2.
I think that the APU would have end in the 300mm^2
Wrt to power, with the down clock of xenos (20%), no matter the extra ALUS and the inclusion of those low power CPU cores, I would no be surprised if such a chip would have burn just a tad more than the Xenos we got (GDDR3 memory controller like their GDDR5 successors may burn more power than their DDR2/3 counter parts).
The smart eDRAM would have been a tad tinier.
-------------------------
The reason behind those choices:
1) speed demons for the CPU were a bad choice, they burn a lot of power, it still show after many shrinks. That was not know at this point in time by the brightest minds be it Intel or IBM, etc. were hoping for 10GHz monsters to take over computing.
2)They were quiet big on top of it, it is tough to get close to their peak throughput, etc.
3) The WiiU is sort of a testament to that, 4 cores (even clocked lower than Expresso 1.2 GHz) could have done the job. In some regards it would underperform Xenon (not too mention the Cell) but when it is all said and done looking at both the silicon and power footprint... Well Nintendo made a wise choice. It is a worthy trade off.
4)Looking at this gen and the PC world, more RAM was better than faster one. In the life spam on the system DDR2 (even 1GB) should have been cheaper than the GDDR3. I could not find when DDR2 PC6400 were release, if not available I guess slower DDR2 could have worked. Looking at that hypothetical system, CPU could not go through as much data as Xenon (not too mention Cell...) freeing more bandwidth for the GPU. Overall I think it was a worthy trade off.
5) bigger but slower GPU=> power consumption would have been improved, not too mention reliability would have been less of an issue.
6) cooling, heat, reliability: we would have a single chip burning north of what Xenon or Xenon alone burnt. It would have been bigger so it would have been easier to dissipate heat, there would have been from scratch a single cooling solution.
7) the eDRAM size. Clearly 10MB looks a lot like a relic of a system aiming at SD resolution +AAx4. Looking at how AA evolved (toward not hardware based solution...) I think that 8MB was just fine, along with a down clock (synchronized with the GPU speed /400MHz).
8MB is all that you need for a forward renderer @720P. A trade off is that even tight G-buffer close to 720p would not have fit. Imo looking at the perfs of a lot of games this gen and resolutions devs chose I don't think it would have been an issue, something like 720x960 could have been the de facto targets for G-Buffer. Looking at the perfs of say Crysis... I'm not sure if it would have been such an issue. PR wise the free x4 AA proved more of a clusterfuck than anything else, actually the whole mandatory 720p requirement was more a hindrance than anything else.
8) Design main focus would not have been about adding crazy custom (2 wide /paired single) SIMD unit to the CPU but on the memory subsystem. Good cache hierarchy, fast access to the memory controller and implementing the silicon on a bulk process.
9) Xenos was the right architecture, nothing to change here. Memexport was forward looking would have been great in an APU with fast communication between the CPU and the shader core (/in a APU)
Overall in some regard that system would be inferior though I think the extra RAM would have made a lot for the system.
