I was thinking about about the PS4's GDDR5 memory.
174GB/s of bandwidth. I think it's a load of crap and another case of Sony providing us inflated theoretical best case numbers.
GDDR5 moduels I/O is 32bit. 16bit is for reads and 16bit for writes. In many high end PC GPUs the GDDR5 modules bus is split. There's two memory controllers, one handles the 16bit input for writes, and the other controller handles the 16bit output for reads.
The downside of the above of course is that you're doubling up for the most part on busses, running two 16bit busses instead of one 32bit. You're also running two memory controllers, one for reading and one for writing. These two factors significantly increase the manufacturing costs, as you have to double the memory controllers and busses increasing die size, complexity, and PCB layout.
In cheaper GPUs there's simply one memory controller, and it has to handle reads and writes. It cannot read and write at the same time, it can only done one at a time. This design is a lot cheaper, simplier, and cost effective to manufacturer, but the downside is bandwidth is significantly reduced.
I doubt very very much the PS4's GDDR5 is running with its bus split to allow for simultanious reads and writes.
The 174GB/s would be the theoretical max, in a ideal read or write only situation. In the real world when you have to factor in real I/O use, overheads, memory controller efficiency, i doubt the PS4's GDDR5 can pull anywhere near 174GB/s
Which also makes sense given the GPU is stated to be only 1.8 teraflops.