theoretical peaks are just that
it is what you would get if there was never a miss of any kind, if there is a miss, there is a penalty. And well, there are always misses, which is why he states it is more around 140/150GB/s in real gaming terms, because there is a miss every 8th cycle or so.
I still call BS on that number. Maybe in in house tech demos using extremely controlled environments will there be a miss only once every 8th cycle, but in the real world gaming environment I think that number will be higher. The more misses, and the higher penalty, the lower your GB rate becomes. The GDDR5 number is 178GB/s peak, but that goes down pretty quickly the more misses you have. Which is why hUMA was important, because it allows the GPU AND the CPU to " see " the thread schedulers and queued jobs and know which one it needs to pull next. So the amount of " misses " are lowered significantly, and also takes some load off the CPU. The fact that the GPU is also side by side with the CPU also means the GDDR5's usual high rate of penalty on a miss is much lower then you would find on a standard PC where everything is separated.
If I had to take a wild guess I would say on a normal PC with 178GB/s memory bandwidth from GPU to GDDR5, your real world bandwidth would be more like 70-90GB/s, maybe lower. But on the PS4, it is likely more around 110-125GB/s. Due to the lower cycle miss penalty and the drop in actual misses and such. Of course developers can write benchmarks if none are provided to run tests on their product to see what the actual performance is bandwidth wise.