| EricHiggin said: Either way SNY should end up with considerably more usable silicon overall, which could very well keep the price down. It may also be part of the reason why they can crank the GPU clocks so high. If the dies are separated on the same interposer, or completely separate on the mobo, that would also help with heat dispersion. I would also think this should make the PS5 upgrade version(s) down the road simpler and cheaper for SNY. |
Let me explain my "waaay too expensive" bit a little further.
First of all, "expensive" in the technological sense does not only mean $ spent, it also means power required (when chip designers talk, they usually think of the second part solely).
Chiplets come into play when you think in terms of a "construction kit" kind of building a whole series of systems. Depending on what the target is, you add chiplets to your designs until you reach your goal. If you have a goal in your mind like "Different (amounts of) cpus interact with different (amounts of) gpus", then chiplets is the future way to go, no question about that.
A console has one single goal, and one single goal only. It is the exact opposite of a "construction kit". It consists of a single set of masks for an SoC or apu, whatever you want to call its core piece. Any set of masks is a very expensive thing to calculate and manufacture, so you want to get away with as few sets as possible.
Now if you propose to make separate dies on an interposer, or even worse, separate on the mobo, the first thing is you need to have is three sets of masks. One for the cpu, one for the i/o (a rather simple set though), and one for the gpu. Notice in a console, none of those mask sets are off-the-shelf things, as all three dies are custom tailored. So on the $ spent side, we are already worse off with separate dies (we leave out the interposer thingie which also addds costs). Next is the fact that an i/o chiplet is just that, a thingie that basically has tranceivers that can send/receive data to/from the other chiplets. That thingie needs power to work. If you put your chiplets completely separate on the mobo, you can save the interposer, but you need a much more powerful i/o chip - as soon as you go "outside" of a chip, your tranceivers require a lot of power. Simply put: More distance=more power required.
Short summary: a single SoC requires less power than a chiplet design, is easier to "maintain consistency", and all things considered costs less to make (even if there might be a small redundancy loss when you have to toss away an entire chip if either cpu or less likely the gpu part malfunctions).
It will be interesting to see what's in store in a few years if there is another generation of consoles. I think we will likely still see SoCs with lots of coprocessors in them. Unless AMD starts building "coprocessor chiplets", something like i/o combined with Sonys ssd/Tempest coprocessors, and other yet to be invented sutff.
