Costs of developing packaging for a new cache die, the packaging and design of all the vias and such that were not there before would be substantial and only for one die type they can sell.
I don't think they are at the point with these chips where that makes sense. On package ram would be far cheaper and simpler and probably help more with 3d work.
It's almost like a custom package has to go through all these various steps for a singular die to begin with. Maybe because it's a custom solution for a custom design. Mind-blowing!
It's almost like a custom package has to go through all these various steps for a singular die to begin with. Maybe because it's a custom solution for a custom design. Mind-blowing!
We are talking things of different scale here though and development costs. You need to retool your 3d packaging facilities, get the cache die and go through all the verification. That's on top of all the usual processes. It would make them a lot more expensive and think about it, for 3d applications you benefit more from bandwidth and size over latency which on package memory would serve a lot better.
That vs sticking your current chip on a different package for BGA and never putting a lid on it for mobile applications is very different.
With the current chiplets that same die will be used in threadripper and epic for the 3d vcache so more and higher margin markets amortise that cost.
It's almost like a custom package has to go through all these various steps for a singular die to begin with. Maybe because it's a custom solution for a custom design. Mind-blowing!
Oh interesting, on the other hand the strix halo is a new I/O die they can connect to, so they would be able to pair that with x3d CCDs, I think it's unlikely they will still as it would only improve CPU performance which they will have enough of already. But they could if someone actually released competition.