Hacker News Reader

It’s a massive supply crunch. More production will come online.

Probably not, at least for DRAM. The demand has historically been very variable, and building production capacity takes multiple years. Also, spare capacity is really expensive. Thus the memory manufacturers don't want to expand, betting on it being yet another temporary bubble.

Also, DRAM fabs are not really usable to make compute (CPU, GPU, etc in this context) silicon. The production lines and tech have diverged some decades ago. So unlike TSMC which can relatively easily retool for another customer, no such luck for the big DRAM manufacturers.

How did Apple figure this out? Isn't the solution to this to sort of evolve to a unified memory architecture, which wins in speed and cost anyway?

I haven't looked at apple specifically, but generally the approach taken with HBM on GPUs etc is to make multiple chiplets and connect them together very close and with very high bandwidth. The chiplets are made with different processes, optimised for each specific use case.

There are a few different variation, such as having a carrier chip below (as an interconnect), having a small PCB that everything is mounted to, or even vertically stacking chips (AMD's 3D v-cache does this on some of their CPU models).

Reply on news.ycombinator.com