7500+ pins!!!
How long until they start ‘sandwiching’ the CPU in the socket to have contacts on both sides of the chip?
You forget about those massive amount of pins on the top side has to be connected to something too. You’ll ended up with a sandwich with large amount of connections on PCB on both sides which makes cooling worse.
Better sticking to PCB on one side and heatsink on top.
I don’t see that happening anytime soon, if ever. Cooling would be an extreme problem.
Medium term, we’ll probably see at least 16 channels and 128+ PCIe lanes and TDPs up to 600-800W in the Venice/Diamond Rapids gen. Should push pin counts to ~9000. Beyond that, it becomes tricky. At a certain point, there’s so much IO you literally cannot fit that many standard DIMMs lined up in a typical ~20in server blade. Will need some way around that.
Long term, optical might be a solution. Have some super high bandwidth optical links on the main package and break out PCIe or whatever else on some external chipset.
Fan-out PHYs will likely make a return, in some form or another. Optical would be a good route.
How the hell do you cool it lol. Maybe something vertical around the circuit?
a layer of heat pipes in the cpu ofc
Can’t wait for Turin-dense vs Clearwater Forest. These sockets are even bigger than this.
At this rate…one day in the future dropping a paper clip can potentially bend the pins…
/s
Note, the title says Sierra Forest, but this is presumably the Birch Stream AP socket, so it will be used for GNR-AP as well.