3 ms·
The next "total rethink" of GPUs is multiple dies on a single package. Not just for HBM, there's no reason you can't have the dies "hanging off" the interposer
by paulmd 10y ago
The next "total rethink" of GPUs is multiple dies on a single package. Not just for HBM, there's no reason you can't have the dies "hanging off" the interposer (supported by an inactive substrate), only overlapping the interposer on the regions necessary for chip bumps. Essentially each die becomes a large SMX Engine or other hierarchal unit that is marshalled by some central memory controller or other controller that presents the illusion of a single GPU rather than SLI/Crossfire.
This is actually the direction AMD is going with Navi, not because of the performance gains per se but because it helps yields big time. You can pre-bin your chips, then stitch a bunch of small chips into medium chips while keeping your yields high.
In theory you can scale up for quite a while. At the long term, you will eventually be limited by clock degradation/signal propagation time however.
The short-term problem is heat and power. This doesn't help efficiency gains per se. If you are stitching together four 600mm^2 dies then you are going to be pulling 1000-1200W and dumping that back out into your cooling system. For US consumers, their circuit breakers are a much more immediate limit. Most household circuits are 15A @ 120V, and that's an instantaneous limit. You are not supposed to continuously pull more than 80% of a circuit's instantaneous rating, so that's 12A (1440W at the wall). Factor in the losses from the PSU's 80% efficiency and you are now talking 1152W continuously. Again, that works out to about four 600mm^2 GPU dies, plus some power budget for the CPU and so on. And you'd probably need to take drastic measures to keep that cool - that's a lot of heat in a small surface area.