We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor. We ask that you disable ad blocking while on Silicon
Investor in the best interests of our community. If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.
One of the articles I linked recently, directly commented on that issue, but it was wrt to the chiplet die attached to the substrate, in order to get all the chiplet surfaces at the same height for the heatspreader. The way they did it was via careful control of the copper posts they are using for the interconnects. They claimed it gave better stack control than plain solder bumps.
I think by far the big reason for chiplets is yield. It allows a 64 core chip to be implemented on 7nm but with only 70mm2 die or so. That can yield, whereas Intel's 500mm2 monsters can't.
As to whether there is any thermal advantage, that I'm less sure of, because a little thickness in the heatsink at the contact can spread that laterally too. As far as the die is concerned, the heat per unit area is approximately independent of adjacent heat sources, its assumed the heat all goes normal to the surface and into the heatsink. Its just the delta between the silicon and the heatsink that counts.
IMO the best reason for chiplets is to get heat out of the area of the CPU.
I'd say that is a really good benefit, but the best reason is chiplet yield -- vs monolithic large dies -- and all of the binning options that come with it. Also, chiplets allow you to update one kind of chiplet in the package without having to update an entire large monolithic die.
Not to mention, that the same CPU chiplet is used in all of AMD CPUs from low end DT to high end server. This time around it might be true of the APUs as well, but we don't actually know that yet. AMD has saved a lot of money having one common die.
It also makes "custom" chips a lot easier to do. I have no idea what the NREs are on an organic interposer, but it has got to be a whole lot less than a new mask set for a processor. Even if a new chiplet needs to be designed, the total costs are still a lot lower.
in principle, yes. I don't know what the software stack is, nor what the workflow through the tooling is like. In general, I suspect it is somewhere between a 10 layer PCB and those multilayer ceramic modules the military at least used to use all the time. It isn't like I can fire up Kicad with a plugin for these things...
I doubt if it is cost-effective to do, say 12 units unless it is for a military contract. You probably don't need a volume of millions of units a year for it to make sense. Does it make sense for a thousand units? Ten thousand? For those numbers, that opens a lot of possibilities. And could spell the end of the traditional merchant chip model for servers. Especially if other companies can use Infinity Fabric for their own designs.