Reading conflicting info on limits online. Sources that spout we can still cram even MORE transistors into finite space cite 3D constructions, but wtf have we been building all along then? 2d?
>>17048658>wtf have we been building all along then? 2d?Yes.
basically, new tech is what they call all around gates
>>17048658the handwavy answer is semiconductor devices are made from a chunk of silicon crystal by structuring the top one layer at a time.primitive cpus had under 3-10 layers (depending on how you count). modern cpus have 30+ layers.fabricated 3d structures using lots of layers is challenging because the layers have to be very well aligned. 3d structures also have more difficulty diffusing heat.
>>170486582032 is when they will start to build up on the node. Future chips aren't pancake.
>>17048658>Sources that spout we can still cram even MORE transistors into finite space cite 3D constructionsIn theory, sure, in practice your thermal budget would blow up. Cooling is proportional with surface area, heat production scales with area and active layers.
>>17049392Newer chips have reduced total power consumption from parasitic losses and decreased the thermal resistance from the chip to the surface by moving all of the power interconnects to the bottom.
>>17049392How about using something besides silicon that can handle the extra heat better?
Hand the 3D All-Around Gate problem to the AIs and let them improve themselves for a few generations. Win-win.
>>17049392Hey AI, I need a fractal 3D structure, all surface area, little volume, full of logic gates. Go.Next.
>>17048788retard here, most of those layers are a bunch of fuckoff wires in the BEOL right, my understanding was the layer count isn't actually the complexity but its the underlying design of the die that requires all those layers
>>17051198Optical computing has been in R&D for decades. Like so many alternatives, it's difficult to get it beyond the infant stage because silicon is so far advanced that optical computing and other alternatives would have to be subsidized to an extreme degree for decades before they could hope to catch up. There's also Gallium Arsenide (GaAs), which also has been in R&D for decades, with some limited uses. It handles heat better than silicon but is three orders of magnitude more expensive. Maybe if it was used initially instead of silicon, it would be the standard now, but like optical computing, silicon has such a huge gap in knowledge and production that it's hard to develop anything else.
>>17051274Yes, transistors are only in a single layer and that is where most of the heat is generated. The rest is wiring, passivation etc. and is shown in >>17049422. So if you now have two transistor layers, things will heat up quickly.>>17051198Sure, look up isotopically pure silicon. Very, very expensive but super performant. I do wonder who uses this.
>>17051326how about Tin? that's in same column as Silicon and likely more heat tolerate.
>>17051420>look up isotopically pure silicon. according to online searching, primary demand for Si-28 comes from Silicon Quantum Computing. Quantum processors use spin qubits that are sensitive to magnetic fields. Natural Si-29 has a nuclear spin ($I = 1/2$) that acts as magnetic noise, causing qubits to lose information. Si-28 has zero nuclear spin ($I = 0$), creating a magnetically silent environment. The interest in Si-28 is driven almost entirely by quantum coherence, not cooling.
>>17048658Moore's law is death, parallelization and self optimizing algos are giving productivity returns that are the equivalent of 5 OOMs impeovements each 2 years
>>17051667>according to online searchingI would not expect NSA or Pentagon to announce the use of this material but I would expect them both to have plenty of use for such devices.
