192-layers NAND doesn't have 192 stacked dies. It has the NAND storage array patterns repeated 192 times on a single piece of silicon, no need for TSVs. If DRAM could be designed for manufacturing in a manner similar to NAND, that would allow over 128GB per DRAM chip even if half the DRAM cell density was sacrificed along the way.https://videocardz.com/newz/samsung...emory-modules-with-up-to-512gb-total-capacity
Samsung already plans 8 layer stacked RAM using TSV on DDR5 Memory modules.
I hope you have the $$$ to pay for all that RAM.
Key word in his description: "per package"So, your challenge is to explain where you get the additional density to push DRAM capacities that far.
You can stack multiple dies of conventional RAM inside one package, be it using TSVs, shingled wire bonding, mezzanine interposers or other means.
An active DRAM chip uses about 250mW while idle chips are around 10mW. If an E3.L card is connected to a 5.0x4 interface and populated with hypothetical 256-layers / 512GB DRAM chips, I doubt more than 1% of it would be active at any given time, so we'd probably be talking just over 1W per chip on average. Fine, that would be only up to about 60 chips per 70W E3.L card instead of 80I think you still have a problem with cost and power, though.
Not to mention the processing resources and bandwidth needed to actually use it. We're talking RAM - not storage, remember? If you don't have enough bandwidth to even populate all of it between reboots or whatever, then it's pointless to have so much.
Cost-wise, 192-layered NAND is found in TB-scale SSDs under $100, so stacking repeating layers of memory cells clearly isn't particularly expensive for NAND structures. If a similar multi-layered construction was developed for DRAM, I'd say there is a pretty good likelihood it would become much cheaper than any form of die-stacking, especially if it has ECC so the OS can flag pages with flaky bits to avoid them in the future and we accept something like 0.1% capacity loss as normal.
As for actually using this much memory, Frontier has 9.4k nodes at 512GB per CPU + 4x MI250. We are 1/200th of the way there!