Post #3137704 · Elektrine

2025-12-18 11:47 UTC

@mike805@noc.social You can't do that. On a wafer you always have x% of chips that are defective, and they're all tightly packed together. You HAVE to cut it, put it in a nice little package so it's protected from dusts and so on. A wafer that's been exposed to dust is just an expensive frisbee @MostlyHarmless@thecanadian.social

Replies (2)

@Eldeberen@social.middleearth.fr 2025-12-18 11:56

@ScriptFanix@maly.io @mike805@noc.social @MostlyHarmless@thecanadian.social You may be able to make waffer-scale chips, ignoring the defective areas/chips on it. That's almost what Nvidia do for its GPUs : high-end have a lot of working cores, low-end are the same but with more defective cores. *But* for RAM this make so sens has it must be put along some computing chips to be useful, and I don't think the wafers bought have been design in that way… And with waffer-scale chips you need to deal with a lot more issues such as heat management, dedicated encloruses (won't fit in a standard server) and so on

Open ##3137705
@mike805@noc.social 2025-12-18 16:47

@ScriptFanix@maly.io @MostlyHarmless@thecanadian.social If you have a bunch of RAM and a bunch of GPUs you can just not use the bad cores and the bad blocks. They already do that inside the RAM chips; there are extra lines that can be swapped in to substitute for lines with bad bits. A GPU cluster is a good candidate for wafer scale since it's made up of a bunch of identical units and it's ok if a few of them don't work.

Open ##3137707