We won't be working on an AI processor for a number of reasons,...

We won't be working on an AI processor for a number of reasons, not withstanding that you shared that this memory is apparently vulnberable to temperature (clarification needed from 4DS).

Lets break the the semiconductor industry into a number of diffirent types of companies.
1) IDM
2) Foundries (Imec provided this service)
3) Fabless
4) Designers (Imec provided this service)
5) IP (4DS)

All 4DS is here to do, is to sell its IP, thats it. 4DS are not a design house, not here to package it, not here to manufacture. We want to offload it, get paid, Ting moves on to his next contract / project, David goes back to building more renos.
Now.... if 4DS decides to pivot towards a licensing model, similar to Weebit - that changes things but so far there has been no indications of this.

So who would be interested? Companies in the (1), (3) and (4) category - which of the companies in that category, well that depends on what it can be used for.


This forum has gone from throwing around universal memory to flexible memory... but I've never heard of a temperature sensitive semiconductor as being "flexible". It completely rules out anything AI related as it would need to be away from everything else and therefor not solving the bottleneck issue.

There's a reason why DC's and NoCs have been forced to start using custom server racks from companies like https://submer.com/ - they're struggling to keep them cool while a normal DC / NoC likes to sit around 20-24 degrees, anything processing AI is in a whole other class of power consumption and heat. Server racks are now being kept in a viscous coolant to manage the heat problem - just like a car engine needing coolant... but sure, we're aiming for the AI market when we aren't suitable for MCU's.


This is why I stated that its the difference between buying a lemon and buying a ferrari. If we are that heat sensitive that we cannot be used in MCU's (which generally sit on the other side of the firewall in a vehicle, shielded from the engine), then our use-of-product drops drastically to the point of where we would be lucky to get 15-30c per share. If we're not heat sensitive, then $1 per share minimum with no upper range.

On the whole thing in regards to retention - the presentation clearly states from hours to days. Now Vinneh to play devil's advocate, "days" could be taken to mean 2 days, 40 days, or 360+ days. Depends on how much kool-aid someone has been drinking. What it does mean is that this cannot be used for SCM and arguably JD was never really targeting this space as "NAND will always have its place" and was more so to augment it with ReRam instead of DRAM currently, with its ability to scale down significantly you could have a significantly large cache. People talking about it being able to emulate SCM and other random quotes from David... non-sense. What are they going to do, package it up to a CMOS or lithium battery to send a pulse every day or so?


Lets focus on what we do know - the target is embedded memory one way or the other where retention is less relevant. How much of that embedded memory that can be targetted now depends no functional integrity in regards to temperature - because this may rule out SRAM, SoCs, MCU's, etc.

I'm all for everyone drinking the kool-aid, but this is a issue that needs clarification.
Personally hoping that Peter was mistaken in what he said or that he was misunderstood.