I was at Semicon Japan last year in December and I learn about the coolest semiconductor company (non profit research organization) ever - MinimalFab [1]. There isn't much information their website but this [2] video explains what MinimalFab is about. Essentially, it is a cleanroom-free, modularized fab where each process step in a fab is like a little ATM machine. Miniaturization of complex fab processes is mind blowing and everything is contained inside the machine including a Class 100 environment. Load a tiny quarter sized wafer in a cassette to process and move material from one machine to another. This kind of a fab setup would be incredibly useful to R&D fabs in universities and small scale fabrication for military, space, defense and perhaps even hobby use.
It seems like this fab is a demonstrator, and is planned for
sale. This makes me wonder if/when this will be available for maker style setups, at which process nodes & price per piece. Would be nice to be able to design some fully custom ip, have it tested, bonded and packaged, and optionally soldered, like you can have it with pcb services today.
Which i think of as more than sufficient to finally being able to implement something like SCED, WAM, CHERI, Applecore, whatever in whichever way one is able to wrap his brains around it. Asynchronous, fully static?
That's cool, but your [1] seems to say that the machine uses a wafer 12.5 mm in diameter, which is is probably less area than any of the CPUs in your [2].
Maybe not comparable in die size for all the chips mentioned there, i don't care so much, because i don't want to clone or emulate them. I want to go simpler. Rebranch from the 70ies so to speak, to take all the roads not taken since then. Just to see what's there :-)
That's 122mm², which is larger than some Pentium II dies, so if you're okay with designing to a circular die[0], you could fit a Pentium II-grade CPU on each wafer.
0: The main reason to use a rectangular die is that they tesselate better to fit many dies per wafer.
That is exceptionally cool. They have taken existing best practices in semiconductor fabs and migrated them to an interesting place.
What is missing of course is what their feature sizes are and the part at the end where you cut the die and package it. It is those things that would define the set of things you could put on a single chip.
Ok thanks for that. So 500 nm today, moving to 250 nm by the end of the year or early next year.
So your 12.5mm wafer can have a 8.8 x 8.8 mm square inside of it, or 78.125 mm^2. If I did the math right that is on the order of 156M transistors given a 4t ram cell that is about 39 million bits of RAM. So basically a pretty useful amount of space for "jelly bean" type applications. A synchronized fab line with a median processing time of 1 minute can produce 60 dice per hour. Assuming a physical plant cost of $8M US (that is "several million Euros + the office space to hold it) and a depreciation cycle of 12 years that is about $500 / day for the machinery we can add another $500 / day for staff + electricity, figuring 8 hour days, that's $125/hour to operate for 60 chips is a bit more than $2/dice.
Well the pencil math works (with all of those assumptions) but even assuming its off by an order of magnitude, $20/dice isn't a deal breaker for your own custom chip that does your special thing. You'll also notice that the 50 weeks a year 40 hours a week assumption. I'm guessing you can get better utilization than that which would offset your depreciation costs.
Here is a PDF from 2017 from someone involved with a project wich seems to intend to use this in an (large scale?) innovation initiative in Japan.
[1] www.lip6.fr/public/2017-03-17_Shimizu.pdf
Mentioned are workflows, used tools, intendend audience and goals,
estimated prices, nda-freeness of spice models and design rule check, open source, open-cores, github, and so on.
If they really make this widely available, then my mind is blown...
That looks very exciting for some research environments, and for some prototyping. But a 12mm wafer is certainly going to limit what you can do, and they make no mention of feature size. In the end, what you can accomplish is determined by the process and how easily your design tools let you target the process. I am sure there is a lot of academic research that could benefit greatly. As to prototyping, I have concerns that it would be difficult to do a part on this process and transfer it to a different process for scale.
[1] https://www.minimalfab.com/en/ [2] https://www.youtube.com/watch?v=WsOVbmfYxoM