Chipforge opensource foundry [pdf] | Hacker News Reader

Chipforge opensource foundry [pdf]

github.com

75 pointsbaybal220 comments

It's awesome people are working on this. Their progress is unfortunately is pretty hard to follow or even support.

Their low-tech process steps: https://github.com/libresilicon/process/raw/master/process_s...

Lightning talk during the 34c3: https://media.ccc.de/v/34c3-9256-lightning_talks_day_2#t=178...

For an idea of how you could build transistors in your (admittedly pretty large) basement: https://www.youtube.com/watch?v=TrmqZ0hgAXk

HN thread with article about a home-made litographically-fabricated IC: https://news.ycombinator.com/item?id=16955549

I wish I could be pleased. Summary:

- they are offering a 1 um process

- process specs contains very little information and a scary number of "(??)"

- test wafer specs are at https://github.com/chipforge/PearlRiver

- classically they have built (or will build?) a ring oscillator

- no results published for test wafer measurements

- again NO test data published

- fab is a space at Hong Kong University of Technology

- university fabs are often run on terrible equipment and by (to put it generously) trainees

- every time the university trainee operator changes you're probably qualifying a new process on your design (yes this is a bit snarky)

Based on my experiences getting stuff out of a university fab "process"... I'd rather sign the lot of NDAs and pay the money. Wait time might be faster out of a production fab, and you can usually bank on the process being followed.

In their defence, the sole fact that they actually managed to get started is already a miracle.

Whatever they present now, is not even "stage 1" of what it is to become once they will get any much substantial cash flow.

Their long term goal is to get to 180nm, or with a stretch to 130nm.

How they will get there without an actual fab, you may ask? The amazing thing is the fabs are actually there: HK has a number of abandoned fabs built up to nineties technology level. They were once built by no name electronics companies in nineties, but they all went bust during the Asian crisis. And HK never managed to emerge as a player in semiconductors.

HK government is facing a dilemma, they have to do something with them, but bulldozing a FAB, even a nineties era one, is not an option. And here the Chipforge comes.

I don't have any knowledge, but wouldn't running an ancient full-scale commercial fab academically/for small volumes be very uneconomical even if they got the fab itself for free.

Even 600nm is commercially viable for specialty applications, 300nm for low-end general purpose, and their aim is 180nm. So there is nothing about the process itself that disqualifies it.

Heck, with immersion lito you can even use DLP chip(s) with 405nm and then-trivial multi-patterning to the 180nm... likely even close to 130nm, but that's a stretch that might be too hard on the compensation. This sort of node, with no minimum order quantity and pricing within 2~5x of mask-based fab (slightly slower steppers at these scales, but 600nm might be still reachable in 2 minutes on a 12" for the exposure if you use some resonance or so to get sawtooth-motion of the head to e.g. compensate for a spinning wafer (think like a record/CD player unless viewed in slow motion). Resonance is surprisingly easy to keep steady-state with just a piezo and maybe another one if needed that provides all high frequency content that's not handled by the fundamental and 2 or 3 low harmonics of the scan frequency.

There are applications thanks to modern automation as far as hard-wired logic instead of a via mask ROM, and this should be able to eat much of the low-end FPGA business due to even just much higher design flexibility.

> HK has a number of abandoned fabs built up to nineties technology level. They were once built by no name electronics companies in nineties, but they all went bust during the Asian crisis.

This is __absolutely__ the most cyberpunk thing I've read all year. Truly, we live in the future.

A question arises - how much work would it actually be to bring up a fab that's been dead for several (much less a dozen) years? Wouldn't the fab basically need to be ripped apart and completely refurbished by now? Or are the most important pieces fairly immune to physical degradation when abandoned without regular ongoing maintenance?

Quite a lot, but still incomparable to doing it from scratch. FAB require very unique equipment, not used in any other industry. Like water filtration plants that can make ultra pure deionised water in tenths of cubic metres per hour, process gas recycling plants, industry specific automation equipment, etc

Even a nineties tech level, a FAB still costs a megaton of money.

The most unique proposition here is that a prospect of getting an actual FAB with a lot of automation, and not a laboratory coaxed into manufacturing role is within the grasp. You can't have that anywhere else in the world.

Going through a company like MOSIS is the best bet for university research. You get access to modern processes and reduced cost by sharing a wafer. Running your own out of date fab is dumb and expensive. One of the cheapest processes they have is ON Semi 0.5um. That is the one my school offered for free to students taking a CMOS layout class for undergrads. After the final project was turned in, the professor would off anyone with a working design the chance to get their layout made. You'd have to write a report on how well the run went (for MOSIS, the school didn't care since this was well after the semester was over).

Yep, you are right. But the point of this enterprise is not RnD or anything, and "doing the dumb, expensive thing" with hope of making it being less so, is the very raison d'etre for the Chipforge

Hi all. I'm leviathan, the guy from the lightening talk and the talk here: https://youtu.be/kiTcaRqzOmY

In fact you really can get factory buildings basically for free from Hong Kong government because they don't know what to do with the building, but there is also a fab which actually is still in use (not heavily though) We're in talks with some of the folks who know the owner of the fab and are having a meeting with him very soon.

They actually have automation of the production line to a certain degree, so we can most likely stem it if we can get two more members into the team.

For now we are just planning the layout and mask set and making sure everything is correct, so that we can provide reliable measurement data in order for you guys to start designing.

Since the HK dollar stays for ages already on 8HKD ~ 1USD but the Reminbi is getting more expensive, we will at some point be cheaper producing here in Hong Kong.

You can however just sign NDAs and shove the money up into the ... wallets ... of big companies like SMIC.

If however you wanna be sure that there are no CIA backdoors sneaked into your mask set I'd suggest our service.

I also accept BitCoins BTW as payment for the chips ;-)

I think the vision of creating an open set of processes and standard cells that every foundry could offer is a good one, but they aren't particularly motivated to sign up.

I am not confident it will motivate the foundry that this will get any additional customers or will (as a process) be able to fill extra capacity.

That said I tried briefly to get some folks together to buy the old Atmel Fab building as a working fab post Microchip acquisition (as I recall it was a 1u fab) but people were talking 10's of millions which wasn't what I was thinking especially since I believe the plan of record is (or was) to just scrap it anyway.

Wow, any background info on that?

On what? The increasingly short supply of fabs in the Bay Area? Apple bought Maxim's 6" line, Microchip closed Micrel's old 6" line. Intel pretty much shut down everything they had in the early 2000's (there are data centers now on the land previously used for the fabs) I am pretty sure that Zilog used to have an R&D Fab over on Bubb road (not sure what happened to it).

I started on the quest because I like to wander around the various auction liquidation shops in the Bay Area which often clean out a building and just sell all the gear they pull out for pennies on the dollar (the recent round of biotech closings lead to lots of refrigerators and incubators and microscopes and such). As I saw 4" and 6" wafer equipment go through these places I asked "What would it take to 'rescue' one of these facilities intact?" (basically a clean room with a full set of machines inside). And made inquiries.

Occasionally I would know someone on the inside who can route my query to the right person but if not one typically ends up at either the corporate development group or the corporate assets group. The sad reality is that the land/space is more valuable than a working, low volume, semiconductor line for 'older' processes.

As a result you have to pay the land use cost rather than the disposal cost and it really isn't an economical sort of thing at all (you'll never be able to make back the money you invest even if you sell the chips). And what that means is that even when you have friends that think it would be 'neat' and could fund such an effort, the math just doesn't pencil out.

If you run at the problem the other way and offer to buy all of the gear, and basically reconstruct the line in a new location, then you have to build a clean room, get a bazillion permits for the chemical storage, build a LNOX plant, and pay to upgrade power to support what the fab needs. Again, the math doesn't pencil out.

What that experience taught me was that in the Bay Area at least, you're not going to be able to purchase/save a fab that the owner no longer wants for any sort of a reasonable price. This is especially true if all you intend to do with it are low volume and open source chips in it.

Did you consider using modern e.g. DLP chips or so for a maskless fab? It would not work on too modern nodes anyway, just due to the restriction to ~405 nm by the aluminium reflectors on the chip turning i-Line and deeper into more and more heat you can't get away from the chip. 180 nm seems to be what one could hope for with an immersion stepper at 405 nm without too much trickery, but the next one up should be possible with dry exposure. Hit me up if you want help or so with a maskless, sane-priced-chips yielding fab. There are just so many applications once you have the same price per chip at 1 or 5 or 100. Things like custom hardwired monolithic analogue compute could enable e.g. micro drones and likely lots of other stuff if it's cheaper to use some tool to generate the analogue or mixed circuit that will replace all logic parts on some circuit board, leaving only power handling components and those that don't work inside an IC. Maybe the freedom to create a monolithic version of an HB100 radar module, allowing a heterodyne design due to the savings from integration, with the needed experimentation being possible just because iterations are neither expensive nor long (comparable to Chinese PCB manufacture, so something under a week).

This is a good thought. The short answer is of course no, I didn't because I was looking at preserving an existing line rather than build a new one. However I think there is a lot of merit in penciling out a closed cell[1] (mostly) dry line. The cost to develop the equipment would be pretty high though, perhaps prohibitive relative to the ability to resell that gear.

[1] Where wafers are always in a closed container, no clean room required.

Oh, I very much meant using the chemical technology, and just replacing the steppers in one of the lines with the maskless ones.

For the closed cell line, which problems are you seeing? Any special use cases for making chips on a process that emphasizes uniquness so much? I mean, apart from maybe being able to do AOI after the front of line processing to hard-route metal around (presumed) defects, in order to reduce the cleanliness requirements for actual monolithic wafers. I mean like if you have a 2D, routing mesh of compute, and you need to worry about routers having defects that tank your yield in wafer-scale chips, you can design the front of line masks to allow wiring one or maybe even more clock delays into the mesh link (without significant cost to the attainable frequency/minimum delay) and also just not connect the link from one or more neighbors over (on the other side of the defect router) to the defect circuity, so that routing between the nodes around the defect one isn't impacted, and neither are any delay cycles. The one node is broken, but it'd just be something a compiler would have to work around.

Other things like hardwired crypto keys might be of use, or hardwired logic for these <15 ppi very, very wide aspect ratio LED screens that are anyway configured for some fixed text to scroll by. Or some analogue IC that is detuned from theoretical values to compensate for the not-precisely-manufacturable discrete capacitors (and inductors), as e.g. the needed Q-factor/voltage of the resonance doesn't allow mechanical trimming components to be used, as they'd induce corona discharge if close enough to not murder the Q-factor.

Digital control is in nowadays even for fast feedback loops, but you can't beat the latency of analogue PD controllers. Good luck compensating single-digit MHz vibration actively with a piezo and e.g. optical interference pattern contrast sensing with some photodiodes like done with 4 of the latter for a CD drive. I'm confident that the power driver is a harder part, given that you want it to follow lumped-element behavior and not distributed-element effects, so you can cope with random noise of sufficiently low frequency to fall within the bandpass range of the control circuity.

If DIY chip fabrication takes off even a little, I hope they use safer chemicals than the industry did back on those nodes. Safety and simplicity should be top priorities IMHO which would also help such a thing proliferate.

I like the idea of having a local fab for my own designs. Even if they're not that complex.

Are there currently safer chemicals available for semiconductor fabrication?

The dangerous part is stripping the oxide layer off the wafer. That's currently done with fluorides, which are really nasty to work with.

Their website

http://libresilicon.com/

Reply on news.ycombinator.com