Making RAM at Home [video]
youtube.com
youtube.com
This channel has another video where it shows how the clean room lab is created starting from a basic backyard shed, and that was truly astounding. The positive pressure to keep the number of particles low in someone’s backyard is almost mystical to me.
To put it that into context, some tags count as upvotes, others count as downvotes, "Troll" is a downvote. So to have your post labelled as "Troll" with a positive score, it has to have enough upvotes to compensate the penalty from the "Troll" votes, but without having another tag dominate. 5 is the maximum score.
"Score: 5, Troll" is therefore the mark of a very successful troll.
Given that the shed in this guy’s backyard is already approaching the entire national technological output of any country in the 1970s I think he may get there.
Ok, maybe I'm being a bit cynical. Stories about bikeable cities are welcome too. And wasn't Soylent popular for a hot minute back in the day?
Indistinguishable From Magic: Manufacturing Modern Computer Chips
https://www.youtube.com/watch?v=NGFhc8R_uO4&t=2070s
It's quite old but I think there is no modern version of it.
I've tried posting to HN a few times but it hasn't gained traction for some reason, but I find it absolutely mind blowing.
Maybe it’s just I come here for the old web feel when video was costly, rare and short.
I'm totally with the text folks on the 5 hour Fossdem sessions, though. Give me an accurate transcript I can grep or don't even bother.
However I've discovered wonderful gems like this RAM video.
Asianometry[0] has a number of videos on EUV lithography that cover some of the mind-blowing advances in the years since.
Veritasium[1] recently also made a video on the subject.
[0] https://www.youtube.com/playlist?list=PLKtxx9TnH76RYHY7L1YzE...
Absolutely insane stuff.
why even allow the cognitive overhead to worry about such a thing? it's not for you to decide anyway - let the users decide using the voting system that's built to task
1999. We will have flying cars
2024. LLMs - there will be robots
2026. How to make your own RAM 2028. How to forage for edible plants and identify safe water sources
2030. How to make horseshoes, nails, and plate armor
2032. How to send smoke signals to neighboring villages
2034. How to defend against barbarian incursionsI'd pay less for RAM that wasn't "certified" in some official way, at least it works.
nothing, except the terrible yields that they would obtain, and the lack of scale making the entire enterprise not profit generating (as the amount of profit per sale is too low if it even is positive, but you can't set it higher as there's cheaper, "better" ram available from pre-established fabs that do have economies of scale).
You could play the artisanal angle, and market it as home grown, organic ram. Not sure how much real buyers of ram care, but might get a few hobbyists in the market.
OpenAI should do their own production, I say slightly bitter because I'm in a health care sector that might be affected because we can't scale up or repair our infrastructure due to their massive pre-orders.
The clean room isn't the hard part about being competitive. It's using advanced lithography to cram billions of cells into a single chip. If you want make DDR2 chips on a 90 nanometer process, that is accessible to a whole lot of companies, but nobody will buy the product. And in the micrometer range you can DIY like this guy.
I'm an infrastructure architect and work with health care, local tax agency, they're all getting 25-30% inflated bills now for new hardware.
And what happens when smaller companies have to repair or scale their infrastructure and can't get affordable RAM?
I'd say if people aren't desperate already, they're about to be.
Correct, nobody will buy your 1GB stick of DDR2-speed RAM for the $100 it cost you to produce it.
> And what happens when smaller companies have to repair or scale their infrastructure and can't get affordable RAM?
That situation sucks but bringing up more obsolete fabrication isn't going to help. They can't compete with modern chips even when those modern chips have a 5x price penalty.
Just imagine if electricity costs were trending towards 0. Instead of e-waste run all those machines till the chips burn out.
The real problem with the RAM business is that it was commodified; normally manufacturers make a relatively small margin. No incentive to build a factory for that. These are not normal times because (a) someone has bought all the RAM and (b) someone has blown up a whole load of globally critical infrastructure in the Middle East.
The risk the existing RAM manufacturers are being cautious about is the risk of normal: if you start building a factory now, will you be selling into a RAM glut?
i mean, you'd think if someone is willing to through 60B (or 10B for an option to buy at a 60B valuation) at Cursor, then other people would give it a shot at starting a SotA semiconductor fab, but apparently the profit expectations are not there
https://en.wikipedia.org/wiki/Extreme_ultraviolet_lithograph...
It would be difficult to break into the RAM business with that sort of product as most of the demand these days is for higher capacities.
1. How is the value read? Is it reading the leak?
2. How is recharging done? I guess the leak itself (assuming my guess in 1. holds) could provide charge for some logic that checks "if has charge then recharge else nop".
I still don't really get transistors :P, but this was cool.
It turns out they intentionally drain a bit of the storage capacitor, and amplify that weak signal. Some of that amplified charge is then fed back to storage.
And you are right, that charge on the drain can then be used both to drive some logic and to activate the recharging of the capacitor that was just discharged.
By the way I am being handwavy about "charge" moving about, if you really want to learn the electronics it is more correct to call it a voltage relative to some ground that the charge always moves towards.
Measuring the charge also removes some of the charge -- fast, compared to the leak spead.
DRAM chips have a circuit that writes the value back -- charges the capacitor up if there's supposed to be a charge, drains it if there's not supposed to be a charge.
Refreshes and normal reads are the same, except that normal reads sends the value(s) to the output pin(s) on the chip.
He has "only" shown the basic grid of capacitors and transistors. The chip he shown has no circuit to measure charges or to write them back afterwards. This makes it easier to test the basic grip and the basic capcacitors.
Pretty sure the proper read out and write back circuit comes in the next video.
The capacitance of the wires themselves is typically an order of magnitude greater than that of the storage capacitors. So when the memory is read, the wires are first precharged to some standard voltage. Then the desired row of storage capacitors is connected to the wires, and the charges from the storage capacitors spread onto the wires, changing their voltages very slightly. These voltage deviations from the standard value are amplified by the "sense amplifiers". The amplifiers are sort of like flip-flops. Once they start in a state which "tilts" slightly to "zero" or slightly to "one", they go all the way to the full magnitude zeros and ones. This not only amplifies the signal, but also automatically brings the voltages on the wires and the still connected to them capacitors to the full magnitude, thus "refreshing" the data. The row is disconnected, and the next read cycle can start for some other row.
In the video, an array of 4x5 capacitors and the associated with them switches was fabricated. The capacitors in the video are several hundred times larger (12400 fF) than typical capacitors in a 64 Kbit DRAM (about 50 fF). I assume this is done so that in the later episodes the author could implement the readout electronics outside of the chip.
The important bits here are:
- any two conductive plates close together with an insulator between make a capacitor (1)
- when a capacitor is charged, the energy goes into the electric FIELD
- that's the FIELD of the field effect transistor
- if the field is strong enough, that causes conduction between source and drain (along the "channel")
- the insulator is nanometers thick, so current leaks across it; at that scale, you can detect individual electrons quantum teleporting straight through it.
(1) technically like gravity there is a capacitance between any two objects in the universe, but it's only significant when you have relatively wide and close conductors
tldr it's not theorical, some madman actually did that!
That said, I could actually kind of see it being useful for stuff that doesn't need to be super fast but does allocate a lot of RAM (though I'm drawing a blank on what that would be at the moment).
... and sufficiently interesting to more than a very small group of persons.
No one is expecting them to make a fortune over night, but it at least shows there is space for new people. This video was actually push into my feed by youtube which is rare for new creators. So they clearly are doing pretty well in the grand scheme of things.
The comment of the video also says "Thanks for the support!" whereas I (maybe you are a subscriber to their Patreon but I know I'm not) have not contributed anything.
Does this look normal to have locked videos behind a paywall then have it public and free elsewhere?
PS: I obviously didn't expect them to make this amount (even though it would be nice) only trying to highlight that there is a mismatch and thus trying to understand what the problem is, namely is that video actually meant to be public.
Yes, this is very normal on youtube with both patreon and nebula.
Can you please give me an example of locked content that is openly shared outside? I guess I don't understand the monetization in that situation so I'd appreciate if you could help me get what they are doing here.
This guy is not exactly a regular person. He is a pretty unique case of a talented semiconductor engineer who has a home lab for side hassles. It is not a low effort thing. He runs the equipment 24/7, scrubs all the surfaces in the lab daily to keep it clean.
Still, with the lab and all of the equipment already at hand, it cost him several weeks of work to produce this demonstration of transistors and capacitors, which kind of work, but are still long ways from a "completely complete" 20 bit DRAM chip.
Unfortunately it is simply too much work for one person to maintain a viable semiconductor fabrication process, even when it is done semi-professionally.
Much more impressive are the modifications to the microscope, transforming it into an improvised lithography machine, and the home made plasma etching machine, cobbled together from surplus components.
Of course, the whole thing, starting from the clean room, is extremely impressive -- Intel started their business in a much simpler facility.
Would be interesting to know where the biggest bottleneck shows up first: materials, lithography, or error rates?
RAM at home:
Those AI companies and hardware manufacturers lost all right to further dictate and increase prices. Capitalism does not work as de-facto blackmail monopoly - or should not. If a state fails to protect the people, such as in the USA right now under the orange king, then the people need to insist on change. ALL steps against this tyranny from a few superrich needs to end.
Right now the legislation is going in the way how lobbyists want this, e. g. trying to make 3D printing illegal, but I think technology will obsolete such illegal laws eventually. Tyranny will eventually fail.
I get my DRAM needs at the RAM ranch.
18GB at a time
It's notable in comparison that all the deadly organics lived together in an unlocked cupboard (vented OFC). I think the only thing I ever saw treated as more of a pariah than HF were radioactive isotopes. Those generally get an entire dedicated room with restricted entry and a tedious mandatory cleaning procedure.
[1] https://www.science.org/content/blog-post/things-i-won-t-tou...
Now, perhaps this chemist meant that no chemist in their right mind would physically touch HF--in that case, I agree completely!
It sure is, if you ignore the fact that you have to have a factory to make it where one of the *nicest* things around is the fucking hydrofluoric acid, and most of the rest will kill you instantly in trace amounts.
The propane is going to get burnt anyway. May as well extract some useful work from it, and when you run a car off it they become ultra low emission.
But in practical terms the Z80 was a 4 um process node so unless you're willing to go back to the proverbial stone age it seems like you need semiconductors and lithography.
The DDR market will adapt, as China grey market state fab smells the opportunity. They have been counterfeiting cmos chips for decades already, and dram is not as complex as people like to assume.
Neuromorphic computing will likely kick over the LLM sand pile at some point, and all that discounted hardware will need re-homed. We can wait for the bubble to run its course, and actual investors realize they were conned. =3
I don't see a lot of work going on in neuromorphic - there was some work at Intel, IIRC. Not saying you're wrong, but just wondering where you think it's going to come from?
Will likely evolve like any regular biological system, and consume translated LLM weight sets into its initial training condition 3D propagation structure.
The speed at which this occurs will likely initially be measured in weeks due to slower growth state writes, but once bootstrapped the GC is self-propagating.
I normally don't like to speculate, but the barrier to entry would actually be much lower than traditional silicon fabrication processes. It was an old idea from science fiction, that until recently was highly impractical. Asimov was likely wrong about the physical process, but not about how it is made.
https://en.wikipedia.org/wiki/Positronic_brain
Due to theoretical constant morphological changes under GC, one must acknowledge the inherent lack of safety such systems would pose. Have a nice day. =3
On the academic side of things there's a steady drip of papers on things like spiking neutral networks so I'd say the general theme is being explored.
I figure if a breakthrough happens it will be overnight just like what happened with transformers.