[HN Gopher] A 32-bit processor made with an atomically thin semi...
___________________________________________________________________
A 32-bit processor made with an atomically thin semiconductor
Author : PaulHoule
Score : 215 points
Date : 2025-04-08 13:08 UTC (4 days ago)
(HTM) web link (arstechnica.com)
(TXT) w3m dump (arstechnica.com)
| amelius wrote:
| I'm still waiting for that inkjet printer that can print
| transistors.
|
| https://www.nature.com/articles/s41598-017-01391-2
| godelski wrote:
| Has anyone tried to replicate this? Seems like it would be very
| useful for amateur makers/hackers were it not for the $23k
| printer cost (no idea for the cost of the discussed silver
| ink). But surely someone crazy had access to one and tried or
| has tried to replicate on a cheaper printer? I figure HN has a
| decent chance of helping find said persons?
| superb_dev wrote:
| I don't think they've tried it yet, but it's seems up the
| alley of Applied Science on YouTube
| godelski wrote:
| I'm not sure this is Ben's forte, but you're right that I
| wouldn't be surprised if he tries it, though he has done
| some circuit stuff[0,1] so nothing would surprise me from
| him. (Hi Ben! Love the work!) BUT I do think this is
| something Sam Zeloof[2] try. He's done some lithography
| using a projector[3]. Also there's Jeri Ellsworth, but I
| think she's shifted to mostly working with her AR project.
| Tons of old videos on that stuff if you're into it.
|
| Side note: I'm assuming anyone who knows any of these
| people would be interested that a new Dan Gelbart video
| just dropped[5]!
| -----------------------------------------
|
| Other side note: @YouTube people (and @GoogleSearch), can
| we talk about search? The updates have been progressively
| making it harder to find these types of accounts. People
| who do * _highly*_ technical things. I get that these are
| aimed at extremely specific audiences but this type of
| information is some of the most valuable information on the
| planet. Lots of knowledge is locked into people 's heads
| and these classes of videos are one of the biggest booms to
| knowledge distribution we've ever seen in the course of
| humanity. I understand that this does not directly lead to
| profits to YouTube (certainly it _DOES_ for Google Search),
| but indirectly it does (keeps these users on your
| platform!) and has a high benefit to humanity in general.
| The beauty of YouTube and Google was you could access
| anything. That we recognized everyone was different and we
| could find like minded people in a vast sea. The problem
| search was meant to solve was to get us access to hard to
| find things. To find needles in ever growing haystacks!
| Please, I really do not want to return to the days of pre-
| search. Nor even early search! It should be _easier_ to
| find niche topics these days, not harder. LLMs aren 't
| going to fix this. This is becoming an existential crisis
| and it needs to be resolved.
|
| [0] https://www.youtube.com/watch?v=UIqhpxul_og
|
| [1] https://www.youtube.com/watch?v=FYgIuc-VqHE
|
| [2] https://www.youtube.com/@SamZeloof
|
| [3] https://www.youtube.com/watch?v=XVoldtNpIzI
|
| [4] https://www.youtube.com/@JeriEllsworthJabber
|
| [5] https://www.youtube.com/watch?v=OuZjjActWmQ
| PaulHoule wrote:
| LLMs could help if they were specifically applied to the
| task [1], however people are actually applying them to
| the generation of countless slop videos. Google's
| problem, which I think there is no cure for, is that
| Google believes it is #1 and to quote Fatboy Slim "We're
| #1 why try harder?" If in some way they feel they have
| competition it is to be a 2nd rate TikTok, not be a
| better version of what made YouTube great.
|
| In the meantime, for everybody that's been turned on to
| something really awesome and creative on YouTube somebody
| else got turned on to something really toxic.
|
| [1] Something significant happens every 10 years in
| search relevance, and SBERT was one of those.
| ezst wrote:
| Again someone mistaking LLMs with knowledge bases. Must
| be a day finishing in `y`
| PaulHoule wrote:
| The original misunderstanding behind "knowledge base" was
| that, in the 1980s, it was an idea in symbolic AI that
| you'd develop a set of facts against an ontology designed
| for accurate inference and somehow by the 1990s it became
| a text repository with a search engine that may or may
| not work. Occasionally useful, sometimes hard to
| distinguish from a trash can. See _Confluence._
|
| Prompt engineers with their decoder models are going to
| always be wondering why they are always a bridesmaid and
| never a bride, with encoder models you can attain the
| holy grail of the system where you put text in one side
| and get, within calibrated accuracy, facts to put into
| the first kind of knowledge base. Or, for that matter, a
| good search engine for the second kind of knowledge base
| which could raise it above the "trash can" level.
| ezst wrote:
| "Funny" how that reminisces of the whole blockchain
| discussion. If the need is fully satisfied by a "boring"
| and cost-effective "facts" database, why would an
| adequate engineer push for (blockchain/)LLM instead?
| PaulHoule wrote:
| There were several reasons why "expert system" were
| rejected in the 1980s including competition with
| programmable calculators and spreadsheets and no correct
| paradigm for reasoning with uncertainty but the one most
| quoted was that the creation of that kind of database is
| not cost-effective.
|
| I spent about 10 years working (sometimes for myself,
| sometimes for employers, sometimes part time, sometimes
| as a software developer sometimes as a business
| developer) on the problem of turning a mass of text into
| facts into text to solve problems like:
|
| - Doctors write copious medical notes from which facts
| would be useful for themselves, payers, researchers,
| regulators.
|
| - An accounting or legal firm may need to scan vast
| numbers of documents and extract facts for a audit or
| lawsuit
|
| - An aerospace manufacturer has a vast database of
| documentation and maintenance notes (even from the teams
| at the airports) that it needs to keep on top of
|
| - A fashion retailer wants to keep track of social media
| chatter to understand how it connects and fails to
| connect with customers and answer questions like "should
| we endorse sports star A or B?"
|
| - Police and soldiers chat with each other over XMPP chat
| about encounters with "the other" which again are rich
| with entities, attributes, events, etc.
|
| Tasks like this need an interactive system but you face
| the problem that people have an upper limit of 2000 or so
| simple decisions [1] in a sustainable day. The problem is
| large but it is not "boil the ocean" because you can set
| requirements for what gets extracted and use the
| techniques of statistical quality control as in Deming to
| know accuracy is in bounds.
|
| You can give people tools to tag things in bulk, you can
| apply rules, you can give the people tools to create the
| rules. I worked on RNN and CNN based models, SVM,
| logistic, autoencoder and other models and before BERT
| they all sucked. If you have the interactive framework
| you can put encoder or decoder LLMs in and it is a
| revolution that makes systems like that much cheaper to
| develop and run for better effects.
|
| [1] hot dog/not hot dog
| godelski wrote:
| I'm highly skeptical. The current ML paradigm is highly
| reliant on aggregating data, but the issue we're
| discussing is about distinguishing subtle details over an
| extremely large search space. Sure, you can probably
| scale your way there but even accounting for
| superposition we're talking about an extremely large
| number of parameters because you aren't performing
| search, you're performing compression. You need to also
| remember the curse of dimensionality. The problem is that
| as the dimension increases the ability to distinguish the
| nearest neighbor from the further neighbor decreases.
| Effectively the notion of distance becomes undefined.
| (The dimensionality increases as parameters increase). So
| now you have to perform search over your compression.
|
| This is why ML is so fucking cool but it's also why they
| are really bad at details. Why you have to really wrestle
| with them to handle nuance. Easiest to see in image
| generators but they're much smaller. Do remember that
| these things are specific trained so that their outputs
| are preferential to humans. The result is that errors are
| in the direction of being difficult to be detected by
| human evaluators. Deciding if that's a feature or bug
| requires careful consideration.
|
| This is not to say that LLMs and ML is useless or trash.
| They are impressive and powerful machines but neither are
| they magic and the answer to everything. We got to
| understand the limitations if we're to move forward. I
| mean that's the job of us here as researchers, engineers,
| and developers. Using a keen eye to find limits and then
| solve them (easier said than done lol)
| philipkglass wrote:
| It's possible that the inkjet printed transistor is both
| replicable _and_ impractical for building a full
| microprocessor.
|
| The inkjet transistor article says "A total of 216 devices
| were tested with a yield of greater than 95%, thus
| demonstrating the true scalability of the process for
| achieving integrated systems." But 95% yield on the
| transistor level implies vanishingly low yield at the device
| level when you need thousands of transistors to build a full
| microprocessor.
|
| Even the new MoS2 microprocessor discussed in the Ars article
| wasn't fabricated all at once. It was built up from sub-
| components like shift registers containing fewer transistors,
| then those components were combined to make a full
| microprocessor. See for example "Supplementary Fig. 7 | Yield
| analysis of wafer-level 8-bit registers." in the
| supplementary information:
|
| https://static-
| content.springer.com/esm/art%3A10.1038%2Fs415...
|
| _The yield of 8-bit registers, each consisting of 144
| transistors, can reach 71% on the wafer._
| godelski wrote:
| My knowledge of transistors is pretty limited[0]. Does the
| yield percentage refer to number of successful chips on a
| substrate or look more at the total number of successful
| transistors? (Or confusing hybrid-term like rain forecasts)
| I believe your comment implies the latter? So the number of
| successful processors is quite low? How many failed
| transistors can you have in a working microprocessor?
| (Probably not an easy to answer question?)
|
| [0] Am I remembering correctly that this is your area?
| Out_of_Characte wrote:
| Yield would be the amount of functioning chips. This may
| be chips, entire packages or even a complexer answere
| where good chips also need to be below certain leakage.
| Cores and caches could be disabled and the list of
| potential yield increasing tooling is always increasing
| when wafers nowadays costs thousands of dollars.
| notjoemama wrote:
| I'll add chip designers add redundancy. When errors are
| detected in testing, they can disable sections of the
| chip by lasering fuses. That allows routing the circuit
| through a higher quality area. Quality is measured by not
| only that the circuit produces correct data but also
| within tolerances of timing and voltage. IIRC RAM is
| approximately 10% redundant. A good quality chip will use
| what meets the spec and leave good transistors unused. A
| poor quality chip will disable bad ones and only use the
| ones that meet spec.
| exe34 wrote:
| if you could print transistors, you could make computers
| the way Wozniak made them - a bunch of chips with a ton of
| wiring.
| chongli wrote:
| You can do that easily and cheaply today without a fancy
| transistor printer.
|
| You can find Apple II schematics easily enough online.
| All the chips are common, off-the-shelf parts still
| available today. You can send the KiCAD drawings (also
| available) to a company like PCBWay and have PCBs made
| very cheaply and in small quantity. Then all you have to
| do is solder in the chips and other components and
| connect the board to a power supply.
| exe34 wrote:
| I think the appeal is the you can print out a couple of
| pages of chips and wire them up, not send out for chips
| and PCBs.
| chongli wrote:
| You can order a whole batch of chips and wire them up on
| breadboards without sending away to have a PCB made. The
| PCB step is the last one when you want to finalize your
| computer and package it up.
|
| Ben Eater actually has a free course on YouTube [1] all
| about building a breadboard computer!
|
| [1] https://youtube.com/playlist?list=PLowKtXNTBypGqImE40
| 5J2565d...
| fc417fc802 wrote:
| There's just a different emotional sense between
| manufacturing the lego bricks yourself versus mail
| ordering the magic blocks that you can assemble into a
| finished product.
| chongli wrote:
| Lego bricks is an apt analogy. I don't know how many
| people would actually care to manufacture their own Lego
| bricks but millions of people enjoy putting Lego
| together.
|
| Sam Zeloof [1] actually went through the exercise of
| making his own semiconductors from scratch. It's a lot of
| chemistry and experimentation and quite interesting as an
| exercise, but not at all practical for building your own
| computer.
|
| Printable transistors would take away the nasty chemistry
| bits that Sam had to deal with but otherwise wouldn't
| help much with making practical devices. Computers have a
| lot of very standard, "Lego brick" or jellybean
| components. Stuff like muxes/demuxes, shift registers,
| adders, and the like. These are the components you can
| buy off the shelf to build your own computer. Building
| these yourself on giant sheets of paper with a printer
| might be interesting but you'd get a far less practical,
| usable computer out of the deal.
|
| [1] https://www.youtube.com/@SamZeloof/videos
| rgzz wrote:
| I don't know much about this topic but you still need the
| magic bricks in the printer to make the magic bricks
| though, no? I guess this can be either depressing or
| relieving but I'm in the former category, I wish you
| could do this stuff from sand or something, without
| relying on modern technology, would be fun.
| fc417fc802 wrote:
| In my mind the printer counts as a tool so it's a
| different category. Also you could always do the same by
| hand with a mask. The feature size might be a bit larger
| though.
|
| As to doing it all "from sand". You can! At least sort
| of. It's always a question of how far down the stack you
| want to take it. After all you probably need to source
| rare earths from somewhere that isn't your backyard.
|
| Check out pictures of the old processes before automated
| VLSI. It was all done by hand including crystalizing the
| silicon. You'll need a clean room and a bunch of weird
| supplies though.
| doublepg23 wrote:
| You can even make a Mac SE/30 "from scratch" - it's mind
| blowing how many PCBs and chips people have made for
| retro computing.
| https://youtu.be/zc3sPoqOFG8?si=iIamSEB00mnxfQdL
| chongli wrote:
| Wow, thank you for this! At some point I really want to
| get my own Mac SE/30. I have a Mac Classic (inherited
| from my uncle) I still need to work on. This video is
| really exciting for anyone who wants to fix one of these
| vintage machines but ends up with a motherboard PCB
| that's been severely damaged by battery leakage.
| bombela wrote:
| I am building one. Right after I find out where to buy liquid
| semi-conductor paste.
| lsllc wrote:
| If it's anything like regular InkJet printers, the liquid
| semiconductor would paste be all dried up every time you went
| to use it, or the capacitor cartridge would run out long before
| the resistor one did!
|
| /s
| ohazi wrote:
| I suspected that this was the case when they mentioned adding
| "one bit at a time" -- the CPU design that they implemented is
| Olof Kindgren's SERV [0], a tiny bit-serial risc-v CPU/soc
| (award-winning, of course).
|
| From [1]:
|
| > Olof Kindgren
|
| > 5th April 2025 at 10:59 am
|
| > It's a great achievement, but I'm of course a little sad to see
| that it's not mentioned anywhere that Wuji is just a renaming of
| my CPU, SERV. They even pasted in block diagrams from my
| documentation.
|
| [0] https://github.com/olofk/serv
|
| [1] https://www.electronicsweekly.com/news/business/2d-32-bit-
| ri...
| koverstreet wrote:
| That sort of copying without attribution should be considered
| outright misconduct; it certainly would be in academia.
| lambda wrote:
| Huh? This is a paper published in Nature, and it does cite
| Olof Kindgren and SERV in the references:
| https://www.nature.com/articles/s41586-025-08759-9#Bib1
|
| The paper itself is behind a paywall so I can't see it, but
| it looks from the references like they provided proper
| attribution.
|
| It's unfortunate that some of the articles around it don't
| mention that, but it seems like the main point of this is
| discussing the process for building the transistors, and then
| showing that can be used to build a complete CPU, not the CPU
| design itself which they just used an off-the-shelf open
| source one, which is designed to use a very small number of
| gates.
| reaperman wrote:
| > The paper itself is behind a paywall so I can't see it
|
| https://archive.org/details/s41586-025-08759-9
| lelandbatey wrote:
| Thanks to the Archive.org link, we can see that indeed they
| link directly to the SERV github in reference 38:
| 38. Kindgren, O. et al. SERV - The SErial RISC-V CPU.
| GitHub http:/github.com/olofk/serv (2020).
| chmod775 wrote:
| They do mention SERV in their references (38).
|
| https://www.nature.com/articles/s41586-025-08759-9
|
| Sadly I can't access the full article right now.
| metalman wrote:
| I like where they say "a sheet that is only a bit over a single
| atom thick, due to the angles between its chemical bonds" it's
| funny that material science has achived ultimate precision, but
| it can only be talked about in general terms Is there any exact
| way to describe the thickness of molebdium disulfide sheets?,
| beyond "a bit over one atom thick" clearly they are etching parts
| of the sheet, and somehow attaching leads, but is it done
| strictly in two dimensions, ie: litteral, flat land?
| fc417fc802 wrote:
| > Is there any exact way to describe the thickness of molebdium
| disulfide sheets?
|
| It's the same set of issues that you'll run into if you try to
| precisely quantify the thickness of a sheet of printer paper.
| It really depends on what you mean when you ask the question.
| The geometry of the electron shell, the minimum theoretical
| width once assembled into the theoretically optimal sheet, the
| impact of various imperfections in practice, the potential for
| more than a single layer to exist (and the associated
| averages), and a number of other things that aren't immediately
| coming to mind.
|
| It's an issue of precision on the part of the party asking the
| question. We usually work on scales that are so large that such
| details aren't meaningful (if you can even measure them in the
| first place).
| roywiggins wrote:
| Looks like a monolayer is about a nanometer thick.
|
| https://www.acsmaterial.com/monolayer-molybdenum-disulfide.h...
| neuroelectron wrote:
| Intel and CEA-Leti Collaboration: Intel and the
| French research institute CEA-Leti are jointly developing 2D
| transition-metal dichalcogenides (TMDs), such as molybdenum
| disulfide (MoS2) and tungsten-based materials, for integration
| into 300mm wafers. These materials offer sub-1nm transistor
| channel thickness, making them ideal for extending Moore's Law
| beyond 2030.
|
| [29 June 2023]
| https://compoundsemiconductor.net/article/117047/CEA-Leti_an...
| mikewarot wrote:
| I wonder if Sam Zeloof and Atomic Semi are trying this out? It
| would be an excellent match for their "build in one atom at a
| time" approach.
| ChrisGammell wrote:
| What is this, a MCU for ant(man)?! It needs to be at
| least...three times that thick!
| gcanyon wrote:
| > It's slow and inefficient
|
| Is there any reason to think this won't improve with time? The
| Intel 4004 was "slow and inefficient" too?
| gcanyon wrote:
| Since it's a single-molecule thick, could this potentially be
| stacked thousands, or millions, thick to deliver ridiculous
| capacity? I assume heat dissipation would be a factor, but the
| article doesn't mention it.
___________________________________________________________________
(page generated 2025-04-12 23:01 UTC)