I worked hard to get my list of favorite songs of 2024 to under 3 hours. The list really doesn’t do 2024 any justice, nor does it do my own faves within it much thereof, but it tells a story.
This February I read CHIP WAR by Chris Miller. The book follows the semiconductor industry form the invention of the transistor, to the trade war between the USA and China. The book is mainly about the economics of electronics manufacturing and geopolitics, but has some technological insight nuggets. I highly recommend it.
“Cores Lose Price War to New Chip” – Intel 1103 DRAM (1971) Ad showing the main allure of their DRAM chip: It’s cheaper than core memory.
Aggressive price-cutting of ICs was a deliberate Fairchild strategy to bootstrap a consumer market. This was a deliberate choice, and not an obvious one to make, given the abundance of demand in DoD funding. It definitely wasn’t Bill Shockley‘s choice in his own company.
1969 Commemoration brochure showing Fairchild’s Shiprock (in color, no less), NM facility. Courtesy of the Computer History Museum
Although accepting DoD contracts, Fairchild Semiconductor was keeping defense-related development at an arm’s length. Had they leaned into being a military contractor, development would not be incentivized to happen fast or cheap. Instead, they kept looking for consumer applications and opened factories wherever was cheapest. Fairchild opened a plant in Shiprock, NM in 1969 (employing primarily Navajo Women) for DoD business and a plant in Hong Kong for TV & Radio parts.
In both cases, factories brought rural communities into tech work (still low-wage manual labor). Later in Southeast Asia, this formula repeated, as semiconductor factories seeking lower margins increased subsistence farmer intake, and in turn promoting urbanization.
Comparing the US market (in which Fairchild was the only profitable player in the 60s) to other semiconductor efforts, the obvious takeaway is cost-cutting and a large market enabled faster technological growth.
Japan becoming a consumer electronics empire was an insanely successful US foreign policy move that survived multiple governments and regime changes; Taiwan’s strategy to fill the same role has been its’ biggest growth engine too.
Side note: with Israel’s current state (gestures vaguely at everything going on in 2023), one might forget that it also tried to have the same geopolitical moat in the early 90s with Tower Semiconductor (ex-National Semi, later Jazz Semi) and later Intel’s fabs. It only worked because of a massive Soviet Jew influx; The only hi-tech manufacturing to exist substantially before that was military – a big change came after the 1987 axing of the Lavi jet fighter, which coincided with the Soviet Jew immigration, and flooded the market with skilled engineers. Although there are fabs in Israel, it currently concentrates foundry-model design centers more than a manufacturing center. There are fabs (3 Intel fabs, currently), but Tower-Jazz (2 fabs in Israel) now makes high-precision CMOS sensors and other high-end products more robust against supply chain disruptions.
Morris Chang‘s prediction that supply chain conditions are unique in the East is still undefeated.
The main driver of US wafer industry was practical experience passed by word of mouth:
Know-how was colocated, often undocumented.
Majority of know-how was not theory, but engineering: melting temperatures, photoresist exposure times.
No noncompetes preventing spread of knowledge.
In the 1980s, this knowledge leaked to Japan quickly (legitimately through academic exchange and some notorious industrial espionage cases). Japan’s strategy wasn’t word of mouth: all semiconductor manufacturers were part of a consortium, working together to deepen Japanese dominance in semiconductor manufacturing.
The Semiconductor Industry, as we know it, is a Supply Chain
A robust supply chain has always been a success enabler – not just today, even when things were smaller: it’s hard to imagine that there was a lot less stuff in the world. The USSR could source machinery and components, but hit a wall as soon as its’ process wasn’t good enough: a single supplier won’t improve their yield or cost without real competition. For comparison, when TI needed purer chemicals or sharper lithography masks than the market had, they made them in-house, as a competitive move. In a planned economy, that was harder to do.
“Twenty-seven years have passed and [the semiconductor industry] witnessed a big change in the world, a big geopolitical situation change in the world. Globalization is almost dead and free trade is almost dead. A lot of people still wish they would come back, but I don’t think they will be back.”
The US is obsessed with tracking and analyzing the semiconductor supply chain. Detaching countries that the US views as threats from the supply chain is really difficult – but so is keeping tabs on every chemical process at risk due to shortage. The discourse forces players in the market to address this publicly. TSMC releases some press material (for public opinion management? not sure) and does publish some reports on their critical sources (I have less than zero expertise on this, mind you).
Some dude showing off his Eurorack Extreme Ultraviolet Lithography Tool, one of the core technologies ASML leans on, Wikimedia Commons
A Western[-ized] fab ecosystem may happen now that a full-on semiconductor war is on with China. CHIP WAR says ASML’s supply chain is the real cause: it’s got thousands of ad-hoc parts sourced from single suppliers. If ASML machines (hella expensive, require ASML technician onsite) are blocked from entering countries the US views as rivals, those countries won’t have access to Extreme UV Lithography for a while (it took the US around 2 decades and multiple nations from research to production).
On the other hand: in the US, unit economics will be pricier, meaning fab margins may suffer – which can lead to many side effects. As I’m writing this, Apple is reportedly lowering the expected production quota of Vision Pro due to Micro OLED panel yield – a technology that’s so far been a challenge to produce consistently. The US hasn’t shown any manufacturing edge yet (Noah Smith has a nice piece on this), and constraining supply channels may be a short-term challenge.
The CDMA/UMTS/5G lineage is the same story of Western strategic importance that x86 and ARM are. Foreign SIGINT code running inside western infra isn’t the only issue, efforts to harm channel capacity happen on simple crappy wifi routers. RF and network safety is an unfortunate frontier – it’s wholly unsexy, but unlikely to go away from defense requirements.
After running into issues bringing Alvy Ray Smith into PARC to as an “Artist In Residence”, “Temp” or “Contractor” to work on Superpaint, Alan Kay succeeded in bringing him over as “Furniture”. Alvy Ray Smith was looked down upon by every senior at PARC outside Kay’s group, who didn’t get why color mattered (memory is expensive!). KCET TV featured Superpaint, Xerox heard and wasn’t having it. Smith, who was at PARC as “Furniture”, had his purchase order cancelled.
The Alto 1 was a prototype built in a handful of units to experiment. v2 was an experiment in productization: it was built in the SDS plant in LA, focusing on assembly. There was an Alto 3 as well, which was meant for mass production. It failed to clear bureaucracy. Many times. In both cases, the Alto was better, further reaching, and deliverable faster than anything Xerox had been making anywhere else. Departments scaled quickly with unfitting org structures to deliver mass market products, and they all preferred their own stuff over Alto 3.
(Suffice to say I’ve seen this behavior happen multiple times since)
The first round of Alto 3 was canned for the Xerox 8xx “Word Processor” system, which had a proprietary OS just for word processing. This series later included the Xerox 820, which ran CP/M.
The second round of Alto 3, with a plan to manufacture entirely from commercial components, was canned in favor of the Xerox Star – a scope-crept project which included an OS and hardware, later only the OS with matching hardware.
(Jim Clark later founded SGI, which at the time of the book’s writing, was a multibillion dollar company lol)
Figure 1 from the Dorado Technical ReportFigure 3 from the Dorado technical report: How to make a CPU
The first issue of INTRODUCTION TO VLSI SYSTEMS is still up on Conway’s UMich website. It’s an incredible book with fundamentals.
The Notetaker
Alan Kay tried to do the opposite of the Dorado and made a computer called Notetaker – a PC for schoolchildren (ok king). To put ethernet inside it, they studied Bob Metcalfe’s ethernet board and made it software-based, shrinking from 80 to 24 chips.
The Notetaker, courtesy of the Computer History Museum
Alan Kay was so burned out from the project that he said he was “going to LA to take Organ lessons” and never came back. Unrelated: Among Kay’s contributions, he appears to also have invented the soyjak meme:
Alan Kay with a working Alto, Cookie Monster displayed on it. Source unknown 🙁
Apple
The demo Apple saw at PARC was a Xerox corporate decision that made no sense to anybody in the center. They came back 2 days later for more demo (likely at the behest of Jobs), then another time for the confidential stuff. Apple Lisa engineers read every PARC paper beforehand.
Side note: the book epilogue mentions “today [1999] even Apple’s existence is in doubt”. lol
The Xerox Business Model: Clicks
This is something I only vaguely knew but shocks me to my core even now: Xerox didn’t sell copy machines. It leased them and charged for photocopies. The counter clicked with each page, and salespeople would nickname their commission size “clicks”. HP ink biz model.
Doing Backflips to Avoid Selling Computers
The printers initially sold by Xerox were essentially teletype machines with a larger buffer. This is at a time where Xerox PARC had a working bitmap laser printer (more similar to Xerograhic copiers!), but Xerox wouldn’t put it to market due to internal politics.
The perception at the time among corporates was hardware needed to be part of a software product. This is a result of software being unpatentable (USPTO wouldn’t accept an application on software).
But We Totally Want to Sell Ethernet. Cables.
Xerox wanted to control all Ethernet products, a worse ROI than licensing photocopiers. IBM had token ring networks, patented and deployed by IBM. A really obvious thing to do was to give ethernet to the world and not worry about mfg, but “OEM” was a vague concept at the time.
How Research Departments Work Today
All in all: It took Xerox 12 years from buying a computer company (SDS, 1969) to selling a first personal computer (Star, 1981). Along the way, PARC developed almost every component of a minicomputer, without successfully productizing any. Star was a financial failure.
PARC drew prestige at first by having an ARPA association. ARPA drew prestige because of a charismatic leader who understood academics and sought their eccentricities.
At the time of the book’s writing (1999), Microsoft Research was the closest thing to PARC, what with connections to academia, and in spite of a lot of “practical” research, still a publishing cycle. Several companies have the same impetus today (2023), where I work included.
Sometime in Oculus Research’s expansion (now Facebook Reality Labs…Research? or Meta? idk) Michael Abrash told employees that the intention was to become PARC. I wonder which part though.
I had the honor of speaking at CVPR’s CV4ARVR workshop this fall. Among many topics covered, I explained a bit about Omniverse XR (the project I’ve been working on at NVIDIA), and the challenges of moving VR and AR into the cloud.
I gave a longer version of this talk at Sandia National Labs’ XR conference and SIGGRAPH 2022, to be published soon.
Yes, I have (had) a music blog, but that’s its own thing. I decided to publish my summaries whenever I make them here. Here’s my 2022.
Spotify:
I assembled almost all the releases on Bandcamp as well, so you can go buy them (If you’re here to buy music instead of streaming it, try to do it on Bandcamp Friday):
Mike Acton’s CPPCon 2014 talk – Programming for the data, not for the program. You may have heard the term “Data-Oriented Design” thrown around; That term has also been tarnished a bit by its’ eagerest adopters, so let’s just say “Designing for the data”.
SIMD Programming Guide for the Cell architecture. Beginners: This architecture is not in use. Some of this knowledge is purely arcane, some still relevant.
For VERSIONS 2016, hosted by Killscreen and NEWINC, I gave a workshop about perception and designing for the senses in virtual reality. The results were positive and I got some emails asking me to share this, so I wrote a short post on what it’s about. Here are the slides:
After the presentation, I asked all the participants to exercise designing for VR, without a headset. The purpose is to use body intuition before using a computer, because VR is focused around the body. I’ve derived some of these methods from design methods rumored to be used in Mark Bolas‘ lab at USC (If anyone at the lab is reading this, please confirm!). In order to design what feels good in virtual space, start with shutting down the most obvious inputs.
To make that happen, you need:
Blindfold
Laser pointer
Sharpies
Large drawing pad
One million post-it notes.
Divide the participants into groups of 3. One is the designer, the other two help out. The designer’s role is to point at things and say what they want to see there, the helpers create large post-it notes and stick them to walls, chairs, etc. There are three rules:
The Designer is blindfolded. They only have two chances to lift their blindfold. After the third, the design is final.
All of the designers instructions are given when blindfolded. Lifting the blindfold is only used to looking around. No talking, not pointing, certainly no touching.
No talking. Only the designer can speak, and they should feel completely alone while doing so.
Photo by Killscreen
With this, some hidden features emerge:
The designer is capable of spatial reckoning without audiovisual cues.
Our proprioception and motion planning skills can do a lot of design work, on things from menus to spaces to interactions.
Interaction and experience come before graphics.
Photo by Killscreen
The interesting thing I find with this is how many, many, people have good spatial reckoning skills embedded in them. They use them daily for processing instructions, planning routes to walk faster in the subway, reaching behind things, driving – these are core human skills that map trivially into VR.
Here’s something I did with Surya Mattu for James George’s video art class at ITP. For now, it’s a program that takes feeds from traffic cameras, extracts cars and turns them into sprites. There’s some fun stuff coming up in part 2, so watch this space.
tl;dr: I made Wikipedia:Random even more satisfying. It’s at randomfax.net.
Maybe it’s just the way I select the people I hang around, but it seems to me I know nearly no one who can say they aren’t the target audience for Wikipedia:Random. Personally, I keep my access to it as close as I can. For a while I even had typing facebook.com in the address bar redirect there.
Wikipedia:Random is an amazing phenomenon. While it may turn out rubbish 50% of the time (unverified statistic), it also turns out pure gold, more often than I’d expect. That means, I guess, that the general importance of all the knowledge in the universe only has a partial ordering, and it’s way more dense in the upper echelons than what you think. There is, however, a minute difference.
Wikipedia:Random isn’t Wikipedia.
It’s just not the same thing. People don’t use that as a feature of Wikipedia, they use it recreationally, in a completely different context. It’s a random fact generator, only better informed and better verified.
randomfax.net is a wrapper for Wikipedia:Random. Only it has two things that make me like it better:
It has a big refresh button, with a pigeon on it.
Instead of a topic, say “Medulla oblongata”, it would display the title as “What the fuck is a ‘Medulla oblongata’?”. I’m calling that a feature.
That’s kind of the point, I guess. Making Wikipedia:Random as fun as it’s supposed to be. Oh, and it’s indexed by google, which means that I have a lot of weird traffic. But more on that later.
randomfax.net is currently available in Hebrew. Once I get a strong enough server, I’ll make it available in English as well. If anyone wants to help open it in another language (depending on the language, it requires a bit of NLP, but I can help with that), hit me up we’ll make it happen.
