| HN Mirror

I don't think that's correct. They didn't have the knowledge to produce a typical ARM CPU, or, for that matter, any CPU; they didn't know what computers were, or why they were important, nor did they have the quantum theory or materials science necessary to fabricate useful silicon chips. Probably a collapse would lose the knowledge locked up inside TSMC, Samsung, and Intel. But we'd still know about zone refining, ion implantation, self-aligning gates, hafnia high-κ dielectrics, RISC, superscalar processors, cache hierarchies, etc.

If we forget about typical ARM CPUs for the moment, and just look at ARM CPUs in general, the ARM 2 was supposedly 27000 transistors according to https://en.wikipedia.org/wiki/Transistor_count. If you had to hand-solder 27000 SOT23 transistors onto PCBs at 10 seconds per transistor, it would take you a couple of weeks of full-time work to build one by hand, and probably another week or two to find and fix the assembly errors. It would be maybe a square meter or two of PCBs. At today's labor prices such a CPU would cost on the order of US$5000. At today's 1.3¢ per MOSFET (LBSS84LT1G and 2N7002 from JLCPCB's basic parts list a few years ago), we're talking about US$400 of transistors.

(Incidentally, Chuck Moore's MuP21 chip was 9000 transistors, so we know how to make an acceptably convenient-to-program chip in a lot less space than the ARM. It just does less computation per cycle. A big chunk of the ARM 2 was the multiplier, which Moore left out.)

It probably wouldn't run at more than 5 million instructions per second (maybe 2 VAX MIPS, slower than a 386), and because it's built out of discrete power transistors, it would use a lot more power than the original ARM. But it would run ARM code, and the supply chain only needs to provide two types of MOSFET, PCBs, and solder.

US$5400 for a 2-VAX-MIPS CPU is not a competitive price for computing power in today's world, and if you want to drive the cost down, you need to automate, specialize, and probably diversify the supply chain. If you were building it out of 74HC00-class chips, for example, you'd probably need a dozen or so SKUs, but each chip would be equivalent to about 20 transistors, so you'd only need about 1400 chips, currently costing about 10¢ each, cutting your parts price to about US$140 and your assembly time to probably a day or two of work, so maybe US$500 including debugging. And your clock rates would be higher and power usage lower, because a gate input built out of 2N7002 and similar power MOSFETs will have a gate capacitance around 60pF, while a 74HC08 is more like 6pF. We're down to US$640, which is still far from economically competitive but sure looks a lot better.

The 74HC08 and family are CMOS clones of the SN7400 series launched by Texas Instruments in 01966, at a time when most of the world electronics supply chain (both providing their materials and buying the chips to put into products) was inside the US. It couldn't have happened in Cameroon or Paraguay for a variety of reasons, one of which was that they weren't sufficiently prosperous due to a lack of international trade. But that's somewhat incidental—what matters for the feasibility is that the supply chain had the money it needed, not where that money came from. Unlike the SR-71 project, they didn't have to import titanium from Russia; unlike the US ten years later, they didn't have to import energy from Saudi Arabia.

In his garage, using surplus machinery and wafers from the existing semiconductor supply chain, Sam Zeloof has reached what he says is the equivalent of Intel's 10μm process from 01971 http://sam.zeloof.xyz/category/semiconductor/ in his garage.

On this basis, it seems to me that making something like the 74HC08 from raw materials is something that a dozen or so people could manage, as long as they had existing equipment. It wouldn't even require a whole city, much less a worldwide supply chain.

So why don't we see this? Why is it not happening if it's possible? Well, we're still talking about building something with 80386-like performance for US$700 or so. This isn't currently a profitable product, because LCSC will sell you a WCH RISC-V microcontroller that's several times that fast for 14¢ in quantity 500 (specifically https://www.lcsc.com/product-detail/Microcontrollers-MCU-MPU...), and it includes RAM, Flash, and several peripherals.

If you want to build something like the actual ARM2 chip from 01986, you'll need to increase transistor density by another factor of 25 over what Zeloof has done and get to a 2μm process, slightly better than the process used for the 8086 and 68000: https://en.wikipedia.org/wiki/List_of_semiconductor_scale_ex...

Now, as it happens, typical ARM CPUs today are 48MHz, and Dennard scaling gets you to 25–50MHz at around 800nm, like the Intel 80486 from 01989. So to make a typical ARM CPU, you don't have to catch up to TSMC's 6nm process. You can get by with an 800nm process. So you might need the work of hundreds or even thousands of people to be able to make something like a typical ARM CPU, and it would probably take them a year or two of full-time work. This works out to an NRE cost on the order of US$50 million. Recouping that NRE cost at 14¢ per chip, assuming a 7¢ cost of goods sold, would require you to sell 700 million chips. And, using an antiquated process like that, you aren't going to be able to match WCH's power consumption numbers, so you probably aren't going to be able to dominate the market to such an extent, especially if you're paying more for your raw materials and machinery.

So it's possible, but it's unprofitable, because the worldwide supply chain can make a better product for a lower cost than this hypothetical Silicon River Rouge plant.

Make no mistake, though: if the worldwide supply chain were to vanish, typical ARM CPUs would be back in less than a decade. We're currently watching the PRC play through this in real time with SMIC. The USA kneecapped 天河-2, the top supercomputer on the TOP500 list, in 02015, and has been fervently attempting to cut off the PRC's semiconductor industry from the world supply chain ever since, on the theory that the US government should have jurisdiction over which companies inside the PRC are allowed to sell to the PRC's military forces. They haven't quite caught up, but with the HiSilicon CPUs used in Huawei's Mate60 cellphones, they've reached 7nm: https://www.youtube.com/watch?v=08myo1UdTZ8