IRC channel logs
2023-09-26.log
back to list of logs
<oriansj>muurkha: not to mention the considerable amount of manual skill that needs to be developed to produce just a couple MB of Core Memory by hand. <oriansj>Lithography has a huge cost advantage the second humans are not needed to manually make the masks <oriansj>but getting to that point is rather expensive as you'll need to make hundreds to thousands of cells perfectly to produce the masks for the 256bit to 1024bit memory chips (which you'll need to produce a great many before you'll be able to support the software for mask layout generation) <oriansj>and that is assuming the masks don't wear out before then and you'll have to do it several times. <muurkha>if you're doing a small number of chips, and you have computers, it might be advantageous to use a maskless setup <oriansj>assuming you have a projector and proper lenses for doing so <muurkha>people do direct e-beam lithography in labs today; you spin on the resist, pop the chip in your electron microscope, and start scanning an electron beam across it. you can do smaller devices than EUV photolithography that way <muurkha>the disadvantage is that an exposure takes hours rather than seconds <muurkha>Sam Zeloof's setup is maskless; he bolted a DLP projector to a microscope <oriansj>hence the have a projector and proper lenses bit <muurkha>yeah, but Leeuwenhoek had the proper lenses <muurkha>not to disparage Leeuwenhoek's manual skill, which merits great honor <oriansj>well to be fair, I don't think I could make lenses of equal quality to Leeuwenhoek's from scratch and I can just buy a projector. <muurkha>but he didn't need the Industrial Revolution to get within a factor of 4 of the fundamental limits of that kind of microscopy <muurkha>(he got to 275×; without resorting to things like phase-contrast microscopy and confocal microscopy, which aren't useful for photolithography, you can maybe get 1000× with visible light with an oil-immersion objective) <oriansj>you only need 1000nm resolution to build a 64bit processor and a couple MB of RAM <muurkha>machinists routinely attained 2500nm resolution for mechanical positioning by the late 19th century, though most work was done to the looser standard of 25000nm <oriansj>after which you can build the systems able to drive a VGA projector; which could be done with minimal circuits. <muurkha>Michelson's diffraction grating ruling engines required more precise control, and they really pushed the limits of what was possible 125 years ago <muurkha>some of them are still in use today to blaze research-grade master diffraction grating masters at the Grating Lab <oriansj>look, once you bootstrap lithography to 130nm (or 65nm if you want to build your own x200 laptop to parity); the rest is just gravy on top <muurkha>I think that's enormously more difficult than getting a single device built <muurkha>but almost 34 years after Eigler and Schweizer, photolithography or any other kind of mass production is still orders of magnitude away from their feat <muurkha>I don't think you even need 1000nm resolution though. an ARM2 32-bit processor is 27000 transistors; a 64-bit equivalent can't be more than 100k. a megabyte of memory is 100 times that size, so that's sort of the tall pole in the tent <muurkha>if you're not going to be using the moral equivalent of a disk or drum or delay line or something <muurkha>27 million transistors, say, three or four per bit. that's 300×300×300; if each memory bit was a cubic centimeter, that would be 3 meters by 3 meters by 3 meters. it would fit in your garage <muurkha>I think the hard part is getting it to work, and to work fast enough, and for long enough <oriansj>well fast isn't a huge problem as a 500khz machine is fast enough to drive a vga projector for lithography purposes <muurkha>if you're doing maskless lithography, a scanning laser might be a reasonable approach <muurkha>but fast *is* a huge problem; it's trivial to write a program that takes a billion, a trillion, or a quadrillion instructions to run <muurkha>getting it down to a million often requires a lot of work <oriansj>but as the ask is non-optimized lithography mask generation and to run a lithography process to produce a single layer on a wafer at a time. <muurkha>you can do very useful levels of industrial process control and automated fabrication on deep-sub-mips machines <muurkha>but I'm typing this on a laptop that typically runs about 20 billion 64-bit instructions per second when it's loaded, and can do several times that with AVX or its GPU; an SDS 940 or Commodore 64 might manage the equivalent of 50000 such instructions per second <muurkha>it can do things in a second that would have taken one of those machines a week <muurkha>it's easy to accidentally get dependent on being able to do those things <oriansj>indeed; and going backwards in terms of hardware requirements is quite hard to do <charlotte1123581>hi guys, i want to bootstrap a linux distro like guix did. i vaguely remember they did some hex0 ... mes ... tcc gcc type of thing. is there any writeup of this process? <stikonas>though this is slightly different from guix <stikonas>but guix process is not suitable for other distros as they wouldn't start with guile <stikonas>if you start with hex0/kaem, you need to change some steps <stikonas>(or if you start with builder-hex0 for kernel bootstrapping too)