r/c/s

2026-11-18

How to Kill God: Evolving the PC to crush Unix 1993-2003

Contrary to popular belief, the Catholic Church wasn't as opposed to the existence of the printing press as the later Protestant reformation lore tells you - the Gutenberg Bible was printed chiefly for Catholic monasteries, clergymen, and the already literate (which, in per-reformation western Europe, basically means wealthy and in cahoots with the Church). It got to a point where these bibles had buyers before they were complete (sounds kinda familiar if you ask me).

Nevertheless, the Protestant framing of the printing press as a divine implement to break Rome's grip was very much useful to the Lutherans, Anglicans, and later anti-theists who started the Enlightenment. Latching onto the one writer who wrote that the Pope "must destroy the printing press least printing destroys the Church" thus becomes an extremely useful narrative device for many. Technology has always had an undertone of it being utilized by a lower, oppressed class to ultimately usurp the "order of things," which is why I find the framing so fascinating.


By 1993, the Amiga and the Atari ST computers were breathing their last breaths. Apple was like a fish on land not knowing how to exist despite their formidable creative base, and the market more or less coalesced into two camps.

The first camp were the sages - these were people who lived the computers in the minicomputer days and finally found themselves with their Suns and SGIs and RS/6000s on their desks. Sophisticated boxes designed for reliability and performance similar to the VAXen and PDPs that they were used to, shrunken to the desktop format. Most of these boxes lived in university and commercial research labs, but that was also where the "hacker culture" formed in the first place - where the practical jokes and ethos was shaped.

On the other, the unwashed, unenlightened masses. At this point the only game left in the market was the IBM PC compatible. A machine for the paper-pushers. To the hacker, the PC was a dead end. Spreadsheets, word processors, a good game or two on occasion, but nothing like how we romanticize DOS gaming now.

Just to put into perspective, this was what the hacker saw in 1993:

Unix WorkstationIBM PC Compatible
CPURISC Superscalar CPU 50-200MHz, out-of-order386-486 @ 33-66MHz
Pentium @ 60-66MHz
RAM16MB - 64MB4MB - 16MB
HDDSCSI Bus Mastering 400MB-1GBIDE, ~100MB-400MB
GFX16-bit or 24-bit color, >= 1152x900ISA or VLB VGA, 256 colors, 320x240, 640x480
3DSGI Indigo2 XL, XZ, Extremelmao, no
NET10BASE-2 or 10BASE-T network2400 - 14.4k modem
always on, NFS storagemostly BBSes, maybe NetWare in an office
OSPreemptive multitaskingNo multitasking or cooperative multitasking
Paged virtual memoryNo paging
Kernelspace/Userspace separationlmao, no. you have DOS.

The future really did seem, if you were one of the sages, to be that a scaled-down Unix machine would be the machines that everyone would have on their desks, and not the other way around. Here's a quote from a famous mailing list thread:

Of course 5 years from now that will be different, but 5 years from now everyone will be running free GNU on their 200 MIPS, 64M SPARCstation-5.

But of course, we know that the PC was the platform that won the war, and not the Unix machine. How? How do you bring platform with over a decade of baggage to the modern era?

Step 0 - a proper bus for the next generation of PCs

Before anything else down this list has any chance of succeeding, the PC had to reckon with its bus. The most common was the 16-bit ISA bus, cheap, but woefully inadequate. IBM tried and failed with MCA (because they pissed off the clone makers), the Gang of Nine tried and failed (EISA was still tied to the 8.33MHz ISA clock), and VESA tried and failed (VLB was tied to the 486's characteristics and nobody implemented bus mastering properly).

Enter the PCI bus. This is the true step up from the "simple" trick of glorified extension cables between the CPU and expansion card. PCI offers proper arbitration that made bus mastering real, first-class plug-and-play (goodbye jumpers!), and most importantly, a host bridge that decoupled the CPU FSB from the downstream devices. Without this decoupling, every bus would have been as quirky and one-generational as VLB.

The final thing that cemented PCI as the bus that will carry the PC to domination was openness. Intel created the PCI-SIG industry group and gave it the spec as soon as the first revision was done, allowing for chipset makers to pump out PCI motherboards for 486 and Pentium sockets as soon as they were ready to. The openness and platform agnostic design also allowed other platforms to adopt it. Notably, the PowerPC camp switched to PCI from NuBus as well.

Step 1 - IDE Bus Mastering

While a multitasking OS doesn't require a DMA hard disk, it would sure be nice if reading and saving files isn't 100%ing the CPU with rep insw and rep outsw calls. This is the premise of bus mastering - allowing the drive to read and write data from/to RAM while the CPU goes to do other things.

IDE, as originally implemented, is kinda dumb. It's an ISA subset extension cable: D0-D15 from ISA, R/W strobes, IRQ14, and reset. The old 8237 DMA chip was also not suitable for bus mastering from a large hard drive, thanks to its 64KB limit.

Enter the PIIX - the PCI-ISA-IDE Xcelerator, a southbridge that implemented two bus-master IDEs channels. It retained the I/O port path, but adds a PCI BAR that exposes the PIIX's new bus-mastering IDE mode. Legacy code still did the strobe-and-rep dance, but competent OSes could flip some bits and hand hard drives a memory window and let it do the rest.

Now your :wq! or Ctrl+S doesn't stop your modtracker, and starting a new program didn't bog down your email client!

Step 2 - The P6 and its consequences

Probably the most important thing in the evolution of PC sits right here with the 686 architecture. If anything was the "most important" evolution that led to the PC killing the Unix workstation, it was this one. So why the P6?

Research into CPUs in and around this era was very much focused on the idea of RISC, or Reduced Instruction Set Computer. It would not be an exaggeration to say that all the Unix workstations in 1993 would be RISC machines, or at least RISC-inspired. Part of the view from the ivory tower was that the PC, if it had any ability to survive, had to do the same painful transition that all Unix platforms had going from the Motorola 68000 to {SPARC, MIPS, PowerPC, PA-RISC}.

The P6 was the solution to that.

For comparison, the P5 looked like this, and was a typical pipelined design not too dissimilar from the "textbook MIPS pipeline." In fact if you are familiar with that design, you should be able to identify which units correspond to their MIPS counterparts. The issue here is the U/V pipe in the execution unit - ideally, you'd want to utilize both pipes here, but the P5 Pentium had arcane pairing rules for utilizing the V pipe that made pairing (and thus, simultaneous execution of two instructions at once) not very practical.

p5 block diagram

The P6 looked like this. Needless to say, a lot more complicated. There are three things here that make the P6 a true leap from the P5 and put the x86 on equal footing as the most advanced of RISC machines.

  • Decoder: the P6 decoded the x86 instructions into smaller micro-ops. This turned the P6 into something a lot more RISC shaped. Intel bet (correctly) that Moore's Law means the decode circuitry will shrink relative to the total usable size, and basically destroyed the RISC argument right then and there.

  • Reservation stations: Oh yeah, we're fully out-of-order on the P6. Goodbye WAR and WAW hazards, and hello fearless 100%ing of the underlying execution engine.

  • Reorder buffer: With x86, instructions are supposed to be retired in-order. The reorder buffer allows the CPU to lie about the underlying execution of instructions and essentially look like things are going in-order.

p6 block diagram

The Pentium Pro, the first CPU that used this new P6 architecture, was expensive and not exactly within-reach for most. But it was proof that a powerful RISC-like CPU, combined with a competent OS, could upend the "natural order of things" while keeping backwards compatibility with existing software. So where are these competent OSes?

Step 3 - Moving away from DOS

While all this was happening, the evil overlords at Microsoft have been priming developers of PC software with a shiny new API called Win32. It was the easiest way of writing GUI apps and targeting a large market of Windows users. Wouldn't it be nice if Win32 applications could run on a proper kernel?

Enter Dave Cutler's revenge - NT is essentially the "proper kernel" I was eluding to - Cutler came from the world of DEC VMS, where he experimented with preemptive multitasking, paging and virtual memory, and an explicit kernel/userspace separation. By the time Windows NT 4.0 came out in 1996, there was already a wealth of Win32 apps shipped for Windows 95 (and thus plug-and-play on Windows NT 4.0), and a competent processor with the Pentium Pro. A rock-solid platform for server and mission-critical workstation tasks like CAD and Win32 programming.

The sages of Unix, however, have their own platform, it's called Unix! So, a Unix OS was needed to entice them over. A shrewd programmer in Finland had been writing a kernel for the PC, with the 386 in mind. The sages of Unix did have something the PC crowd could use, however - a sect of these sages had source code for most of a Unix userland ready to go. The GNU userland was already popular with many a Unix sage, so combining it with this newfangled kernel provided a lawsuit-unencumbered Unix environment (Unix itself was going through a rough lawsuit period between Novell and the UC regents). The PC suddenly had an enticing platform for these sages - and it was cheap. The economics of PC's enormous scale hit the sages like a disputation on a church door.

Step 4 - Graphics

While P6, PCI, IDE Bus Mastering, and a choice of NT and Linux sufficiently chipped away most of the Unix workstation's advantages, they still had their golden child in SGI. These machines were slick, new-age devices with gorgeous UIs, was where 3D research lived, and it even had cultural cachet ("It's a UNIX system, I know this!" refered to an SGI machine). How to slay this beast?

The first step was the 3Dfx Voodoo. 3Dfx was founded by ex-SGI engineers who wanted to build hardware-accelerated 3D cards not unlike the ones that were in the SGI machines they helped develop. The 3DFx Voodoo was essentially, a card that implemented the bare minimum of OpenGL (the API that SGI spun out as an open standard) needed to allow for video games to run. It wasn't quite enough for work stuff like CAD (which was the original goal for OpenGL) but the sheer scale of Voodoo and Voodoo2 meant that PC suddenly had 3D investment money in the form of Voodoo card sales and video games. Money that could rival any amount of research-based funding, especially after the end of the Cold War.

The next step, IMO, is the Geforce4 Ti, at the tail end of our era of discussion in 2003. Prior to this, cards like the Voodoos and RIVA TNT2 used a fixed pixel pipe. The Voodoo even separated rasterization and texture mapping into two discrete chips. Going from the fixed model to a shader model was the first step in transforming the GPU from mere raster acceleration to GPGPU like we have with CUDA and Ray-Tracing now in 2026. The seeds of computing now was laid right here at the grave of the Unix workstation.

Conclusion

So why did I write this? Well, I find the narrative fascinating, especially laid out like this, because there was never really any unified "let's kill the Unix machine" call from the PC manufacturers (other than at Intel for a brief period). It truly is beautiful. The other part of me thinks that us, the consumers of computers, may have to take up the implements to kill God (or a upperclass of sages) yet again, and I'm not sure if the PC is the platform to do so in 2026. Thoughts?



macos debian ffmpeg firefox csh wii color xp xenia rainbow bi trans enby dolby web3 mt32 atom firefox2 archive ublock nixos computer chill-pill momg