These were very common, and not bad goers at all. Notice the second-sourcing; Harris, AMD, and Siemens all made the 286-12 under license. Motherboards had improved a great deal by the time these became common: simpler, more highly integrated, and much more reliable.

The 286-12 was the first 286 to be a really practical, affordable part for most people, and eventually became the second most common of them all.

→ Illustration: an unidentified, fairly late model 286 main board from the days when the 286-12 was the volume seller. There are several things to notice about it: first, the much lower chip count — those three large Headland chips you can see (one almost out of shot) are pretty much all there is to it, aside from RAM and BIOS chips. Next, the CPU itself is AMD manufactured: AMD had yet to venture out into designing its own X86 CPUs but was manufacturing under license from Intel in considerable volume. Third is the empty 80287 socket on the right. Though one was always provided, it was very rare to see it utilised.

68-pin DIP or PGAIntelIntel, AMD, Harris, Siemens198380287 or Weitek
Internal clockExternal clockL1 cacheWidthTransistor count
12MHz12MHznone16-bit134 thousand

The definitive 286, and easily the most common one. We still used to see them from time to time right up to the end of the century, plugging away at DOS wordprocessing, running cash registers, still going strong. The one illustrated was manufactured by Harris Semiconductor, one of the several second-source makers.

This was the 286's high-water mark; the -16 was almost universal: the recipe was 286-16, 1MB RAM in DIPPs or SIPPs or sometimes SIMMs, a 256k VGA card and a 40MB IDE stepper hard drive from Miniscribe, Seagate or Western Digital.

(A short PS: we last re-wrote this entry in the evening of 7/7/98 — and fate being what it is, guess what came into the workshop for an upgrade the very next day: an immaculate AMD 286-16; Headland chipset, 1MB RAM, nice little 40MB Western Digital Caviar hard drive, Paradise 256k VGA card. It seemed a shame to gut it. The system soon became a 200MHz C6 WinChip with 32MB of RAM and a 1.7GB hard drive. We threw the 286 board away, but not without regret.)

68-pin DIP or PGAIntelIntel, AMD, Harris, Siemens198380287 or Weitek
Internal clockExternal clockL1 cacheWidthTransistor count
16MHz16MHznone16-bit134 thousand

The 386 was a huge advance but you'd never know it from one of these little slugs — they were usually out-performed by the better 286s. Technically, it's possible to run Windows 95 or 98 on an SX-16 — we've seen it done — though the proper term is probably not 'run', it is 'crawl'.

The original 386 (after the debut of the 386SX Intel renamed the original to 386DX) was a full-blown 32-bit monster, and very expensive. This, the 16-bit/32-bit hybrid 386SX, came along three years later. Just like the 8/16-bit hybrid 8088 a few years earlier, the 386SX was designed to provide a way of using existing motherboard components. The SX chip was much cheaper than the DX (though still quite dear) and boards for it, being being only 16-bit and essentially the same as 286 boards, were cheaper too.

Most SX-16s were shipped with 1MB of RAM and were used to run DOS applications like Word Perfect 5.1 or Lotus 123. They did this quite well, but no better than a similarly-clocked 286. Judging by the retailer's warranty sticker (just out of shot) the one illustrated became some wealthy family's pride and joy for Christmas 1991. This was the time when the Intel stranglehold on CPUs was just starting to break: the AMD 386DX-40 was in production but not yet common and Cyrix was hard at work on its own in-house designs. Before another year was out, the average price of a CPU would have halved and the performance more than doubled.

Even with 4MB or 8MB RAM, you wouldn't want to run Windows 3.1 on a 386SX-16 though. The SX-33s and DX-40s that followed soon after were vastly faster.

132-pin PGAIntelIntelJune 1988387SX
Internal clockExternal clockL1 cacheWidthTransistor count
16MHz16MHznone16/32-bit hybrid275 thousand

Originally just called the 386-16. This was the first 386, and is very rare.

The Intel 386 was easily the most significant X86 CPU of them all. Essentially, modern chips like the Athlon and the Pentium 4 are just very fast 386s. Modern software uses the 386 architecture, the newer chips just do much the same things only faster. The 386 was a huge advance over the 8086 and the '286. In fact, it was designed before the 286, but was too difficult to make and use at first. Intel had to wait and let the technology mature for a few years before the world was ready for the 32-bit 386.

The 386 had vastly better memory management than the 286, and built-in multi-tasking features to allow the development of the powerful modern operating systems that we still use today under various names: OS/2 (which became EcomStation), Windows NT (which became Windows 2000, then Windows XP) and Linux.

Unlike the 286, the 386 could switch back and forward from Protected Mode to Real Mode easily, and it introduced a wonderful new mode called Virtual 8086 Mode. It was this Virtual 8086 Mode that set the 386 apart: the chip could run multiple different programs all at the same time, each program 'thinking' it has a complete 8086 all to itself. This is how OS/2 achieved its incredible crash-protected DOS multi-tasking. Windows NT (and to a lesser extent Windows 95/98) followed suit. Virtually all modern software requires a 386, and none of the later X86 CPUs have made the same sort of quantum leap into the future that the 386 did. It was Intel's finest achievement.

132-pin PGAIntelIntelOctober 1985387
Internal clockExternal clockL1 cacheWidthTransistor count
16MHz16MHznone32-bit275 thousand

Very rare, very quick. AMD and Harris made these little-known racehorses after Intel decided it could live without its former partners and refused to license them to manufacture 386 parts. Harris went on to do its own 286-25 (see below), then exited the CPU making business.

AMD took a different view of matters. So far as the Texan firm was concerned, it had a contract and it was sticking to it. AMD went on to make its own mostly Intel-designed 386 parts, and eventually would follow Cyrix's lead and do its own design work too. This was just the very modest start for AMD as a major CPU manufacturer. For the rest of the AMD story, read on.

We didn't do much formal benchmarking in those days — most of the time it just wasn't needed as you could tell two different speed grades apart at a glance — but we always had a strong suspicion that a good 286 was actually faster than a 386SX at the same clock speed, certainly for DOS applications, which was what everybody ran in those days. We should have measured some while they were still available. If we get an old 286-20 come in one quiet day, we'll take a 386SX-25 to compare it to and try the theory out.

68-pin DIPIntelAMD, Harris, SiemensAbout 1990287
Internal clockExternal clockL1 cacheWidthTransistor count
20MHz20MHznone16-bit134 thousand

These were a little faster than an SX-16. But then, so was custard.

We vividly remember how they seemed to be slower than a 286-20 too, which they had no right to be: but this was probably more a matter of psychology than of science. After all, you expected a 386 to be fast — it was the very latest technology after all — and you expected a 286 to cruise along at a more modest pace, because it was an old design and slated for the axe before too long. So when you ran them side by side, the 286 perhaps seemed faster than it was. Actually, there was not much difference: both parts could only run the software that was available at the time, and that meant that for all their fancy features they were both operating as mere fast 8086s.

Like all the earlier X86 chips, the 386SX had provision for an external floating-point maths co-processor. For an 8088 you could buy an 8087 co-pro; for a 286, a 287; for a 386, a 387.

The purpose of the co-pro was (and still is) to do complex maths. A standard CPU like a 386 was already very good at doing integer (whole number) maths (at least by the standards of the day), and the vast majority of computer tasks involved simple integers. When some fractional calculations were required, the 386 (or 8086 or 286) was quite capable of doing them too, though more slowly. But some uses require very many complex fractional calculations — examples are scientific and engineering work, CAD, and more recently 3D graphics. The co-pro (or NPU) is specifically designed for floating point maths, and is typically eight or ten times faster at it than the stand-alone CPU. Nearly all 386 boards had an empty socket for an NPU. Current CPUs, of course, all have the NPU integrated into the main chip. (In fact, most computer tasks are integer-based even today — although the role of the NPU has increased over the years, it is still much less important than the integer unit.)

132-pin PGAIntelIntelJanuary 1989387SX
Internal clockExternal clockL1 cacheWidthTransistor count
20MHz20MHznone16/32-bit hybrid275 thousand

We used to think these were a high-performance part once. It seems hard to believe now — it was only by comparison to a 286 or 386-16 that these were anything special.

→ Until the AMD parts brought competition to the market, the 386SX was very hard to justify in price-performance terms. The big price drops that followed made the SX much more attractive for six or twelve short months until the long, long reign of the 386DX-40 began.

In reality, the 386SX did not last well. It began as a cheap way to gain access to the 386 instruction set without having to go to the expense of a 32-bit mainboard, but it soon became an excuse for Intel to shake off the second source manufacturers and charge inflated prices for a rather ordinary CPU.

Although it was substantially dearer than a 286, it was just as crippled by its slow 16-bit memory access. It did offer the benefit of the 386 instruction set — which was vastly better than that of the brain-dead 286 — but it didn't offer any particular performance advantage. By the time software packages arrived that could actually use the 386 instruction set, they demanded performance that was far beyond the ability of even an SX-40, let alone an SX-25.

132-pin PGAIntelIntel, AMDJanuary 1989387SX
Internal clockExternal clockL1 cacheWidthTransistor count
25MHz25MHznone16/32-bit hybrid275 thousand

We only ever saw two of these. They flew! We had one in the shop for a while, and used to ask friends to guess what was inside it without looking. Most people thought it was a 386DX-40.

We saw an 8088-12 once too. (Actually a V20.) The old XT had a fixed frequency ISA bus, in other words, the video and controller cards had to run at the same speed as the CPU. (In a 286, or a Pentium II, the ISA bus runs at 8MHz no matter what speed the CPU does.) This hot-rod XT used to have to boot at 4.77MHz with the turbo switched off, otherwise the CGA video card would hang! Looking back, we wonder what you were supposed to do if you wanted a maths co-pro chip for a 286-25 — there were several 20MHz 287 chips available but so far as we know nothing at 25MHz. Perhaps you just got a 386 in the first place.

68-pin DIPIntelHarris1989external
Internal clockExternal clockL1 cacheWidthTransistor count
25MHz25MHznone16-bit134 thousand