Gregory's ServerThe Intel 386 processor (1985) was the first 32-bit processor in the x86 line. Let's take a close look at the processor dies, seeing how Intel shrunk the chip, created new versions, and why the 386 SL jumped from 285,000 transistors to 855,000 transistors. 1/9
 21 comments
Chip designers often puts their initial on a die. The 386, though, has many initials. I zoomed in on the initials, which seem to be next to the functional unit that person worked on. I couldn't identify most of the people, but "SEC" is head mask designer Shirley Carter. 3/9 In 1987, Intel improved their process from 1.5 µm to 1 µm, making a much smaller 386 die. It's not a simple, mechanical process: they rotated the instruction decode unit and changed standard-cell generation, but left the datapath mostly the same. Left: 1.5 µm, right: 1 µm. 4/9 The 386 die cost dropped until the ceramic package cost as much as the die. Intel introduced the stripped-down 386 SX with a 16-bit bus that could go in a cheap plastic package, replacing the 286 in the low-end market. The original 386 was renamed the DX. Photo: 1.5, 1 µm SX. 5/9 The 386 SL (1990) packed a 386 core along with various controllers onto one chip, more compact for the laptop market. The 386 core itself is in the middle of the chip. While the regular 386 had 285,000 transistors, the 386 SL had a whopping 855,000. 6/9 The 386 was critical to the computer industry. While IBM tried to break free of clones with the PS/2, Compaq went its own direction, making the first 386 PC, the Deskpro 386. The Deskpro was a big success and the PS/2 was mostly a failure. 7/9 Because of the 386, Compaq gained control of the PC architecture while IBM ended up eventually exiting the PC market. The 386 led to Intel's first billion-dollar quarter in 1990 and the enduring success of the x86 architecture. 8/9 Credits: thanks to https://twitter.com/Siliconinsid for the die images. The wall-sized 386 photo is from Intel's 1985 annual report. Thanks to Pat Gelsinger who sent me copies of his 1985 papers on the 386. 9/9  @kenshirriff Amazing that intel only discontinued the 386 in 2007 and that it lives on in the embedded market.  @kenshirriff i'd like to argue that IBM still did very well in the PC market, particularly with the Thinkpad series. technically they didn't exit the PC market until 2005. @kenshirriff it would be interesting to learn more details about how the Motorola chips, which you say were more technically advanced, did not win out over Intel. Was it pricing, good luck for Intel, bad luck for Motorola, differences in strategy? @wtfrank Much of Intel's success was due to "Operation Crush", where Intel had a huge marketing campaign to get design wins over Motorola. The "luck" of having the 8088 selected for the IBM PC was also huge. @kenshirriff thanks, I read a bit about that Operation Crush, it seems like a smart organisational & management strategy that managed to motivate the entire organisation to beat Motorola.  @kenshirriff I strongly believe the PS/2 was a failure because they went all-in on Microchannel, rather than have a transition period when their PCs would have both MCA and ISA slots. All the other PC bus transitions had "bridge" motherboards with both buses. I remember ISA and PCI, then PCI and PCIe. IBM f-ed that up big time not giving people a chance to migrate to new peripherals.   @kenshirriff *really* handy for vendors with existing 80286 boards. It gave them a fast track to a (slower, more affordable) i386sx AT. @kenshirriff This is something I've been curious about for a while. Why did chip makers use those fancy ceramic packages during the first few years of a chip's production? It was always like that... New CPU comes out, it came in a pretty ceramic package with a gold cover, etc... then later in its life, cheap plastic package. Did the ceramic package permit easier debugging for early dies? Lower failure rate when packaging dies? @kenshirriff One aerospace project I worked on used obsolete x386 uP's that were running out of stock. Rather than upgrade and recertify, they bought the mask(s) and had enough fabricated to support to EOL. @kenshirriff I haven't skipped ahead but I'm guessing cache and Service Mode added transistors. #i386 |
I created this diagram showing the main functional blocks of the 386. Lots going on: the Data Unit (lower left) does actual computation. Microcode in lower right. Memory paging, segments, and prefetch upper left. Instruction decode center. Bus interface upper right. 2/9