Windows CE

» see microsoft.com/windowsembedded as it lived, on the wayback machine

~ the eulogy ~

Windows CE was the Windows that was not Windows. Microsoft shipped version 1.0 on 16 November 1996 under the codename Pegasus, and the thing it ran on was a clamshell about the size of a paperback with a keyboard nobody could touch-type on. The name was borrowed. None of the NT or 95 source was in there. It was a new kernel, written to run in a couple of megabytes with a real-time scheduler underneath it, and the reason it said Windows on the box was that the shell put a Start button in the bottom left corner.

The launch devices were called Handheld PCs and they cost around $650, which in 1996 also bought a serviceable desktop. They did not sell. Microsoft kept going anyway, because the kernel was the actual product. CE 3.0 became the base for Pocket PC 2000 and for everything Windows Mobile did afterward, and the same code was quietly going into machines that never advertised what they ran on: the barcode terminal in a warehouse, the panel bolted onto equipment that had no other computer in it.

The versions kept coming and the name kept changing. CE.NET in 2002, CE 5.0 in 2004, then CE 6.0 in 2006 with a rewritten kernel that gave every process its own two gigabytes of address space instead of the 32 slots the old design allowed. Microsoft then renamed the line Windows Embedded Compact, shipped Compact 7 in 2011 and Compact 2013 on 13 June 2013, and stopped. Mainstream support for that final version ended on 9 October 2018. Extended support ended on 10 October 2023. OEM license sales run until 2028, so there are plants building new hardware right now on an operating system that has been out of support for three years.

I owned none of the pocket devices. A $650 handheld in 1997 was not a purchase anyone around me was going to consider, and the machine that mattered was the desktop with the modem hanging off it. Where I actually met Windows CE was later and from the other direction. I work on embedded vision now, and the shape of the CE problem is the shape of the job: a fixed device with a small memory budget and a deadline the software has to hit every single time. CE lost the pocket and won the machines nobody calls computers, and it held that ground for twenty years because the alternative was writing your own scheduler.

~ the dossier ~

~ what it actually was ~

CE was sold to manufacturers, not to people. The manufacturer opened a tool called Platform Builder and chose which components went into the build. What came out the other end was an operating system compiled for exactly one device. Two CE machines sitting next to each other on a shelf could have almost nothing in common: different processor architecture, different set of drivers, different subset of the API. That was the pitch and also the trap. Nobody could write a CE application and expect it to run on CE generally.

The other half of the pitch was determinism. CE had bounded interrupt latency and a priority scheme you could reason about, which is why it went into things that had to respond on a schedule rather than eventually. Microsoft also let licensees see and modify a large part of the source, which was unusual for the company at the time and was the only way an OEM could ship a device it could not patch later. Once a CE device left the factory it was finished. There was no update channel, and mostly there was no expectation of one.

~ the dreamcast detour ~

Sega licensed a CE build for the Dreamcast so that studios porting from the PC would find DirectX where they expected it. Microsoft announced the Windows CE Toolkit 2.0 for the console on 16 September 1999. It was never the console's operating system in the way people remember it: the machine booted Sega's own environment, most of the good games used Sega's Katana SDK directly, and the CE layer was an optional runtime shipped on the game disc by studios that wanted it.

What survives is the logo. A stack of Dreamcast cases carries a small Windows CE mark on the spine, and it is the only place most people ever consciously saw the name. A console with Microsoft branding on the box in 1999 looked at the time like a partnership. Two years later Microsoft shipped the Xbox and Sega left the hardware business.

~ the long tail ~

The reason the 2023 date matters to anyone is that CE ended up inside equipment with a twenty-year service life. Infusion pumps, patient monitors, ticket machines, industrial handhelds sold by the pallet to logistics companies. Whole categories of device were designed around it, and when support ended the vendors had no drop-in replacement to offer, only a migration: rewrite for Windows 10 IoT, or rewrite for Linux, which is what most of them did and had been doing quietly for a decade already.

That is the real cause of death and it happened long before the support page changed. Embedded Linux got a real-time patch set and free toolchains, and it was cheaper than paying Microsoft a per-unit royalty on a device selling for two hundred dollars. CE did not lose an argument. It lost a price comparison, one product line at a time, over about fifteen years.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »