Transmeta Crusoe, 20 years later
There is a saying: Oldies, but Goldies. 20 years ago Transmeta was clearly onto something. When we look back at what they achieved with THAT technology one can only wonder on how would the computing landscape look today if Intel did not use anything and everything in its anticompetitive armoury to kill it.
I realised that when rearranging the computing at home while trying to chase down a rather elusive hang specific to 6 core/12 thread Ryzens. In order to minimise the amount of factors under investigation I decided to move the printer, scanner and other stuff off the Ryzen desktop to another Intel machine. There was nothing else available in a room which had place for the rather unwieldy HP CP1025 as well as the ancient Epson GT8200, so I decided to move both of them to a dedicated print/scan server. This is trivial on a modern Linux - network printing is mostly plug-n-play and the scanner can be easily exported over network using saned.
I would have done it using a Razzie, if not for the x86-only binary plugin for the HP1025 (thank you HP). So I grudgingly pulled the old x86 thin clients pile out of the dusty drawer. All, but one, showed unacceptable power consumption ~ 15-25W. That is 25 North Atlantic Pesos per year for the honour of hooking up two USB devices. Only the Transmeta showed 9-11W despite using an obsolete low efficiency 12V brick. Switching to a modern power supply dropped that to 8-10W.
I decided to investigate if I can get it any further. The machine was reporting single frequency and no power management. That for Transmeta Crusoe is actually false and nobody has bothered documenting it - you can crank down ANY Crusoe by telling the Linux kernel that its minimum frequency is let's say 300.
echo 300000 > /sys/devices/system/cpu/cpu0/cpufreq/scaling_min_freq
It will not allow you to go under some (machine dependent) frequency. It will work fine and will nicely oscillate between frequencies in small steps even on the ancient 3.16 kernel which comes with Debian Jessie. This is something which other CPUs and SoCs have barely managed to achieve in 2017+ - 15+ years later and they still require extensive OS software support for that. Transmeta had that in 2000!!!
This dropped the power consumption to 7W. With 2000 tech. If we had that today we would be way beyond what Arm delivers per W and per performance unit. It is a pity Eu got on Intel's case only after AMD asked for it. It should have been on its case day one when Transmeta, or even earlier in the days of UMC U5 and IBM BL.
Unfortunately whatever happened happened and we are where we are. It is a pity that it is also being dropped by major distros due to GCC changes. It still runs and it is still fit for purpose where you need something small, cheap and reliable. 20 years later. Amazing... To say the least.