So, why is it not on the table for... anything in FOSS? Doesn't seem rational. Even a bit hypocritical given vendors like Gaisler and nonprofits like SPARC International have met FOSS halfway or almost wholly. Unlike the others that sue FOSS developers.
Far as power efficiency, kristoffer might be able to chime in as it's not in the data sheets for Gaisler. That's suspicious: either the numbers are bad or they leave it off given its meant for customization. Anyway, the Leon4...
http://www.gaisler.com/index.php/products/processors/leon4?t...
...uses 30,000 gates per core. Same ballpark as ARM and MIPS. Power use should be similar or at least acceptable if comparing ARM, MIPS, and Leon on same ASIC process. They often do rad-hard given it makes it resistant to SEU errors. That takes plenty of extra circuitry. Numbers I have for those, the high end, are 15mW per 1Mhz for Leon3RadHard and for Leon4RadHardQuadCore max was 6watts per one slideshow.
I'll take 6watts consumption in a router in exchange for quad-core, IOMMU-enabled, fault-tolerant, open CPU. What about you? Would 6 watts kill it for you?
The router vendor surely looks at the cost of the CPU - ARM and MIPS cores are probably much cheaper.
Keep guessing.
So, it's mainly the ecosystem with companies and FOSS people wanting to benefit from what's already there instead of improve FOSS HW ecosystems. There's currently, but not indefinitely as you said, a cost advantage for the mass market SOC's as well for PPC, ARM, MIPS, and possibly SuperH.
$4 per 100 units (4 pennies each), or $4 per unit in quantity of 100?
Compare things that are alike. If you take a manufacturer (say: TSMC), pick its node (say: 16nm FF+) and you decide on a package (physical manifestation of RTL primitives in the silicon) you get better performance per watt on one architecture over some other. ARM and MIPS are inherently very power efficient. You can't just take SPARC and make it more power efficient than these two. It doesn't work like that.
It's also not true that ISA doesn't matter. ISA impacts bandwidth requirements heavily. This in turn impacts latency and latency hiding, cache requirements and many other things. In fact data transfer is typically as costly as (if not more expensive than) computation. Getting data to all the right places on the schedule eats power like crazy. This is exactly why ARM has Thumb. It's not like internally core does different things than it would do with wide ISA. It's just that stuff's more densely packed, which helps tremendously.
Which brings me to my last point. There's an open architecture that's quite nice. It's SuperH (or SH2 in its open source form), which in turn is what ARM's Thumb is based on. It's not perfect, but it's pretty solid. Omitting it in the OP makes me think author isn't very thorough with his research. But everything has to start somewhere. ;)
Sure you have Thumb, that saves a little on memory bandwidth which is good. But nobody uses it anyways, and you might run slower so you can't sleep as fast.
I like the J2 (the open source sh) project as well but it doesn't have a MMU which rules it out for anything but simpler embedded projects.