I have a certain sadness as I write this. Data center computational densities have incurred a setback.
There was a time when it looked as though ARM might give power-hungry Intel designs a run for their money in the world of high-density computing. It's the sort of density that cloud providers need: rack after rack, crammed to the gills, chilled, high-speed buses. But power costs a lot of money, alternate energy initiatives aside.
AMD had bought SeaMicro, whose high-density chassis full of power-sipping ARM CPUs form large arrays of calculative strength, without the hefty bill from the power company for oceans of coulombs. HP had initially announced Project Moonshot, the cartridge-based high-density server with ARM, or FPGA cartridges to slowly sip power, but ultimately delivered its chassis with Intel Atom. ARM blades are still available, and FPGAs are said to be shipping.
The folks at Nvidia are shipping high-density, low-power-consumption arrays with a fair degree of success. Taking over even a rackful in data centers is rare. Organizations like HP and Dell are cramming blade chassis into racks with surprising success, given the competition from custom data center designs popularized by Facebook, Google, and others.
The high-density designs are extremely muscular. Some have veneers of hypervisors, which tend to push power consumption to the max on platforms where they live. The once-touted power vs performance initiatives, where an operating system could push back power consumption during idle periods, seems a thing of the past.
Electric cars and motorcycles have the same psychological problem: potential range issues. I dream of a Zero Motorcycle, where I could ride from Bloomington, Indiana, to Indianapolis and back on a single charge. They're available, but they're fifteen grand or more. They're not selling in huge numbers because it takes an hour or so to recharge them. Deciding to hop on and randomly ride to California isn't easy when it requires planning stops for charging.
In the same way, throttling back CPUs so one can save power seems laudable, but the data center cooling capacities aren't entirely flexible, so the data center HVAC blowers will blow the maximum cooling—they're not really able to throttle back in millisecond intervals just because an app went quiet for a while.
For now, cutting back on grid expenses based on lower power consumption of the base platform via native CPU power consumption has hit one more rock on its way to the shoals. AMD slowed it down. ARM initiatives will need a new protagonist, one with enough capital and vision to grow the low-power consumption meme, organically.