| I've heard the assumption that energy costs outweigh cpu costs for TCO, but after some rough guesstimation I'm not so sure. 1MWh of electricity is enough to power a cpu running 114W on average for a year. A 24 core (48 thread) xeon has a 165w tdp, meaning that even such a beast is likely to use less on average (most applications don't have 100% cpu load 100% of the time, and virtualizing up to that load causes serious issues when several clients happen to simulatneously peak). In essence, this should run pretty much any single CPU, and even if you're doing something weird to actually get that 100% load, we're not far off. 1MWh of electricity is around 40$ at wholesale prices (https://www.eia.gov/electricity/monthly/update/wholesale_mar...). Even if google pays 100, then it looks to me as if straightforward server cpu energy usage is unlikely to be a huge issue. Of course, there are power supply losses and cooling overheads, but even if that raises the power consumption by a factor 5 (doubtful), then a several-thousand dollar server cpu is still going to outweigh energy costs for many years. Furthermore, energy costs are trending downwards and cpu's are still getting more efficient; it's a somewhat risky investment thus to buy a more expensive (but more efficient) chip if that investment needs years to pay off. So, it looks to me as though plain old cpu costs certainly are far from irrelevant; they may indeed exceed energy costs over the chips lifetime if the energy is acquired at close to wholesale prices and cooling is efficient (say 3x overhead?). |