3 ms·

"We’ve built our own super-efficient servers, invented more efficient ways to cool our data centers, and even used advanced machine learning to squeeze more out

by jfim 11y ago

"We’ve built our own super-efficient servers, invented more efficient ways to cool our data centers, and even used advanced machine learning to squeeze more out of every watt of power we consume. Compared to five years ago, we now get 3.5 times the computing power out of the same amount of energy." I'm probably missing something, but isn't that just brought by the usual performance improvements between CPU generations?

duaneb 11y ago
> I'm probably missing something, but isn't that just brought by the usual performance improvements between CPU generations? Load balancing arbitrary computation is definitely a software problem as well as hardware. This is basically the halting problem on an industrial scale—you divvy up work without being able to understand it until it's done. At this scale, I'm also betting there is nearly infinite room to move computation around the data center and a lot of room to save money computing the same number of tasks on fewer cpus.
- wmf 11y ago
  Google just published a great paper on that topic: http://highscalability.com/blog/2015/6/4/paper-heracles-improving-resource-efficiency-at-scale.html http://highscalability.com/blog/2015/6/4/paper-heracles-impr...
Filligree 11y ago
Yes, the same chip can be used in more or less efficient ways. What they're describing includes: - Turning off CPUs that aren't in use. Turning off machines that aren't in use, and moving jobs around to maximize that. - Improving the air-conditioning efficiency. Typical data centers use more energy moving heat around than moving bits; Google has improved vastly on that. - And more. Google has published papers on this stuff.
pervycreeper 11y ago
There's more stuff to power in a data center than just CPUs. Not sure what the increase in performance per watt is in the CPUs Google used in 2010 compared to 2015, either.
munificent 11y ago
> isn't that just brought by the usual performance improvements between CPU generations? Not entirely. The quote is about computing power per unit of energy not per unit of time which is what Moore's Law is about. Your Pentium probably ran at about 150 MHz and consumed around 12 Watts. A Core i7 might run at around 3 GHz, but it's also using 120 Watts to do it. Take a look at the W/Ghz and W/Mhz columns in the charts here: https://en.wikipedia.org/wiki/List_of_CPU_power_dissipation_figures https://en.wikipedia.org/wiki/List_of_CPU_power_dissipation_...

by jfim 11y ago

duaneb 11y ago

> I'm probably missing something, but isn't that just brought by the usual performance improvements between CPU generations? Load balancing arbitrary computation is definitely a software problem as well as hardware. This is basically the halting problem on an industrial scale—you divvy up work without being able to understand it until it's done. At this scale, I'm also betting there is nearly infinite room to move computation around the data center and a lot of room to save money computing the same number of tasks on fewer cpus.

wmf 11y ago
Google just published a great paper on that topic: http://highscalability.com/blog/2015/6/4/paper-heracles-improving-resource-efficiency-at-scale.html http://highscalability.com/blog/2015/6/4/paper-heracles-impr...

Filligree 11y ago

Yes, the same chip can be used in more or less efficient ways. What they're describing includes: - Turning off CPUs that aren't in use. Turning off machines that aren't in use, and moving jobs around to maximize that. - Improving the air-conditioning efficiency. Typical data centers use more energy moving heat around than moving bits; Google has improved vastly on that. - And more. Google has published papers on this stuff.

pervycreeper 11y ago

There's more stuff to power in a data center than just CPUs. Not sure what the increase in performance per watt is in the CPUs Google used in 2010 compared to 2015, either.

munificent 11y ago

> isn't that just brought by the usual performance improvements between CPU generations? Not entirely. The quote is about computing power per unit of energy not per unit of time which is what Moore's Law is about. Your Pentium probably ran at about 150 MHz and consumed around 12 Watts. A Core i7 might run at around 3 GHz, but it's also using 120 Watts to do it. Take a look at the W/Ghz and W/Mhz columns in the charts here: https://en.wikipedia.org/wiki/List_of_CPU_power_dissipation_figures https://en.wikipedia.org/wiki/List_of_CPU_power_dissipation_...