Search is a fundamental component of intelligence or even thought. Maybe that’s why Google is now calling it knowledge. Our brains are already good at search. Look around a room and every object your eye passes is identified in your brain. Something out of place? It catches your eye. That is where we are headed with Internet search, though not exactly the way one might expect.
This is where every new generation of computer scientists brings-up the idea of artificial intelligence. If only we made the network smart enough to know not only what we really mean but what we really need. Maybe someday, but for now don’t hold your breath waiting for that one. Starting in the 1980s fortunes were spent and lost developing artificial intelligence that wasn’t, well, very intelligent. Absent some breakthrough that I have yet to see, this is not a good path to follow even today. Fortunately it isn’t really needed, at least not yet.
Google’s approach is leveraging its existing strength, which is hardware optimization. A couple years ago the company did research to figure out at what processor performance level — at what percentage of CPU capacity — data center power consumption was minimized. No other company but Google would consider a strategy of deliberately throttling-back its data centers.
Whether Google even realized it this approach to transport efficiency has been around for a long time… in aviation. What Google separately sought were two very special data center power levels known in aeronautical engineering as the Breguet Number and Carson’s Speed.
If you’ve never heard of a Breguet Number don’t feel bad. It is the power level (typically represented by a cruising speed) at which a particular aircraft will have the longest range on its internal fuel. Fly faster or slower than the Breguet Number and you won’t go as far before running out of gas. Every airplane has a different Breguet Number, though the rule of thumb says that 32 percent power is pretty close. I don’t know what power level Google came up with from its data center research, but it is likely in that 32 percent range.
Google found that by operating its CPUs at very low power levels it could broadly optimize search in terms of total power consumption. Running at higher power levels (faster CPU clocks) could get you more search results but with a total power bill that was higher in simple terms of watts-per-search.
Operating data centers at their Breguet power level means building three times the facility that Google would need if they ran the place the old fashioned way — balls to the wall.
Deliberately having three times the computing power available brought unintended consequences — the same consequences that any 17 year-old experiences when they replace the stock engine in their old clunker with a powerplant three times as big. Google acquired a lead foot. The first result of that lead foot was Instant Search — using extra CPU cycles to prefetch search results in real time. It’s not something Google set out to do but rather an unintended consequence of overbuilt data centers.
This brings us to Carson’s Speed. Bruguet was a French engineer best-known for his family’s fine watches, while Carson was a professor at the U. S. Naval Academy.
The problem with Breguet Numbers for pilots is that airplanes are intended to go fast and Breguet-friendly power levels are slow and boring. Going faster is a constant temptation with airplanes because they are of necessity built with a lot of excess power — power that is needed for climbing to altitude. An airplane built with an engine small enough to only reach Breguet Number speeds wouldn’t have enough power to even get off the ground. If you have excess power (and finite patience) what is the best speed to fly?
That would be Carson’s speed — the speed to get the most extra speed for the least extra cost. Or, as Carson put it, of finding “the least wasteful way of wasting.” For aircraft the speed in question turned out to be 1.32 times the speed for most miles per gallon (the Bruguet Number). Carson’s Speed uses excess power most efficiently.
Other than three G-V’s and one Boeing 767 built for a harem, Google flies data centers, not airplanes. But Google’s situation going into its power experiment was actually very similar to aviation because it was an exercise in reducing power. Google data centers weren’t built to Bruguet specs, they were faster. Given this excess computing power that had already been paid for in capital terms, what was the most efficient way of using it? Carson’s Speed — about 43 percent power — leaving plenty of excess cycles for new services like Instant Search.
But once you enable Instant Search for everyone, the data center is again running consistently above its Carson’s Speed which means you need even more hardware to bring the building back to 43 percent. It’s an arms race that until this moment only Google may have known they were conducting.
Google competitors have been constantly building new data centers, too, but they never knew there was a specific target beyond just keeping up with Google.
Google sees the excess power above Carson’s Speed as a safety margin in case traffic spikes or a data center goes down, which makes sense. But is also a strategic advantage over Google competitors.
Having discovered its lead foot, Google will employ it more and more. We’ll see whole new types of brute force services aimed at using excess CPU cycles against expanded data sets to reduce the distance between searching and finding. If a question is answered as quickly as it is asked and all the answers are cached and analyzed the results may actually start to appear before they are needed.
That, my friends looks a lot like knowledge.