《那些古怪又讓人憂心的問題》第28期:人力計算機(3)
This is tough to calculate. We can easily come up with benchmark scores for various types of computers, but how do you measure the instructions per second of, say, the chip in a Furby?
Most of the transistors in the world are in microchips not designed to run these tests. If we're assuming that all humans are being modified (trained) to carry out the benchmark calculations, how much effort should we spend to modify each computer chip so it can run the benchmark? To avoid this problem, we can instead estimate the aggregate power of all the world's computing devices by counting transistors. It turns out that processors from the 1980s and processors from today have a roughly similar ratio of transistors to MIPS-about 30 transistors per instruction per second, give or take an order of magnitude. A paper by Gordon Moore (of Moore's law fame) gives figures for the total number of transistors manufactured per year since the 1950s. It looks something like this: Using our ratio, we can convert the number of transistors to a total amount of computing power. This tells us that a typical modern laptop, which has a benchmark score in the tens of thousands of MIPS, has more computing power than existed in the entire world in 1965. By that measure, the year when the combined power of computers finally pulled ahead of the combined computing power of humans was 1977.
世界上絕大多數晶體管都封裝在並非專門用於這種測試的芯片裏，如果假設所有的人類都經過訓練能夠進行基準計算的話，那麼需要花多少功夫才能修改每一臺計算機的芯片以使它們能夠進行基準測試呢？爲了避免這種問題，我們可以通過數晶體管的數目來粗略估計全球所有計算設備的總計算能力。結果我發現20世紀80年代的處理器和今天的處理器的晶體管數目與MIPS的比值大致相同——這個比值大約爲每秒每條指令需要30個晶體管，數據可能誤差一個數量級。戈登•摩爾（著名的摩爾定律的發現者）發表的一篇論文中給出了自20世紀50年代以來每年生產的晶體管總量。這些數字畫成圖表之後長這樣：有了這些比值，我們就能把晶體管總數折算成總計算能力。這意味着一臺基準測試結果爲幾萬MIPS的現代普通筆記本電腦的計算能力超過1965年全球總人口的計算能力。按照這種算法，計算機的總運算能力超越全人類的總計算能力應該發生在1977年。The complexity of neurons
神經的複雜度
Again, making people do pencil-and-paper CPU benchmarks is a phenomenally silly way to measure human computing power. Measured by complexity, our brains are more sophisticated than any supercomputer. Right? There are projects that attempt to use supercomputers to fully simulate a brain at the level of individual synapses.5 If we look at how many processors and how much time these simulations require, we can come up with a figure for the number of transistors required to equal the complexity of the human brain. The numbers from a 2013 run of the Japanese K supercomputer suggest a figure of 1015 transistors per human brain.6 By this measure, it wasn't until the year 1988 that all the logic circuits in the world added up to the complexity of a single brain . . . and the total complexity of all our circuits is still dwarfed by the total complexity of all brains. Under Moore's law–based projections, and using these simulation figures, computers won't pull ahead of humans until the year 2036.7
我想再次重申一下，讓人類拿紙筆做CPU基準測試來得出人類的計算能力是一個很愚蠢的方法。從複雜度上來看，我們的大腦比任何一臺超級計算機都要複雜，沒錯吧？絕大多數時這是沒錯的。現在有些項目致力於用超級計算機來完整模擬大腦單獨一個突觸的功能。6如果我們能看到這些實驗動用了多少處理器和時間，我們就能大致猜測出要媲美人類全腦複雜度需要多少個晶體管。2013年日本“京”超級計算機經過測試得出的結果是，每個人腦相當於1015個晶體管。7這樣算來，直到1988年全世界所有的邏輯電路加在一起才能抵得上一個人類大腦的複雜度……而與所有人腦加在一起的複雜度比起來，這些電路的總複雜度根本不值一提。如果摩爾定律預測的趨勢持續保持下去的話，根據這些模擬結果，計算機要在2036年才能超過人類。
