Moore's Law is in the news again since Intel, having maintained for years that nothing is slowing down, finally showed what they really think. It is a golden rule in economics to look at what people do and not what they say, what is known as "revealed preference." On last week's earnings call, the CFO Stacy Smith said: As the economic useful life of our manufacturing equipment lengthens, we are extending the depreciable life of equipment in our factories from four to five years...We did an in-depth analysis based on the cadence of moving from one process technology node to the next. We talked about that at the beginning of 2015, and the third wave of products. And we completed our long-range planning in the fourth quarter, and that's what triggered the change in the depreciation cadence. (It is hard to type the word cadence without my fingers automatically capitalizing it). So what this all means is that Intel accepts that Moore's Law is slowing down, so equipment will be used for longer, and so depreciation needs to be spread out over a longer period. This sort of accounting decision is somewhat arbitrary but can have a big effect on the financials of a company even though nothing has changed underneath. It is like when Hollywood studios change the percentage of expenses to be set against theater screening versus DVD and streaming. Obviously it used to be 100% and it is dishonest to pretend Netflix does not exist, but nobody has clue in advance just how to decide the percentage ahead of time. (How much we will be streaming the latest Star Wars movie in 2018?) It is fun to look at Moore's Law from the really-big-picture point of view. In fact, not even considering just semiconductor manufacturing. Steve Jurvetson (the Jurvetson in Draper Fisher Jurvetson, one of the top VCs here in the valley) calls this the most important chart in the technology business. It is originally from Ray Kurzweil's book The Age of Spiritual Machines . It shows computing performance from electromechanical to ICs via vacuum tubes (thermionic valves if you are British like me). Although this graph ends in 2000, the trend has continued, but whether you consider multicore to be a continuation or something different depends on what you care about. Single-core performance has stalled, but you can buy a lot of calculations per $1,000, so in that sense the line continues straight (or exponentially since it is a logarithmic scale). (Somewhat off-topic, I highly recommend The Difference Engine by William Gibson and Bruce Sterling (actually pretty much anything by either author is worth a read). It is a science fiction novel that takes as its starting point that electronics never got invented but the mechanical gear driven computing of Babbage's difference engine had been perfected (and the internal combustion engine was not invented but steam engines were perfected). In fact, despite the title, it is really Babbage's analytical engine which is perfected, since it is programmable by punched cards.) If we just look at computer performance, the following few lines from a table that Wikipedia put together shows how far we have come since 1960. (The full table appears on this page .) Year Cost per GFLOPS in 2013 $s Lowest cost platform Notes 1961 $8.3T 17 million IBM 1620 @ $64K each Multiply took 16ms 1984 $43M Cray XMP/48 $15M for 0.8 GFLOPS 1997 $42,000 Bunyip Beowolf Cracked the $1/MFLOPS barrier 2013 $0.22 Sony Playstation $400 for 1.8 TFLOPS 2015 $0.08 Celeron G1830 & Radeon R9 295X2 System Intel CPU+AMD GPU I think we can ignore the first row, that requires 17 million computers of a line that IBM manufactured a total of about 2,000. But the second line is one of the famous line of Cray computers that you could actually go and buy (if you were rich). So since 1984, the cost of a gigaflop of computing has fallen from $43M to 8 cents, a factor of half a billion smaller. I am regularly amused when people say things like "Your smartphone has more computing power than NASA had to go to the moon." It is a true statement, but it is a ridiculous one. NASA had a total computing power of 1 MIPS (and that really is the way you should write it, the S stands for seconds and is not a plural, and MIP on its own doesn't even make sense). Your smartphone has a computing power of 25,000 MIPS (iPhone 6, ymmv). The Saturn launch vehicle used for the Apollo program weighed 6.5M pounds and so 25,000 times less than that is 250 pounds. So the equivalent remark would be that the Saturn rocket weighed more than a person (OK, an overweight one). True, but silly. The worrying thing going forward is that Moore's Law has slowed down not from a technical point of view but from an economic point of view. We know how to build 10nm, 7nm, and probably 5nm. The optimistic view is to look at Kurzwei's graph or the table above and assume it will continue, that we will have another 25,000-fold reduction in the cost of computing in the next 30 years. But the pessimistic view is that we will get faster and better computers but they won't get any cheaper. That might not matter much, most of us would happily shell out $500 for the iPhone 20 running at teraflop rates. Plus most markets are like that without being disasters: cars haven't got a lot cheaper, just a lot better. The truth is probably somewhere in between. Electronics will get cheaper, but the rate at which it gets cheaper will slow down significantly. But there is no doubt that semiconductor processes are getting more and more complicated and expensive and the cost of building a state-of-the-art fab now means that there are only three companies that can do it for logic (Intel, Samsung, and TSMC) and a few more companies that can do it for memory (Micron, SK Hynix, Toshiba/Sandisk). A modern fab depreciates at something like $50/second (maybe only $40 after Intel's recent changes in depreciation), meaning that to afford one you had better be doing business at over $200/second. We live in interesting times (and by the way, the so-called Chinese curse "May you live in interesting times" is unknown in China).![]()
![]()
↧