Cheap, Fast, Good

Pick Two

So why can't I have Server Sky Right Now?

Here's the infamous Project Management Triangle. You cannot have something good that is both cheap and fast to develop.


As soon as a potential project meets two of the criteria, one of the seven billion of us is likely to commence doing it. If a potential project only meets one of the criteria, it will wait behind other "two criterion" projects, some of which will improve the technological environment, making cheaper, faster, or better more likely. When it is Railroading Time, there will be a rush to build railroads.

I am not wealthy. And I am committed to doing server sky properly. So I can do cheap, and I can do good, but that means Server Sky won't happen fast. If you are rich and in a hurry, and are willing to do things the right way, let's talk.

The Inevitability of Server Sky

or something like it

Something like server sky is inevitable. Corporations like Intel are committed to Moore's Law. It is in their DNA. The result of Moore's Law is 200 times more transistors per decade, 4.5% more per month. Intel is composed of tenacious and clever people, who will keep turning the crank until somebody else does it better. And they have the technology, either in hand or in the lab. They have many brilliant device designers working on dozens of possible paths towards the devices of a decade from now - the question is not if there is a path to take, but which path will lead to the best result. As soon as they slow down and try to rest on their past accomplishments, some other company will take the lead and leave them in the dust. There is a lot of room to improve.

But powering 40,000 times more transistors two decades from now, or 500 billion times more transistors five decades from now, may be impossible within the constraints of temperature and available power on the Earth's surface.

Transistors can run slower and power starved - this is the new direction Moore's Law is taking, with slow and intermittent structural logic replacing software, idle until demanded by the task. For example, a video codec can be implemented in software, with instruction fetches and speculative branching and all sorts of clever enhancements to get it running fast enough, but this is incredibly power hungry. The same video codec can be impemented as gates, with fixed wiring replacing agile instruction flows; that reduces flexibility, but the hardware approach is much more energy efficient, and that hardware can be shut down and approach zero power consumption when it is not needed. Indeed, the power consumption can be throttled by agile adjustment of clock rate and voltage supply, or reduced by parallelism, to very low maximum values.

But not zero. At some point, thermal noise restricts the minimum power of a very small logic gate. A flipflop can be run very slowly, with very low clock rate, but its operating speed is a function of its "inertia" (capacitance), and very small things have very low inertia. Small things go fast, that is inherent in their nature. And that means a very brief thermal excursion can have sufficient energy to change the state of a bit storage element. Future logic systems designed for maximum speed-power efficiency will be scaled towards this threshold, and will have error detection and correction circuitry to compensate, but at the bottom is the hard wall of the Shannon Limit - the production of one-way computation, propagated out into the real world of results, means that useful results will always have an energy cost associated with them, and that energy cost cannot scale below limits imposed by a real world of atoms and heat and the speed of light, given human associated limits of resources and patience and quality. Cheap, fast, good still applies.

So, after an enormous amount of cleverness, and a torrent of fantastically cheap digital logic from an ever-more-productive "transistor hose", we hit the wall, where all our cleverness produces a multi-terawatt power bill.

In spite of a lot of fancy accounting shenanigans by so called "green power" advocates, that power is robbed from natural systems. We can hope that we can use all that computation to manage nature better than nature manages itself, and use the extracted surplus to power more computation. But our track record so far isn't good. We build wind turbines out of CO_2_ intensive steel, and anchor them in CO_2_ intensive concrete, and their highly intermittent power is leveled by intermittent gushes of salmon-destroying hydropower (from dams made of more steel and concrete). The proud green consumers check a box on their power bill and pay a little extra for "green energy", when in fact no amount of extra payment can compensate for the real damage their ignorant pride does to the real world of power lines, river flows, and a natural environment too complex to be reduced to simple slogans.

Above our heads, our sun puts out trillions of terawatts, high quality low entropy photons that stream into empty space, never to touch matter (much less benefit life) ever, ever again. A gram of indium phosphide, thinned to 200 nanometers, spreads out to a square meter, intercepts 1360 watts, and can turn that into 200 watts of electrical power. Even with expensive rockets, it costs $10 to launch that gram, and an electromagnetic launcher could put it into space for a small fraction of a penny. There's more than enough power in space to power the exponential growth of transistors, for many decades; far from the sun, the logic temperature can be reduced and a watt of high quality photons can be transformed into logical operations. And there's a big galaxy out there - lots of power available if we are flexible in materials and methods, and patient for the results.

Why We Should Not Presume Another Apollo

Cheap, fast, good followed us into space. Space advocates have spent decades hoping for an escape from the iron laws of project managment - they will die of disappointment and old age before the laws do.

The Apollo program, the enduring example of "why can't we ..." applied by space advocates to space settlement, manned expeditions to Mars, and other goals, is a perfect example of those laws in action. Kennedy made his moon speech in 1961 - Apollo 11 landed on the moon in 1969, less than 100 months later. Kennedy did not make his speech in a political or economic vacuum. Space was one of the largest pork-barrel programs in US history (quick returns), with a string of frequent publicity events that sustained interest in continued spending. At no point in the program was there a "publicity gap" as long in time as a manned expedition to Mars, or the construction of a large space habitat, would require. The geopolitical reality was that the space program was a proxy demonstration of nuclear war fighting ability; a nation that could launch a Saturn V with the precision in time and space to land astronauts and return them to earth could certainly deliver enough warheads over the pole to obliterate an opponent.

The "proxy demonstration" aspect follows the space program to the present day. "Successful" war is the precision delivery of destruction. The quality and accuracy required to deliver the Mars Science Laboratory ("Curiosity") reflects the US capability to deliver exactly the amount of energy to disable an opponent anywhere on Earth. MSL's multinational character demonstrates that the US can work productively with a lot of friends, and welcomes more. MSL was far cheaper than Apollo, with comparable geopolitical benefits.

Apollo was not cheap, but it happened because it was good (politically, domestically and internationally) and it was fast (less than two terms in the Senate). Politicians voting for Apollo could expect positive political returns before their next run for office (fast). Taxpayers got national pride with none of the risks of war (good). It wasn't cheap, but it wasn't as expensive as the global nuclear war for which it was a peaceful proxy. And it was not nearly as expensive as the multi-trillion dollar development of nuclear weapons and delivery vehicles - in that sense, it provided a desirable multiplication of military effectiveness less expensive than expanding the strategic arsenal.

The trend is clear - political funding of space will find ever-cheaper ways to deliver the same political benefits. We cannot expect a nearly-bankrupt US government to make larger expenditures of funds for space in the future; in fact, we are wise to assume that Western politicians will never be that stupid. However, nations with something to demonstrate and national cohesion to build may be wise to make prudent investments in space programs. The space race to Mars will probably be between China and India, when they become wealthy enough to pay for the enormous effort, and the result will be political dominance in Asia and the developing world.

Like the war it thankfully replaces, space is diplomacy by other means. Which, if the result is avoided wars and saved lives, is worth a hell of a lot of money. The scientific results are an added bonus. Win, win, WIN!

Making Server Sky Happen

Global internet connectivity with Server Sky will provide a quicker and more tangible return on investment than a manned trip to Mars. India is beset with corruption, and is structured to be politically hostile to infrastructure; New Delhi cannot deliver clean water and 24x7 electricity to all its citizens. Mumbai imposes absurd height restrictions on itself, resulting in impoverished citizens paying more per square foot for living space than anywhere else in the world. Although India has made great strides in abolishing the License Raj, it is a long way from a free economy - the ghost of Fabian socialism and Gandhian ruralism still haunts academia, and iron ricebowl corruption still pervades the bureacracy.

Meanwhile, a new generation of bright, college-trained Indian technologists are mostly wasted in call centers connected over the Internet to the western world. Indian companies like Wipro, and transnationals like Microsoft and Intel, profitably employ a small fraction of these technologists designing software and chips for the Wealthy West, but far more young technologists are available, far more can be trained, and a far larger portion of the developing world could profitably use those services if delivered cheaply and widely enough.

Delivery of services over the internet - human talent multiplied by software - can increase the effective wealth of the third world, providing information abundance in spite of continuing material scarcity. Perhaps we cannot afford to build schools and hospitals and roads and utilities for seven billion people. But cheap information can mostly substitute for all of these things. Even clean water and sewage treatment, because pathogens are fundamentally biological information malware, and can be fought with better biological information ( tailored phages against bacteria and parasites, for example). Bits can substitute for atoms, and terabits can be more effective than tons of steel.

Good, yes. Cheap and fast as well? Er, no. We can make small and relatively cheap experiments, even a few in space, to reduce the risks to deployment of Server Sky. But it will take time to do this right, and to avoid path-dependent traps like those that lock us into fossil fuel dependency. The cheapest way to avoid these traps in 2014 involve nothing more than library research and desktop computation and creative-commons publicity such as this.

It is far too expensive to recruit the world in this effort, nor is it necessary. Smart people will find clever ways to implement the technology, or discover subtle flaws that could derail the project or produce damaging unintended consequences. They will find these pages in time, because they will discover a solution needing applications, or a natural phenomena whose consequences they want to explore. Those discoverers will connect into coalitions, who will build the assemblies that assemble into Server Sky.

When the time is right for a large investment to yield profit quickly, smart investors will grab the brass ring before others notice it. I would like to believe that there are investors out there clever enough to organize this effort in 2014, but Moore's Law will make the task 70% easier next year. The choice between first mover advantage and payback time may favor waiting a year or two or even five, but not as long as a decade, when catching up is too expensive. I hope I am around to help when perception moves from possibility to inevitability.

CheapFastGood (last edited 2014-02-07 19:36:28 by KeithLofstrom)