That is a 9 followed by 18 zeros. A 3.4 GHz Intel Pentium 4 "Prescott" CPU would require 83 years, 10 months, and 18 days (including leap years) to accumulate that many clock cycles, much less do anything useful with them. We were growing alpha-quartz atom by atom and periodically doing monstrous moment of inertia matrix inversions, all in 80-bit long_double_precision. Add a blizzard of arithmetic and square roots. We were progressively creating 760 single crystal solid spheres from 8 angstroms to 0.6 millimeters in diameter as a test of arcane geometric theory calculating the handedness of things atom by atom.
We had 30 days. The work was to be/has been presented on 01 May 2004 at the American Physical Society national meeting. Would a brief scribble of equation be borne out by nine quintillion atoms? 80-bit long_double_precision gave us 9.223 quintillion by the book. Calculations needed a few decimal places' elbow room. We could count high enough, but could we create data?
If ab initio theory and real world quartz snuggled tightly, there was a test on a lab bench that ccould disprove the founding postulate of General Relativity and validate an entirely different class of gravitation theory. Admittedly only six such lab benches exist on Earth, but one of them is at UC Irvine. The argument in favor of doing the experiment faced "Not Invented Here." The keeper of the UCI apparatus was not keen to be convinced.
Little spheres were simple enough. Software ran in fat, slow, buggy, loathsome Wincrap. The first trillion atoms and their numeric analyses oozed through in a week. Data walked all over the predicted line. Nine quintillion atoms and their statistically robust conclusion would require about 172,489 years. Bill Gates' little joke upon humanity was not up to it, nor was Intel's. Pentiums sacrifice throughput for maximum clock speed. Marketing can run with big gigahertz.
Linux stomps Wincrap. Uncle Al with brain in hand surfed Usenet Linux newsgroups to see what was shaking. Time on a 3.2 GHz Intel Xeon box was volunteered and a quadrillion atom run commenced. Time on a 2.8 GHz AMD Opteron box was volunteered the next day. It would be interesting to compare results to 18 decimal places. (They were identical.)
The Xeon needed seven days, the Opteron only five. Clock cycle for cycle, AMD pounded Intel's butt with 60% more throughput. Nine quintillion atoms was only 123 years away in premium hardware. The bigger the radius the slower the calculation, as the square.
Ordinary folk confronted by the seeming impossibility of the endeavor would embrace the rational decision and quit. The Severely Gifted are... different. Intelligence is prowess at rendering novel tasks. The UK programmer's box held two Opteron-244s. AMD is the planetary champion of CISC parallel execution. Could 583 lines of hellishly looped source code be parallelized?
No way! Never. Uh-uh. Couldn't happen. The guy spasmed and churned. The code was woven to a Gordian knot, interpenetrating itself. He tried for three weeks and nothing came of it. What the heck, make the rational decision. Wake up the next morning with the solution. I love genius!
Take 30 days, add two Opteron-244s plus an AMD Athlon to handle system overhead, get 43 quadrillion atoms. 123 years of 9 quintillion runtime shrank to 17 years. Parallel execution is good!
Who has a really big pile of unused AMD Opterons on hand? AMD came to mind. They had conveniently turned over excess inventory by ganging 256 Opterons into the 247th fastest computer on Earth. Nobel Laureate Richard Feynman opined that to secure some nooky on a date, one has to ask. We exercised the remarkable gall to ask for free time on a teraFLOPS computer.
No way! AMD would not donate more than 30 days runtime in two Opteron-848 octapods. That was 16 bleeding edge technology maxed-out CPUs locked in coherent HyperTransport, equivalent to about 700 Intel Pentium 4s running individually. Intel chips could not be efficiently clustered for a number of painstakingly dreadful reasons. Second base is not bad for a first date!
Ten days into the 30 day run we had 0.343 quintillion atoms. The program then shifted gears per specification, spent five days getting its petticoats in order, and began spitting output at the last and largest diameter increments. Against all hopes and expectations, our monstrous Gordian knot CHIpir software accelerated.
In the end we counted to 1.9455 quintillion atoms in 30 days, but meaningful output died at 444 quadrillion atoms. A real world quartz ball 0.22 millimeters in diameter was entirely satisfactory for the presentation. Data snugged theory.
Physics poo-pooed it. Uncle Al has poop-ooed back wth two papers submitted for publication. The first paper, on the pure math, is accepted. ACS, CAS, J. Am. Chem. Soc., NIST, IUPAC, Angew. Chem. each and all could not assign a formal name to [6.6]chiralane, a C27H28 alkane (it is polycyclic) that is perfectly chiral. That was once called "discovery". Now they tell me to go shove it - my own people. Civilization deserves to die.
The physics paper is fun! The journal is having problems with Referees. I suggested eight candidates. The crystallographer said "A-OK!" No physicist denies the physics. No physicist understands the chemistry. Everybody is outraged that the two might be combined - but they can't find an error (other than political errors).
It does not need eight pages and a full page of references. 50 words suffice:
NO "Intel Inside." If Intel takes this poorly, Intel is invited to volunteer time in its 8-core wonder iron. AMD is good people.
