While it’s always fun to compare the speed of modern day desktop computers to, say, those from 1995, it’s even funner with supercomputers. Supercomputers are only super for so long, and after a certain point, your everyday computer becomes much more powerful than older supercomputers. Computers also tend to have hardware problems as they age, and the most recent software they can effectively run after a point becomes obsolete.
In this post, I’m going to be comparing the amount of FLOPS (FLoating point operations per second, a general measure of processor “speed”) in a modern day computer processor – specifically, the Intel Core 2 Duo E6600, our reference processor vs FLOPS in various supercomputers over time. Our reference processor performs at a theoretical max of 38 GFLOPS (38 billion FLOPS.)
The Z1, designed and built by Konrad Zuse between 1935 and 1938, a mechanical machine, and arguably the granddaddy of all modern computers, performed at a measly 1 FLOPS. Our reference CPU is 38 billion times faster than this. A simple, handheld, dollar store calculator beats this by 9 FLOPS (since it’s doing relatively simple calculations at the slow rate humans can enter in problems, it doesn’t need to be any faster.)
Next, we have the first real supercomputer – the Colossus 1, introduced in 1943 and specifically designed by the British during WW2 to break encrypted German messages. It performed at 5 KFLOPS, about 5000 times faster than the Z1. Still, our reference processor beats it 7.6 million times in speed. Though, back then, the idea of a supercomputer was practically non-existent, since there were no desktop computers to compare to, and this particular computer didn’t have much before it.
Now we jump 11 years into the future (1954) and we have the IBM NORC, the first supercomputer to actually claim that it’s a supercomputer, stating that it’s purpose was to surpass in speed all other computers existent at the time. It was built for the US Navy’s Bureau of Ordnance. It performed at 67 KFLOPS, which our reference processor beats 567,164 times over.
Only 10 years later, the newest supercomputer already outperforms the NORC 44 times over, the CDC 6600, a pretty significant supercomputer in it’s time. It’s widely considered to be the first successful supercomputer, outperforming it’s competitors by about 3 times. It remained fastest until 1969, when it was outclassed only by it’s successor, the CDC 7600. Our processor is 12,667 times faster than the CDC 6600.
In 1974, we again revisit CDC, who have just unveiled their STAR-100 system, performing at 100 MFLOPS, much faster than the CDC 7600, at 36 MFLOPS. Unfortunately for CDC, several poor design mistakes in this supercomputer caused it’s performance to be much lower than expected, and ultimately one of the reasons CDC was pushed from it’s dominance of the supercomputer market. On the plus side, it was one of the first supercomputers to use a vector processor for improved math performance. Our reference processor performs 380 times faster than this machine.
By 1985, Cray has come as a leader in the market, releasing the Cray-2/8. When released, it was the fastest supercomputer in the world, surpassing the Cray X-MP. It performed at a max of 3.9 GFLOPS, being beaten by our reference processor 9.7 times.
At this point the world’s supercomputers are catching up with our reference processor’s speed, in this case the NEC SX-3/44R is 23.2 GFLOPS, and the reference processor only beats it 1.6 times over. In 1993, our processor is finally surpassed with the release of the Thinking Machines CM-5/1024, which ran at 59.7 GFLOPS. The Core 2 Duo E6600 is only 63% of this performance.
In the end, FLOPS are only a general idea of performance. The odds are a modern machine with our reference processor could beat the CM-5/1024, because of much quicker speed and design in other areas (I.e, RAM, I/O, etc.)
And just for laughs, I’ll add some statistics about the cost per GFLOP per year: in 1961, this was $1.1 trillion, in 1984, $15 million, in 1997, $30,000, and today, only 13 cents.
