Super fast -- 06/12/08

On Monday IBM announced that it has built a new supercomputer -- one that has achieved the highest performance level yet, one petaflop.

Not that long ago -- or so it seems -- when I was a computer operator -- but somehow three decades have slipped past since those days -- I was working in the computer center at a state university. When I had started the job we were running an IBM 370/158 mainframe.

Then we switched to an Amdahl machine. It was the equivalent of the highest IBM 370 machine, except it was cheaper to run and was also provided at an extremely attractive price (probably because they wanted to be able to brag to potential customers about replacing an IBM machine with an Amdahl in IBM's own backyard, this being less than eight miles from where IBM had been founded).

If I recall correctly, our Amdahl machine was rated at 2.7 MIPS.
  Gene Amdahl is one of those quirky computer geniuses. He came out of a tiny country town in South Dakota to get a PhD in theoretical physics, and then go directly to work for IBM. He had a dynamic (although stormy) career at IBM, but then quit when IBM rejected some of his ideas. He founded his own computer company -- Amdahl Corporation -- with significant financial and manufacturing assistance from Fujitsu -- and competed against IBM with a "plug-compatible" computer system. That is, you could replace your IBM mainframe with an Amdahl mainframe, plug in all of the same printers and tape drives, etc. and turn it on and run the very same software. We may be accustomed to that kind of thing today with our PCs but believe me, at the time it was amazing.

MIPs = million instructions per second. More modern statistics are measured in terms of floating point arithmetic operations executed per second ("flop") because modern machines have multiple processors (mainframes and super computers and, these days, even in higher end PCs) and thus more than one instruction can be completed per clock cycle.

Ten years ago the fastest supercomputer in the world was capable of one teraflop performance. That is, it could perform one trillion floating point arithmetic instructions per second. That's one thousand times one billion.

The new IBM Roadrunner supercomputer is one thousand times faster than that. One petaflop is one thousand times one teraflop -- i.e., one thosand times one trillion. (Yes, the Roadrunner name is a direct allusion to a certain very speedy cartoon character.)

IBM's Roadrunner -- which has 80 terabytes of memory (i.e., 80,000 gigabytes) -- is made of 12,960 IBM Cell chips and 6,948 dual-core AMD Opteron chips. If you've seen any news stories about this they probably described the IBM Cell chips as being just like the ones that power Sony's PlayStation 3 game machine. Well, not quite. It is a Cell chip, but it is a bit souped-up, providing double-precision calculations and is closer to the Cell chip used in IBM Blade servers.

A petaflop is on quadrillion calculations per second. The next thousandfold goal is the exaflop, which is a quintillion calculations per second, followed by the zettaflop, the yottaflop and the xeraflop. The terminology sounds funny, but I'm sure an exaflop supercomputer will be along eventually... and by then our PCs will probably have a thousand times the performance that they have today.

Yesterday at lunch time I dashed over to the Red Cross Blood Center in Narragansett for an appointment to donate blood.

I am not the easiest subject for this. My veins tend to be "slippery" and to "roll" -- well, at least that is how the plebotomists describe it -- and sometimes the needle goes in just perfectly on the very first try... and sometimes they have to try... and try... and then eventually switch to the other arm (doesn't seem to matter which one they start with). Sometimes they laugh when they see me come in. My difficult veins are the stuff of legend. And yet sometimes -- probably almost half the time -- everything goes smoothly on the very first try -- but sometimes I end up with bandages on both arms by the time they are able to get a successful flow going.

They have some kind of semi-automated process called apheresis that is the automated collection of a blood products from a donor -- but they have told me repeatedly that I would not be a good subject for that process, that I should stick to plain old-fashioned whole blood donation. Okay, whatever. But yesterday I was being processed by a woman who thought that a particular vein on my left arm would work fine with the machine.

Much to my amazement -- and to the amazement of many of the other staff members -- she got the needle in perfectly on her first try and the whole thing went smoothly. The machine processed my blood, removed red blood cells and returned my plasma.

It takes at least twice as long as a regular whole blood donation. After a while I was kind of bored and I realized that I had my digital camera in a pocket (wearing cargo style shorts, the kind with lots of pockets), so I pulled it out and took a photo of the process, holding the camera one-handed and off at an angle, just sort of a guess that I was pointing in the right direction.

Even though this process does not leave you as fluid-depleted as a regular donation, it does take twice as many red blood cells, so the interval before you can donate again is also doubled -- to sixteen weeks instead of eight -- so my next appointment is not until October 1st.

previous entry

next entry

To list of entries for 2008

To Home (Index) page


Weblog Commenting and Trackback by HaloScan.com