I've been working hard recently to get back up to speed on the current state of the art in super computing so that I can get a better understanding of the race to Exaflop capable systems. This may sound esoteric but it's important and foundational to understanding many of the developments we'll be seeing over the next several years.
Between 2001 and 2007 I worked at one of the leading High Performance Computing (HPC) centers. I supported the day to day operations and siting of many of the worlds fastest computers including BlueGene/L a system that in various forms held the number one position from September of 2004 until June of 2008 on the top 500 super computers list and is still in the top twenty at number fourteen. That is and was an amazing run that is unlikely to ever be duplicated given how quickly technology is moving forward.
To give an example, another system I worked on was Thunder which was an Intel Itanium based system that first appeared on the top 500 list in November of 2004. Thunder clocked in at 19.94 Teraflops which was good enough for number two. Today that number wouldn't even scrape the very bottom of the list which is held by a Chinese Intel Xeon based system that is over twice as fast.
BG/L at it's peak was bench marked at 478 Teraflops with put it close to half a Petaflop. BG/L is a cousin of the Watson super computer that IBM used to make human beings look mostly obsolete on Jeopardy recently.
The number one computer in the world today is K which runs at 8.2 Petaflops.
A Petaflop is 1024 Teraflops which is 1024 Gigaflops which is 1024 Megaflops which is 1024 Kiloflops which is where this litany will end. A FLOP is a floating point operation. So for instance calculating the answer to
3.4 * 4.4
would be a single FLOP. Floating point is all about the decimal place. A computer capable of doing 1024 such calculations in a second would be a one Kiloflops system. A system capable of doing
1024 * 1024
floating point calculations a second would be a one Megaflops system, and so on. Today's laptops and personal computers are generally capable of several Gigaflops using just their central processors and potentially much more than that if their GPU (Graphic Processing Units) are brought into the picture.
Enough of the boring math. Suffice it to say that hitting the really large numbers we're seeing today requires a lot of computers running in parallel.
So, in June of 2008 BlueGence/L was number one at 478 Teraflops. As of June 2011 the number one system is 8.2 Petaflops, or just over seventeen times faster. That's and impressive leap in three years. The thing is, to reach one Exaflop we'll need a system that is ONE HUNDRED AND TWENTY ONE TIMES FASTER than the current top system.
The three companies most likely to play a big part in getting to that destination in no particular order are AMD, Intel & NVidia. Trying to predict who is going to be the key player in the first Exaflop based system is going to be fun.
Image via Wikipedia
No comments:
Post a Comment