PetaFLOPS performance (1015 Floating Point Operations per Second) has now been achieved and surpassed. On May 25, 2008, the system named Roadrunner from IBM broke the PetaFLOPS performance by reaching 1.026 PetaFLOPS. In November 2009, the Oak Ridge Cray XT system named Jaguar reached 1.75 PetaFLOPS on the Top500 benchmark (HPL).These systems are terribly large, expensive, and power hungry, but the same could be said of TeraFLOPS systems (1012 FLOPS) several years ago. The TeraFLOPS barrier wasn’t broken until June 1997 with ASCI Red. Now you can get almost 3 TeraFLOPS of single precision performance from a single ATI GPU (Graphical Processing Unit) and soon a single GPU from Nvidia will be capable of 0.7 TeraFLOPS of double precision performance. In 12 years the industry has gone from multiple racks to a single processor all with the same level of performance. Performance of processors, networks, etc. is constantly improving so PetaFLOPS may be here faster than you realize.Hams Meuer, one of the founders of the Top500, wrote a paper analyzing the Top500 list and using it to make projections. One of the observations he made was that the last place system on the Top500 will be a PetaFLOPS around 2016. So in 7 years to even make the Top500 list you will need a PetaFLOPS system.
Many people don’t care about making the Top500 list; they just want to get their work done as efficiently as possible. But Dr. Meuer’s analysis shows that all of the fastest systems in the world are likely to be PetaFLOPS systems. Given that people need more computational power, not less, it is very likely that there will be many more than 500 PetaFLOPS systems being used. Consequently, PetaFLOPS will be common place around the year 2016. This is the first in a series of articles discussing PetaFLOPS level systems – where we today, where are we headed, what will PetaFLOPS systems look like in the next few years, what kind of technologies are we likely to see in these systems, and what kind of problems are likely to be encountered.Get ready – PetaFLOPS is coming quickly to everyday clusters.
Next Up ... Part 1 – Who needs PetaFLOPS and why? -- Jeff Layton