11
February
2010
|
05:32 AM
America/Los_Angeles

The Need For A Radical New Type Of Computer Architecture

Irving Wladawsky-Berger is a former chief strategist at IBM and he has also worked as GM of IBM's supercomputer group.

In this post: Extreme Scale Computing he explains the challenges that supercomputers face in reaching the next stage: 1,000 times more powerful than current petaflop (1 followed by 15 zeros) supercomputers.

The challenge in getting to this next stage of 'exascale computing,' is power consumption and heat generation. And that means we will need a new type of architecture.

Massively parallel architectures, using tens to hundreds of thousands of processors from the PC and Unix markets have dominated supercomputing over the past twenty years. They got us into the terascale and petascale ranges. But, they will not get us to exascale.

Mr Wladawsky-Berger says:

"Another massive technology and architectural transition now looms for supercomputing and the IT industry in general."

The reason that this is not just a supercomputer problem, but also one for the IT industry, is because of the rise in importance of cloud computing.

The technology requirements are quite similar, especially the need for low power, low cost components. They also share similar requirements for highly efficient, autonomic system management. One can actually view cloud-based systems as a kind of exascale class supercomputers designed to support embarrassingly parallel workloads, such as massive information analysis or huge numbers of sensors and mobile devices.

If we can build exascale computing supercomputers, we will also be able to build smaller systems that use the same basic technologies and that will have many business applications, not just in cloud computing.

But there is a big problem: the power consumed, and the heat generated by computing components -- tremendous numbers of computing components.

...the only way to now increase performance toward an exascale system is massive parallelism, an exaflop supercomputer might have 100s of millions of processing elements or cores. Such massive parallelism will require major innovations in the architecture, software and applications for exascale systems.

A future exascale computer will likely make use of low power consuming processors used in consumer mobile devices. (Possibly chips such as Apple's A4 chip, which offers ten hours of battery life in the iPad.)

If we can crack the formidable challenges of processors, software architectures, and how to develop applications designed for extreme parallel processing, we will be very close to cracking some very large computational problems that currently are out of reach, from "economics and medicine to business and government."

Fortunately, the Obama administration has recognized the need for exascale computing and has listed it in its Strategy for American Innovation.

You can read the full article here.