First is the possibility of a desk-top supercomputer in a few years. A company called Optalysys says they will soon be releasing a first generation chip set and desk-top size computer that will be able to run at a speed of 346 gigaflops in the first generation. A flop is a measure of instructions per second that can be performed by a computer. A gigaflop is 109 instructions, a petaflop is 1015 instructions and an exaflop is 1018. The fastest supercomputer today is the Tinahhe-2, built by a Chinese university and which operates at 34 petaflops, which is obviously much faster than this first desktop machine.
The computer works by beaming low-intensity lasers through layers of liquid crystal. They say that in upcoming generations that they will have a machine that can do 9 petaflops by 2017 and they have a goal of having a machine that will do 17.1 exaflops (17,100 petaflops) by 2020. The 2017 version will be half as fast as the fastest supercomputer today and yet be far smaller and use far less power. This would make it possible for many more companies and universities to own a supercomputer. And if they really can achieve their goal by 2020 it means another big leap forward in supercomputing power since that machine would be several magnitudes faster than the Chinese machine today. This is exciting news because in the future there are going to be mountains of data to be analyzed and it’s going to take myriad, and affordable supercomputing to keep up with the demands of big data.
In a somewhat related, but very different approach, IBM has announced that it has developed a chip that mimics the way the human brain works. They have developed a chip they call TrueNorth that contains the equivalent of one million human neurons and 256 million synapses.
The IBM chip is a totally different approach to computing. The human brain stores memories and does computing within the same neural network and this chip does the same thing. IBM has been able to create what they call spiking neurons within the chip, which means that the chip can store data as a pattern of pulses much in the same way the brain does. This is a fundamentally different approach than traditional computers that use what is called Von Neumann computing that separates data and computing. One of the problems with traditional computing is that data has to be moved back and forth to be processes, meaning that normal computers don’t do anything in real time and there are often data bottlenecks.
The IBM TrueNorth chip, even in this first generation is able to process things in real time. Early work on the chip has shown that it can do things like recognize images in real time both faster and with far less power than traditional computers. IBM doesn’t claim that this particular chip is ready to put into products and they see it as the first prototype for testing this new method of computing. It’s even possible that this might be a dead-end in terms of commercial applications, although IBM already sees possibilities for this kind of computer to be used for both real time and graphics applications.
This chip was designed as part of a DARPA program called SyNAPSE, which is short for Systems of Neuromorphic Adaptive Plastic Scalable Electronics, which is an effort to create a brain-like hardware. The end game of that program is to eventually design a computer that can learn, and this first IBM chip is a long way from that end game. And of course, anybody who has seen the Terminator movies knows that DARPA is shooting to develop a benign version of Skynet!