IBM presents processor that works like a brain

Hardware manufacturer IBM has built a processor that functions much like the brain. The neural processor works fundamentally differently from current processors and is much more economical than current processors for specialized tasks.

The TrueNorth neural processor has 4096 cores, made on Samsung’s 28nm process. In total, the processor has 5.4 billion transistors. Each processor core has more than 100,000 bits of its own memory to store its state and which connections it can make.

The difference with conventional processors is that the information does not come from the binary display of the on or off of the transistors, but from spikes, where the information is contained in the timing and frequency of these spikes. In addition, the ‘neurons’ in the processor can send those spikes to and receive from 256 other ‘neurons’; this is not possible with conventional processors.

This different architecture makes the processor much more economical when performing certain tasks, such as object recognition in video images. The researchers from IBM and Cornell University say in the scientific journal Science that when analyzing video images of 240×400 pixels at 30fps, the chip consumes 65mW; Current processors require much more energy for this.

IBM has been involved with neural processors for years. It says it started six years ago and already presented a software ecosystem for these processors last year, because current software does not work on them.