IBM Simulates 500 Billion Neurons and 100 Trillion Synapses

In a neuronal simulation unprecedented IBM has managed to simulate 500 billion neurons and 100 trillion synapses . It has used Sequoia which is the second largest supercomputer in the world and half million cores. This is a computational feat but has little to do with neuroscience. Let’s see why.

The Department of Cognitive Computing at IBM Almaden led by Dharmendra S. Modha takes a few years doing amazing simulations in the context of the DARPA Synapse. As part of this project, announced the simulation at the scale of the cortex of a mouse, a rat and then later a cat .

The objective of this program is to create a chip neurosinaptico that marks a break with the traditional architecture of computers. This architecture is called the Von Neumann using all the computers today, including mobile and cards. In the von Neumann architecture is separate from the processor memory and the hardware and the software programs are separated from the data. It has been very successful as they meet the miniaturization of components expressed in Moore’s law : every two years, doubling the number of transistors in a given space. The problem is that we are reaching the limits of the atom and that the law will be fulfilled .

Neurosinaptico the chip is a complete break with the Von Neumann architecture. Design is based on the neurons in which there is no distinction between hardware and software, programs and data, memory and processor . The chip consists of a matrix of neurons and their crosses are made between synapses. Thus, each synapse chip is hw and sw, processing and memory, and program data. Because everything is distributed, it is not necessary one so extreme miniaturization and above all a fast clock. Gigahertz from the current frequency to neurons fire one hertz , and in the case of the chip to 8 hertz. In addition, processors are clock driven , ie they act under the direction of the clock while neurons are event driven , act only if there is activity to perform.

One goal is to reduce power consumption. A brain which consumes a small light bulb, 20 watts. A supercomputer consumes hundreds of megawatts. The new chip has a very low consumption. These chips are built with classical silicon CMOS technology.

The architecture of many chips neurosynaptic States has called TrueNorth . There is already developing a chip of 256 neurons, axons 1024, and 256 × 1024 synapses.

The chip is not yet in mass production. To continue to work in parallel with the development, testing has been done today. For this we have used a simulator called Compass. Compass translates the behavior of a chip neurosinaptico (not Von Neumann) to a classical computer (von Neumann). Using Compass has been simulated ( pdf ) the behavior of two billion chips. This means 500 billion neurons and 100 trillion synapses, completely astronomical figures. The result of the simulation was run 1542 times slower than real time.

For the simulation has been used the second world’s largest supercomputer, Sequoia a Blue Gene / Q of 96 cabinets with one and a half million cores and 1.5 petabytes of memory. One objective of the simulation is to see the scaling. A common problem is that when we add more cores, the system does not work proportionately faster. In the end, adding more cores does not increase the performance: the evil scale system . Imagine a waiter serving behind the bar. If a second waiter, go faster, but not twice. If you keep adding waiters come a time that does not increase efficiency even be reduced. The evil scale system. Well, in the simulation on has been almost perfect scaling which is computationally very satisfying.

What has this to do with neuroscience and brain? Very little. The simulation does not imitate any animal behavior or cognitive or human. To simulate the behavior of the brain, we need to know how it works and that is far from achieved. For when that knowledge must be ready computers to be able to simulate it and in this context that the present investigation. Although not alone, the idea of designing these new chips is put into production in traditional business applications to give high power with low consumption. brain simulation must wait even a decade.