PERSPECTIVES Atoms are differentially combined to OPINION produce a spectrum of molecules, which are qualitatively very different from atoms The Blue Brain Project in terms of their properties and the infor- mation they contain. After all, molecules cannot be understood by the study of Henry Markram atoms alone. DNA molecules can be strung Abstract | IBM’s Blue Gene supercomputer allows a quantum leap in the level of together in numerous sequences to produce different genes, which collectively produce detail at which the brain can be modelled. I argue that the time is right to begin hundreds of thousands of proteins that are assimilating the wealth of data that has been accumulated over the past century qualitatively different from their building and start building biologically accurate models of the brain from first principles to blocks. Different combinations of proteins aid our understanding of brain function and dysfunction. produce qualitatively different types of cell that can be combined in various ways in the brain to produce distinct brain regions that Alan Turing (1912–1954) started off by initiative are to simulate the brains of mam- contain and process qualitatively different wanting to “build the brain” and ended up mals with a high level of biological accuracy types of information. The brain seems to with a computer. In the 60 years that have and, ultimately, to study the steps involved in make the next quantum leap in the quality of followed, computation speed has gone the emergence of biological intelligence. intelligence, beyond the physical structures from 1 floating point operation per second to form dynamic electrical ‘molecules’. The (FLOPS) to over 250 trillion — by far the Concepts of intelligence ultimate question, therefore, is whether largest man-made growth rate of any kind IBM built the computer Deep Blue3 to the interaction between neurons drives a in the ~10,000 years of human civilization. compete against and eventually beat Garry series of qualitative leaps in the manner in This is a mere blink of an eye, a single Kasparov at chess, shaking the foundations which information is embodied to represent generation, in the 5 million years of human of our concepts of intelligence. Deep Blue an organism and its world. As computers evolution and billions of years of organic combined conventional methods from approach petaFLOPS speeds, it might now life. What will the future hold — in the computer science, but was able to win by be possible to retrace these elementary steps next 10 years, 100 years, 1,000 years? These brute force, considering 200 million moves in the emergence of biological intelligence immense calcu lation speeds have revolu- per second using if–then-like routines using a detailed, biologically accurate model tionized science, technology and medicine (BOX 1). Nevertheless, this defeat of a human of the brain. in numerous and profound ways. In par- master by a computer on such a complex ticular, it is becoming increasingly possible cognitive task posed the question of whether Detailed models to simulate some of nature’s most intimate the relevant world of an organism could In 1952, Hodgkin and Huxley published processes with exquisite accuracy, from simply be described by enough if–then the highly successful model of ionic cur- atomic reactions to the folding of a single conditions. It could perhaps be argued that rents that allowed simulation of the action protein, gene networks, molecular interac- artificial intelligence, robotics and even the potential4. These simulations revealed the tions, the opening of an ion channel on the most advanced computational neuroscience emergent behaviour of ion channels, and surface of a cell, and the detailed activity approaches that have been used to model showed how only two types of ion chan- of a single neuron. As calculation speeds brain function are merely if–then-like nel can give rise to the action potential approach and go beyond the petaFLOPS conditions in various forms. Adaptation — the currency of the brain. These insights range, it is becoming feasible to make the and learning algorithms have massively fuelled experiments and simulations for next series of quantum leaps to simulating enhanced the power of these systems, but it decades, and now explain how different networks of neurons, brain regions and, could also be claimed that these approaches combinations of ion channels underlie eventually, the whole brain. Turing may, merely enable the system to automatically electrical diversity in the nervous system. after all, have provided the means by which acquire more if–then rules. Regardless of the Wilfred Rall realized that the complexity of to build the brain. complexity of such an operation, the quality the dendritic and axonal arborizations of On 1 July 2005, the Brain Mind Institute of the operation is much the same during neurons would profoundly affect neuronal (BMI, at the Ecole Polytechnique Fédérale de any stage of the computation, and this form processing, and developed cable theory for Lausanne) and IBM (International Business of intelligence could therefore be considered neurons5 despite fierce resistance from the Machines) launched the Blue Brain Project1. as ‘linear intelligence’. entire community, which argued against Using the enormous computing power of From a biological perspective, there are the need to consider such complexity. Rall’s IBM’s prototype Blue Gene/L supercom- quantum leaps in the ‘quality’ of intelligence framework explains the ‘passive’ spatiotem- puter2 (FIG. 1), the aims of this ambitious between different levels of an organism. poral integration in neurons and is key to NATURE REVIEWS | NEUROSCIENCE VOLUME 7 | FEBRUARY 2006 | 153 PERSPECTIVES inhibitory interneurons15. Since then, increas- ing computational capacity has spawned various multi-neuron, multi-compartment cortical, thalamocortical and cerebellar mod- Rack els from many laboratories16–32. The current 1,024 compute nodes state-of-the-art is a model of a thalamocortical Up to 512 GB memory Up to 128 I/O nodes System column comprising 3,650 multi-compartment Up to 5.6 TFLOPS Up to 64 racks neurons (~100 compartments) represent- Up to 65,536 compute ing diverse types, including superficial and nodes with 32 TB memory deep pyramidal neurons, spiny stellates, Compute card (64×32×32 torus) 2 BGL chips Up to 360 TFLOPS fast-spiking interneurons, low-threshold Up to 1 GB memory spiking interneurons, thalamocortical relay (512 MB per node) Node card 32 Up to 11.2 GFLOPS 16 compute cards neurons and reticular nucleus neurons . (32 compute nodes) Traub’s model has given insight into the neu- Up to 16 GB memory ral properties that underlie diverse cortical BGL chip Up to 2 I/O cards (4 I/O nodes) Dual 700 MHz CPUs Up to 180 GFLOPS circuit operations such as gamma oscillations, 4 MB L3 spindles and epileptogenic bursts. Up to 5.6 GFLOPS These studies provide sound proof of Figure 1 | The Blue Gene/L supercomputer architecture. Blue Gene/L is built using system-on-a-chip principle that multi-compartment, multi- technology in which all functions of a node (except for main memory) are integrated onto a single neuron circuit simulations are possible, and application-specific integrated circuit (ASIC). This ASIC includes 2 PowerPC 440 cores running at give valuable insight into cortical network 700 MHz. Associated with each core is a 64-bit ‘double’ floating point unit (FPU) that can operate in properties. The size of the current models single instruction, multiple data (SIMD) mode. Each (single) FPU can execute up to 2 ‘multiply-adds’ seemed a remote prospect in the early days per cycle, which means that the peak performance of the chip is 8 floating point operations per cycle of modelling. They provide a strong founda- (4 under normal conditions, with no use of SIMD mode). This leads to a peak performance of 5.6 billion tion for taking the next quantum step, to floating point operations per second (gigaFLOPS or GFLOPS) per chip or node, or 2.8 GFLOPS in non- further increase the size of the modelled SIMD mode. The two CPUs (central processing units) can be used in ‘co-processor’ mode (resulting in one CPU and 512 MB RAM (random access memory) for computation, the other CPU being used for network to an unprecedented level. processing the I/O (input/output) of the main CPU) or in ‘virtual node’ mode (in which both CPUs with At this point, some may ask, why not use 256 MB each are used for computation). So, the aggregate performance of a processor card in virtual this computing power to simulate cortical node mode is: 2 x node = 2 x 2.8 GFLOPS = 5.6 GFLOPS, and its peak performance (optimal use of circuits with artificial neural networks, in double FPU) is: 2 x 5.6 GFLOPS = 11.2 GFLOPS. A rack (1,024 nodes = 2,048 CPUs) therefore has 2.8 which the entire neuron is represented by one teraFLOPS or TFLOPS, and a peak of 5.6 TFLOPS. The Blue Brain Project’s Blue Gene is a 4-rack system summing node (point neuron), connectiv- that has 4,096 nodes, equal to 8,192 CPUs, with a peak performance of 22.4 TFLOPS. A 64-rack machine ity is simplified to reciprocal interactions should provide 180 TFLOPS, or 360 TFLOPS at peak performance. BGL, Blue Gene/L; torus, torus-like between all nodes, and functional properties connectivity between processors. Modified with permission from IBM (International Business are simplified as ‘integrate and fire’ types of Machines) © (2005) IBM Corporation. activity. Such simulations provide a powerful exploratory tool, but the lack of biological realism severely limits their biological inter- understanding ‘active’ integration due to In the cortex, a pioneering series of simu- pretation. The main problem is that there are nonlinear conductances in dendrites6–10.