Peter Jin (cuw@rrpf.orexryrl.rqh)
I’m a first year PhD grad student in Kurt Keutzer’s group in the EECS department. I’m broadly interested in parallelism, GPUs, and energy. I like C, CUDA, and LLVM. As an undergrad, I was a physics major, and I still have some residual interest in biophysics and neuroscience. I once worked on the slime mold bacteria Myxococcus xanthus, and I also have experience with visual stimulus experiments on the salamander retina. My favorite color is orange.
The Blue Brain Project is a large computational neuroscience project at EPFL. Its most ambitious goal is the large-scale biophysical modeling of cortical columns in the mammalian brain.
Blue Brain performs simulations using the NEURON software developed at Yale, which models neuron dynamics with Hodgkin-Huxley-type differential equations. The computing platform is an IBM Blue Gene/P supercomputer with 4 racks, totaling 4096 nodes and 16384 cores.
The biophysical basis of neuron simulations is the Hodgkin-Huxley model, which gives an equivalent RC circuit for the neuronal membrane system, consisting of a lipid bilayer cell membrane (capacitance) populated with voltage-gated ion channels (conductances and reversal potentials). It is a differential equation of the membrane potential, and can be thought of as a mean field model, since the physical events of ions entering the cell are in fact discrete and finite. (Synaptic transmission relies on ligand-gated ion channels, which are not explicitly part of the Hodgkin-Huxley equation but can be modeled by hand by manipulating the membrane current term.)
The Hodgkin-Huxley model is important because it correctly models the formation of action potentials (spikes) when the membrane potential rises above a certain threshold. In many neurons, the resting membrane potential is at -65 mV, while the spiking threshold is around -40 mV (depolarized). During a spike, the membrane potential quickly reaches a peak of +20 mV within a millisecond; the fast depolarization is due to rapidly acting voltage-gated sodium ion channels in the neuronal soma and axon. The spike just as rapidly falls, until the membrane potential is around -80 mV (hyperpolarized). The hyperpolarization is due to slower acting voltage-gated potassium ion channels, which activate after the sodium channels.
While the Hodgkin-Huxley model predicts the generation of spikes, cable theory predicts the transmission of spikes along the neuronal axon. Cable theory models are usually partial difference equations in time and spatial dimensions. They also employ compartmental models, which account for the specific cell morphology of the neuron. The Blue Brain Project's neuron model has as many as 10,000 compartments per neuron, with over a dozen ion channels per compartment.
In mammals, there is a distinct ordering of neocortical neurons into columnar structures, with as many as 100,000 neurons per column. The existence of a putative repeating unit of cortical computation makes the cortical column an attractive target for modeling in computational neuroscience. Thus, given an idealized neural circuit for the cortical column, and connecting many of these units together, we end up with a large system of coupled ordinary and partial differential equations. This is essentially the heart of the Blue Brain Project.
The Blue Brain Project uses the NEURON software accelerated with MPI, as well as the NeoCortical Simulator (NCS) software for managing communication between NEURON instances. Simulations are written as configuration or markup files (e.g., in XML) which are parsed and run by the computational engine.
The Blue Gene/P is a distributed memory computer. Its processor cores are a single kind of ASIC, called the “BG/P compute chip” (BPC). There are two distinct nodes, compute nodes and I/O nodes, both composed of the same kind of BPC cores. Compute nodes are connected in a 3D torus topology, and each compute node is connected to a configurable network of I/O nodes. Each compute node is capable of 13.6 Gflops/node; the Blue Gene/P at EPFL thus has a peak performance of 55 Tflops. A 72-rack Blue Gene/P, the JUGENE at FZJ, was ranked 4th on the Top 500 in November 2009, clocking in at 1 Pflops peak performance. Another Blue Gene/P, the Intrepid at Argonne National Lab, was ranked 8th on the same list, at 557 Tflops.
The Blue Gene/P can simulate an entire cortical column of 100,000 neurons at a detailed biophysical level, or it can perform a coarser-grained simulation of 100 million neurons. Using the aforementioned software, each BPC core can simulate 10 detailed neurons, so each neuron is being simulated at most at 1.4 Gflops/neuron. Without a more detailed knowledge of the algorithms employed or the design space explored (i.e., by reading the code), it is hard to say how well or poorly the Blue Brain Project scales.
[1] Henry Markram. “The Blue Brain Project.” Nature Reviews Neuroscience 7 (2006), pp. 153-160.
[2] IBM Blue Gene Team. “Overview of the IBM Blue Gene/P project.” IBM Journal of Research and Development 52/1-2 (2008), pp. 199-220.
[3] Top 500, November 2009 (http://www.top500.org/lists/2009/11/).