News for hackers. Stuff that matters.
Re:IBM announces advances toward a computer that works like ... by a-priori on Wednesday November 18, 2009 @02:03PM (#948864)
Right now, it's possible to simulate one biologically-accurate neuron at a compartment-level (where you model a neuron by dividing it into many connected compartments), on one CPU core in real-time.

Here's the kicker: it's also been shown that these simulations can scale linearly with the number of processors (given sufficient interconnect bandwidth). So with N cores, you can simulate kN neurons for some constant k (and I think that k=1).

This implies that the limit is the number of CPU cores you throw at the problem. The human brain has around 100 billion neurons so, in theory, if you had a 100-billion core supercomputer, you could simulate the whole thing. That's a lot, but my point is that scaling a simulation is an engineering problem.

This doesn't say anything about the accuracy of the model; that depends on accurate characterisations of each part of the neuron, and of how they're arranged and connected. But the key point here is that these are low-level characteristics that can be individually examined, making whole-brain simulations a tractable problem.

Source: Djurfeldt et al. Brain-scale simulation of the neocortex on the IBM Blue Gene/L supercomputer. IBM JOURNAL OF RESEARCH AND DEVELOPMENT (2008)

There are no comments on this comment yet.