On This Page

Von Neumann
GPU & FPGA
Neuromorphic

The Solution:
Don't simulate—emulate!

Large-scale neural simulations are difficult for computers because neural models do not map onto the Von Neumann architecture. While computers operate sequentially, with a few cores executing a programmed set of instructions step-by-step, the brain operates in parallel, with a multitude of neurons processing information distributed throughout a highly interconnected network. The computer can compensate for its lack of parallelism with brute speed, but that comes at a steep cost in energy—putting cortex-scale simulations out of reach for affordable computers.

Jun Makino's GRAPE Faced with the problem of modeling how galaxies evolve, astrophysicists developed their own supercomputer. It is blazingly fast at these simulations—the law of gravity is hardwired into its digital circuits. Its 32 chips perform a trillion arithmetic operations per second—a third as fast as IBM's 2048-processor Blue Gene rack and sixteen times more cost-effective ($42K versus $2M). [D. Normille 2001]

A computer that executes one instruction every time the brain activates a synapse would consume a hundred megawatts!

Alternative hardware solutions are being explored to satisfy the need for affordable simulation platforms. GPUs sporting a hundred and twenty-eight general-purpose processors on a single chip have been programmed to run neural simulations a hundred times faster than a PC. FPGAs sporting a hundred-thousand individually configurable logic gates on a single chip have been configured to run neural simulations two thousand times faster than a PC. Though rivaling a 2048-processor Blue Gene rack, at a fraction of the cost, this performance is several orders of magnitude short of what is needed to simulate multiple cortical areas in real-time.

Whereas simulation refers to software, emulation refes to hardware—a physical realization of a neural model.

Taking a radically different approach, we are building an affordable supercomputer capable of performing cortex-scale simulations by basing its architecture on the brain. This computer's fundamental component is not a logic gate, like in a digital computer, but a silicon neuron whose properties and connectivity are reconfigurable, much like an FPGA’s logic gates. This neuromorphic approach, developed over the past two decades, yields hitherto unimagined levels of efficiency that make beyond-Blue-Gene performance affordable on a Beowulf-cluster budget. It is part of a profound shift in computing, away from the sequential, step-by-step Von Neumann machine towards a parallel, interconnected architecture more like the brain.