July 07, 2011
Under the category of "Grand Challenge" applications, perhaps none is grander than simulation of the human brain. Reflecting the complexity and scale of the brain with current computer technology is truly a daunting task. But a group of researchers and computer scientists at a number of UK universities are attempting to do just that under a project named SpiNNaker.
SpiNNaker, which stands for Spiking Neural Network architecture, aims to map the brain's functions for the purpose of helping neuroscientists, psychologists and doctors understand brain injuries, diseases and other neurological conditions. The project is being run out of a group at University of Manchester, which designed the system architecture, and is being funded by a £5m grant from the Engineering and Physical Sciences Research Council (EPSRC). Other elements of the SpiNNaker system are being developed at the universities of Southampton, Cambridge and Sheffield.
For the casual observer, constructing a facsimile of the most complex organ in the human body from digital technology may see like a natural fit for computers. The view of the brain as a biological processor (and the processor as a digital brain) is well entrenched in popular culture. But the designs are fundamentally different.
Operationally, computers are precise, extremely fast and deterministic; brains are imprecise, slow, and non-deterministic. And, of course the underlying architectures are completely different. Computers relying on digital electronics, while the brain employs a complex mix of biomolecular structures and processes.
The SpiNNaker design meets the architecture of the brain halfway by going for lots of simple, low-power computing units, in this case, ARM968 processors. The initial Manchester-designed SpiNNaker multi-processor is a custom SoC with 18 of these processors integrated on-chip. (The original spec called for 20 processors per chip.) The multi-processor also incorporates a local bus, called Network-on-Chip or NoC, which links up the individual processors and off-chip memory. Each SpiNNaker node is reported to draw less than one watt of power, while delivering the computational throughput of a typical PC.
The design is purpose-built to simulate the action of spiking neurons. Spiking in this context means when neurons are stimulated above a certain threshold level to generate an event that can be propagated across a neural net. But instead of using neurotransmitters to do this, the computer is just passing data packets around.
To be truly useful, the spiking needs to happen in real-time. Fortunately, this is where computer technology shines. Electrical communication is actually more efficient than the biochemical version, so nothing exotic needs to be done in the hardware to make all this magical neural spiking a virtual reality.
And that may happen soon. The design phase of the project is coming to a close and the SpiNNaker team is starting to gather the pieces together. According to a news release this week, SpiNNaker chips were delivered in June (from Taiwan -- presumable TSMC), and have passed their functionality tests. The plan is to build a 50,000-node machine with up to one million ARM processors.
While that seems like a lot, researchers estimate that it will only be enough to represent about one percent of the real deal. A human brain contains around 100 billion neurons along with 1,000 million connections and a single ARM processor in the SpiNNaker chip can only handle 1,000 neurons. The good news is that one percent may be enough to answer a lot of questions about the functional operation of the brain.
Even at one percent, the scale of the machine is probably the trickiest part of the project. With so many processors in the mix, there are bound to be individual failures at fairly regular intervals. To deal with the inevitable, the designers made SpiNNaker fault tolerant at multiple levels. For example, each of the ARM processors can be disabled if they fail at start-up and a chip can remain functional even if "several processors fail." If an entire chip goes south, data can be rerouted to neighboring chips thanks to redundant inter-chip links.
The other challenge to scaling out is power, but here is where the ARM architecture pays dividends. The initial system of 50,000 nodes is estimated to draw just 23 KW to 36 KW of power. By supercomputing standards, that's just a pittance. Of course, judged against the 20 watt version in our heads, SpiNNaker has a ways to go.
The power profile suggests that if there are no inherent scaling limitations in the hardware or software, the design could conceivably be used to build a machine that would support a "complete" human brain simulation for just a few megawatts. With improved process technology, that could easily slip into the sub-megawatt level.
For all that, SpiNNaker isn't designed to simulate higher level cognitive features -- the most interesting function of the brain. Inevitably that will require more complex hardware and software. So even if someone builds a super-sized SpiNNaker, it won't come close to the functionality of the 100 percent organic version anytime soon.
Posted by Michael Feldman - July 07, 2011 @ 7:40 PM, Pacific Daylight Time
Michael Feldman is the editor of HPCwire.
No Recent Blog Comments
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?