From the Editor | Main Blog Index
August 18, 2011
The company that brought Jeopardy cyber-champ Watson to the world has come up with some even brainier technology. IBM has unveiled prototypes of microprocessors that are being dubbed "cognitive computing chips." According to IBM, the hardware is made up of neural circuits and are designed to mimic the brain's ability to perceive sensory input, understand it, and take action based on that understanding.
To emulate human-level thinking, Watson used clever software on conventional hardware, which could be described as sort of a brute force approach. In contrast, these new chips are designed to behave fundamentally like our own brains, being able process sensory input in a vastly parallel fashion, create correlations, learn from experience, and adapt its processing dynamically.
Although the work can be categorized as artificial intelligence, it's actually more general than traditional AI, which tends to be focus on individual capabilities, like pattern recognition. In this latest effort, IBM is attempting to integrate all aspects of thinking, including perception and action, as well as cognition. The goal is to build a truly intelligent machine that is able to perform human-level analytics in real time, and with the power and size efficiencies of a biological brain.
The work is being done with funding from the Defense Advanced Research Projects Agency (DARPA) and in conjunction with a number of US universities. The project, which the agency has titled Systems of Neuromorphic Adaptive Plastic Scalable Electronics, or SyNAPSE, kicked off in 2008, when DARPA anted up $4.9 million to bootstrap the work. In the second phase of the project, they chipped in $16.1 million more to IBM and its university partners. Now DARPA is prepared to kick in an additional $21 million for the third phase.
The defense agency's interest in such technology is understandable, given the DoD's increasing reliance on drone aircraft and other types of unmanned vehicles, not to mention its need to analyze the tremendous amounts of intelligence data. In short, the defense department would like nothing better than to replace its legions of soldiers and analysts with computer chips.
IBM, of course, is aiming at a much larger market. The idea is that system could be hooked up to vast sensor networks, monitoring the environment, cars, homes, even people. Such a product would span every industry and provide the underpinnings of IBM's so-called Smarter Planet, although in this case it's more like Smarter Planet 2.0.
The hardware design is certainly futuristic. The cognitive computing prototype chips are made up of a "neurosynaptic core," which encompass of computational circuits (the neurons), memory (the synapses), and communication lines (the axons). Although this is accomplished with standard digital circuitry, the architecture is unique. From the IBM press release:
IBM’s overarching cognitive computing architecture is an on-chip network of light-weight cores, creating a single integrated system of hardware and software. This architecture represents a critical shift away from traditional von Neumann computing to a potentially more power-efficient architecture that has no set programming, integrates memory with processor, and mimics the brain’s event-driven, distributed and parallel processing.
Getting compute, memory and communication integrated together is central to the architecture's brainy behavior. According to Dharmendra Modha, IBM's lead on the project, the tight integration is key to getting the circuitry to behave like biological neurons and synapses, and do so within a very organic power budget. The power consumption of the human brain is estimated to be between 10 and 100 watts.
To date, IBM has developed two chip prototypes, both of which have been implemented on 45 nm SOI CMOS at the company's fab in Fishkill, New York. Each design contains 256 neurons, one with 256K programmable synapses and the other with 64K learning synapses. The IBM researchers claim to have used them to demonstrate simple applications like navigation, machine vision, pattern recognition, associative memory and classification.
IBM has not revealed a timeline for any commercial products. The company's goal is to eventually construct a system with ten billion neurons and a hundred trillion synapses, while consuming a single kilowatt of power. Using future nanoelectronics, the researchers estimate such a machine will take up less than two liters of volume.
The chip prototypes will be described in more detail at the IEEE Custom Integrated Circuits Conference on September 20 in San Jose, California.
Posted by Michael Feldman - August 18, 2011 @ 7:16 PM, Pacific Daylight Time
Michael Feldman is the editor of HPCwire.
No Recent Blog Comments
The Jordanian government has requisitioned a custom-built supercomputer, which is now being used by universities across the country. What it's made of may be surprising – but it's not unprecedented.
Read more...
The top research stories of the week include novel methods of data race detection; a comparison of predictive laws; a review of FPGA's promise; GPU virtualization using PCI Direct pass-through; and an analysis of the Amazon Web Services High-IO platform.
Read more...
NetApp and SGI made simultaneous announcements yesterday of new storage systems specifically targeted at HPC and big data applications. NetApp launched its latest offering in the E-Series line of controllers, the NetApp E5500, and OEM partner SGI adopted the technology for its InfiniteStorage 5600 RAID system.
Read more...
Mar 05, 2013 |
How does the Phi coprocessor measure up to Xeon "Sandy Bridge" brand-mate?
Read more...
Mar 05, 2013 |
The cloud-based experiment seeks participants from the HPC, CAE, life sciences, and big data communities.
Read more...
Mar 01, 2013 |
Create DIY HPC clusters in the Amazon cloud with the StarCluster toolkit. Here's how.
Read more...
Feb 27, 2013 |
When it comes to Titan's final acceptance testing, ONRL says not so fast.
Read more...
Feb 27, 2013 |
On the lighter side of HPC, here's a video explaining what supercomputers do.
Read more...
03/07/2013 | HP and Intel | Businesses that want to be on the cutting edge of their industries are increasingly turning to high-performance computing (HPC) solutions to handle complex compute processes and speed up their rate of innovation. Download this Executive Brief to see how businesses in energy, life sciences and entertainment put HPC solutions to work in their operations.
02/11/2013 | Allinea | Researchers have no time to lose in their quest to divine the future of the Earth’s climate and its effects on humanity. When a problem arises, one of the world’s top centers for climate research relies on Allinea DDT.
Join Mellanox and Xyratex on April 2 for this educational webinar to learn about the challenges faced in moving to a cloud infrastructure and the storage solutions available to enable success in today’s evolving service provider data center environment.
Highlights:
CFD models of 100 million+ cells are increasingly common. Models of a billion+ cells are emerging. To meet demand for 10 to 100x larger models, the simulations will have to scale to 10 to 100x more cores. Both compute environment capabilities and capacity must be enhanced. Learn how Cray and ANSYS are scaling production models to 1000s of cores and providing the compute environment to take CFD simulations to the next level of extreme fidelity.