I agree Our site saves small pieces of text information (cookies) on your device in order to deliver better content and for statistical purposes. You can disable the usage of cookies by changing the settings of your browser. By browsing our website without changing the browser settings you grant us permission to store that information on your device.
Neuromorphic computing implements aspects of biological neural networks as analogue or digital copies on electronic circuits. The goal of this approach is twofold: Offering a tool for neuroscience to understand the dynamic processes of learning and development in the brain and applying brain inspiration to generic cognitive computing. Key advantages of neuromorphic computing compared to traditional approaches are energy efficiency, execution speed, robustness against local failures and the ability to learn.
In the HBP the neuromorphic computing subproject carries out two major activities: Constructing two large-scale, unique neuromorphic machines and prototyping the next generation neuromorphic chips.
The large-scale neuromorphic machines are based on two complementary principles. The many-core SpiNNaker machine located in Manchester (UK) connects 500.000 ARM processors with a packet-based network optimized for the exchange of neural action potentials (spikes). The BrainScaleS physical model machine located in Heidelberg (Germany) implements analogue electronic models of 4 Million neurons and 1 Billion synapses on 20 silicon wafers. Both machines are integrated into the HBP collaboratory and offer full software support for their configuration, operation and data analysis. The most prominent feature of the neuromorphic machines is their execution speed. The SpiNNaker system runs at real-time, BrainScaleS is implemented as an accelerated system and operates at 10.000 times real-time. Simulations at conventional supercomputers typical run factors of 1000 slower than biology and cannot access the vastly different timescales involved in learning and development ranging from milliseconds to years. Recent research in neuroscience and computing has indicated that learning and development are a key aspect for neuroscience and real world applications of cognitive computing. HBP is the only project worldwide addressing this need with dedicated novel hardware architectures
|The BrainScaleS system is based on physical (analogue or mixed-signal) emulations of neuron, synapse and plasticity models with digital connectivity, running up to ten thousand times faster than real time.||The SpiNNaker system is based on numerical models running in real time on custom digital multicore chips using the ARM architecture.|
|The BrainScaleS system (NM-PM-1) contains 20 8-inch silicon wafers in 180 nm process technology. Each wafer incorporates 50 x 106 plastic synapses and 200,000 biologically realistic neurons. The system does not execute pre-programmed code but evolves according to the physical properties of the electronic devices, running at up to 10 thousand times faster than real time.||The SpiNNaker system (NM-MC-1) provides almost 30,000 custom digital chips, each with eighteen cores and a shared local 128 Mbyte RAM, giving a total of over 500,000 cores. A single chip can simulate 16,000 neurons with eight million plastic synapses running in real time with an energy budget of 1W.|
Whereas the base chips now operating in the two large-scale machines have been developed in various national and European projects over the last decade, next generation chips are now being developed in the HBP. Those next generation chips will be the basis for the next generation of large-scale machines to be operational towards the end of the current project planning around 2023. In the ramp-up phase first prototypes for both complementary approaches, SpiNNaker and BrainScaleS, have been designed and produced. Both chip prototypes make use of the huge technological progress in solid state manufacturing technology and are a result of a close collaboration with neuroscientists in the HBP. Special emphasis is put on greatly improved capabilities for efficient implementation of learning and development.
Layout drawing of the third HICANN-DLS prototype ASIC for the BrainScaleS 2 system.
|Prototype SpiNNaker-2 test chip and small-scale system (Santos)|
A number of demonstrations of the benefits of neuromorphic technology are beginning to emerge , and more can be expected in the short to medium term. Various start-up companies are emerging, in the USA and elsewhere, to exploit the prospective advantages of neuromorphic and similar technologies in these new machine learning application domains. In the HBP, small and large-scale demonstration systems are available and attract an increasing number of users from industry and academia. While these systems are primarily made for basic research on understanding information processing in the (human) brain, efforts are being made to also implement machine learning tasks on them. Next generation chips of the SpiNNaker and BrainScaleS architecture will be available for first users in the first half of 2018.
In the medium term we may expect neuromorphic technologies to deliver a range of applications more efficiently than conventional computers, for example to deliver speech and image recognition capabilities in smart phones. (Currently such capabilities are available only using powerful cloud resources to implement the recognition algorithms.) These will require small-scale neuromorphic accelerators integrated with the application processor, using a fraction of the resources of a single chip. Large-scale systems may be used to find causal relations in complex data from science, finance, business and government. Based on the causal relations detected such neuromorphic systems may be able to make temporal predictions on different time-scales.
In the long term there is the prospect of using neuromorphic technology to integrate energy-efficient intelligent cognitive functions into a wide range of consumer and business products, from driverless cars to domestic robots. While human-level “strong” artificial intelligence remains a mystery, and indeed may depend on the emergence of an understanding of information processing in the biological brain (through initiatives such as the Human Brain Project) before it becomes a practical reality, there are many useful applications that can benefit from more modest cognitive capabilities. The technology is relatively young, and there is much uncertainty as to where it will find its place in the wider world, but it clearly meets a need in the rapidly changing world of computing. The fact that major companies like IBM have defined cognitive computing as their main business for the future makes the development of neuromorphic hardware architectures especially interesting and economically attractive.
The Neuromorphic Computing platform targets researchers in multiple fields, including computational neuroscience and machine learning. Platform users will be able to study network implementations of their choice including simplified versions of brain models developed on the HBP Brain Simulation Platform or generic circuit models based on theoretical work.
The platform also offers industry researchers and technology developers the possibility to experiment with and test applications based on state-of-the-art neuromorphic devices and systems.
Compared to traditional HPC resources, the Neuromorphic systems potentially offer higher speed (real-time or accelerated) and lower energy consumption. The accelerated systems are particularly suited for investigations of plasticity and learning, enabling simulation of hours or days of biological time in only a few seconds.
Users do not need to be members of the Human Brain Project, but need an HBP community account (free of charge). To request an account and then access the systems via the internet, please start here.
The systems still have rough edges, but the platform offers user support and training, and the software supporting the platform is continuously being improved.
Both systems (BrainScaleS and SpiNNaker) have an interface designed for neuroscience researchers, based on Python scripts using the PyNN API for simulator-independent specification of neuronal network models. PyNN scripts also run on the popular software simulators NEST, NEURON and Brian.