Homepage of WAran Research Foundation (WARFT)

  • Increase font size
  • Default font size
  • Decrease font size
Home Director's Vision
Director's Vision
Ever lasting encouragement from my students and my own dear ambition to contribute to the benefit of humanity, both led me to institute the research foundation WAran Research FoundaTion (WARFT). At WARFT I have formed these major research groups - Charaka, Vishwakarma, Marconi, Ramanujan, Hardy, Naren and Bhaskara.

 The following are a few of the major directions in which a rich multi-disciplinary research is being carried out at WARFT.

Combating the Brain Disorders and Diseases:

The fine tuning of a research process involves  deeper thinking associated with  the sensory pathways unified in the cognitive space of internal mental events  - learning, memory, feelings, actions, reactions and so on. What happens if the Temporal Lobe functionally breaks off? How could this happen? Imagine that you can listen to others, but cannot identify the speakers. What has gone wrong with the  auditory cortex? 

Nothing is  disheartening than the plight and sight  of children suffering from brain disorders and diseases. Can Our Science & Research give the powers possessed by all humans to children with brain disorders or other handicaps? Is it plausible to  transform a  spastic child  into the   dynamics  of  life?  WARFT  shall spend every effort through space and time to render love and affection to such children.

WARFT is into a relentless research to reach Beyond the Brain's Horizon. WARFT  is  focusing on  interdisciplinary research areas like neuroscience, brain modeling ,  supercomputing clusters & applications, and deep sub-micron technology issues in billion device processor. Other areas include  image processing, digital signal processing,  mixed signal design. and the associated fields.

It is an Himalayan task for a single organization to venture into such massive socio-scientific pursuit for the benefit of humanity. In this context WARFT has been registered as a charitable trust under the Govt. of  TamilNadu.  

Brain modeling and the Supercomputing Clusters:

There is a  divided opinion among neuroscientists regarding the role  of powerful supercomputing clusters for investigating both the known and unknown regions of the brain.

One class of experimentalists prefer to probe either individual neurons or as many as possible simultaneously and record their spiking activities under different stimulus conditions and during different time intervals to investigate the process of  information encoding and decoding in neuronal pathways. Synapse, scientists spend their life time in  analyzing its dynamics, the intra and extra cellular Ca2+ concentration and the related kinetics. Their efforts have led to breakthrough in research and in invention of new drugs. There are an other class of neuroscientists who try to develop new experimental methodologies and  investigate patients suffering from  unique disorders to  understand the brain functions keeping away complex mathematical, bio-chemical and bio-physical aspects.  Computational neurobiologists research on evolving functional models of single neuron  and  neuronal population and further on to the ion channels and synapse dynamics. 

To date  no research initiative has come up to evolve methodologies involving  theories and experiments to predict biologically realistic  connectivity pattern of millions of neurons of a brain region. This prediction model should help generate and track the spatiotemporal firings of millions of neurons at individual neuron level. This has never been posed as a challenging  research. 

Why do we need to know such complex connectivity patterns?

In depth understanding of the biological connectivity of the millions of neurons is the root  to investigate the information processing responsible for the functionality of  a brain center and modeling the information processing  across the boundaries of these brain centers and their  integrated functionality.

With this interconnectivity, it is possible to tract the synaptic organization of  individual neurons and model the learning process at single neuron level and analyze how each neuron contributes towards the learning process.

For Biologically Realistic Fault Simulation of  a Brain region, its neuronal interconnectivity  is of prime importance. Biological Fault Simulation of Brain regions will greatly help invent new and effective drugs.

Such prediction model for complex connectivity pattern should encompass the dendrites, axons, soma , synapse by arriving at the  Distributions of the associated biophysical and biochemical parameters of the brain region.

The parameters could be Distribution of Dendritic  Branches, their Length, Tapering and Electrical Parameters. Distribution of Axon Length, the Telodendria, Nodes of  Ranvier and R & C Electrical Parameters. Distribution of Geometrical Coordinates for a single synapse contact along the Telodendria and the corresponding Dendrite Branch, Synaptic Distribution of every individual Neuron etc.


Million Neurons  Connectivity OF A BRAIN REGION whose BOLD fmri is known - MMINi-DASS:   A SIMULATION MODEL FROM WARFT (pursued by Charaka Group) 

"It is the complexity of the connections between the many elements, not that of the individual components, that make complex information processing possible. Individual neurons  can carry out computations because they are wired together in organized and diverse way" -Eric R. Kandell, Nobel Laureate

"With the advent of functional imaging studies with positron emission tomography (PET), functional magnetic resonance imaging (fMRI) and electroencephalography, there has been a rapid increase in our knowledge of the functional and the anatomical map of the human cortex. These techniques, however, do not attempt to identify the mechanisms or neuronal circuits responsible for these functions. Thus, the challenge is to discover what is actually happening when different regions of the cortex are activated under different sensory or behavioral tasks. Fundamental to this central endeavor is an understanding of the structure and function of the microcircuits of the neocortex and their components."-Gordon M Shepherd

 This is the first ever attempt to predict the massive neuronal interconnectivity of a brain region based on the BOLD fMRI and the haemodynamics of that region. It is a Great  Expedition Beyond Brain's Horizon. To be successful in this expedition what is needed is not just the conventional Supercomputers, but something  radically  different. 

THE FUSION CLUSTER FROM WARFT: An evolutionary approach towards the design and development of a SUPERCOMPUTER to achieve performance in the range of Hundreds of EXAOPS.(pursued by Vishwakarma Group)

Because of the increasing demand in several application areas like protein folding, galaxy simulation, computational fluid dynamics, brain modeling it becomes mandatory to evolve supercomputing clusters whose node architecture is highly tuned towards the application. This to a large extent alleviates the time complexity which is currently predicted in Petaflop years for these applications running in general purpose supercomputers. But evolving such architectures for specific applications may be cost prohibitive. How do we overcome this problem? 

Make the node architecture heterogeneous involving varied Algorithm Level Functional Units (ALFUs) like sorter, graph partitioner, matrix multiplier, matrix inverter ,square rooter ,adjacency/incidence matrices generator, DFS/BFS units, covering  wider applications. This necessarily leads to the evolution of a new concept of Simultaneous Multiple Applications (SMAPP) mapping on to a cluster to efficiently harness the heterogeneous functionality of the cluster nodes. This simulataneous multiple application mapping is expected to achieve high sustained performance.

To achieve steep increase in processing power in the range of Hundreds of Exaops, Thousands of ALFUs of types listed above need to be organized to build the architecture of the heterogeneous node to reach sub-Petaops at the node level itself. Each node is a SuperComputer On a Chip (SCOC). Billion device technology can be effectively utilized to implement such nodes.

Though the simultaneous multiple applications mapping may appear complex, it is resolved by the higher level library composed of Algorithm Level Instruction Set Architecture -ALISA (due to the presence of ALFUs), resulting in a simplified mapping process. ALISA will drastically reduce  number of memory access over the conventional supercomputing node .On the other hand, the libraries for general purpose supercomputers are designed specific to every application and are bound to be complex due to their generic nature( Arithmetic Instructions) leading to much increased memory access. Thus due to the  general purpose nature of the nodes, the conventional supercomputers Petaops performance can be achieved through employing  large number of nodes running into several millions.     

Wired and Wireless communication across the SCOC node cluster

  1. There are several hundreds of wired MIP SCOC clusters (Islands) and communication across them is wireless. These Individual islands need to be EMI Shielded to help in frequency reusability across islands. Such an approach is expected to provide the following significant advantages. Reduced power over wired networked topology.
  2. Reconfigurability across the wireless islands can be well achieved matching the communication complexity present in the  applications.
  3. Efficient simultaneous mapping of applications  through such wireless clusters.
  4. Reconfiguration, in a wired topology, matching with simultaneous multiple applications will be highly complex.    

The Fusion Cluster will be driven by hierarchically based multiple hosts.     
  1. To manage parallel mapping of SMAPP across million nodes.
  2. To speedup the overall mapping processes
  3. To manage Very Large I/O.

The Fusion Cluster will tackle very wide class of special purpose applications and also in comprehending the complex problem of brain modeling.

THE MIP SCOC from WARFT: (pursued by VISHWAKARMA Group)

These challenges need to be tackled in the most effective way to strongly support the the evolution of Fusion Cluster.

Billion device node processor design and implementation is a major challenge to architects. This involves a number of important factors to be addressed such as testing, reliability, fault tolerance, and at deeper levels - interconnect modeling, efficient clock distribution, hot spots, ground modeling for analyzing ground current gradients and power loss. 

INTERCONNECT DOMINANCE  (pursued by Naren , Marconi ,Hardy groups)

  • Immediate impact of interconnect dominance is on algorithm designers who try to tackle the computational complexity. Under post DSM technology the DFT algorithm is expected to be  more efficient for the hardware implementation than the FFT algorithm.
  • Existing Analog Models of the Deep Submicron Interconnects---complex computations and unstable.
  • Electromigration along the interconnect-- reliability of BUS Structures
  • Detection of WEAK 0s and WEAK 1s-Frequency Domain Testing of Instruction Execution of Multi Ghz Processors. Analyzing the signal integrity of clock signal in multi level H-tree is made simpler through the digital mode
  • Tracking the Electromigration and Interconnect Aging
  • Surprisingly no research initiative has come up to bring in Fault Tolerance in BUS Architecture
  • Interconnect Digital Model FROM WARFT-reduced time complexity, stable and better accuracy over AWE and RICE WARFT is  putting  forth a major research initiative in this direction.
Testing and Fault Tolerance, Overcoming the problem of Electromigration , Cross Talk and Intermittent  Faults, Cross Talk will give rise to Burst Noise under multi Ghz clocking.

Instruction and Data Encoding  -  Which class of Codes, Systematic or Non-Systematic   

Low Power High Performance Mixed Number System                                      

This Number system is a combination of the Sign-Log Number System and the RNS Number System called LRNS. This is found to be very effective in reducing the MIP SCOC power - Low Power Supercomputing.

Special Purpose Matrix Processors --low power has become an important issue in DSP/Image processing applications  Low Power Sparse Matrix Operations.

Ground Plane Modeling

Partitioning has great impact on Ground Voltage Distribution. Highly non-uniform Ground Voltage Gradient along the ground plane will affect the Noise Margin and can lead to Ground Plane Hotspots. Efficient Algorithms for optimizing the number of ground and power pins and their physical distribution along the die area are essential.

Encoding  of spatio-temporal neuronal activities in the Time-Frequency (TF) domain:

A new perspective on Neuronal Information Processing along the visual path way from WARFT. Information encoding and decoding in the brain has been a focus of major research. These mainly concentrate on information processing by the neurons in time domain. Even these time domain information processing have not been well understood in spite of massive research activities and  particularly with regard to the decoding .  Another important observation is, these types of neural encoding of the inputs are totally  disregarded in all object recognition models proposed till now. This becomes biologically unrealistic.

This directed us to think differently, leading to analyze the spatio-temporal neuronal activity in the Time-Frequency (TF) domain. This approach unveils a new direction  in correlating the input stimulus, single, multiple, and mixed objects to a corresponding set of dominant harmonics in the TF domain, after Gabor transforming the Neural Spike Activities of a region of interest. Even the spatial information that gets encoded along with the objects and the object count can be correlated with the Dominant Harmonics present in the TF domain .We strongly believe that this will lead us to build a comprehensive object model involving the neural encoding of input stimulus.  Recent work has been towards encoding in the Time-Frequence-Phase domain and using Gabor Phase for object perception.
Another major research initiative is integrating the spatiotemporal neuronal activities of the sensory cortices to develop an integrated cognitive model to analyze behavior and reasoning.  

A recent interview where Prof. N. Venkateswaran talks about the importance of research can be found here
Title Filter     Display #