Research Interests

  • Signal Processing & Machine Learning
  • Cellular & Wifi Standards - PHY systems
  • CPU - GPU Architectures for ML
  • System Design for Distributed Processing
  • Information and Learning Theory
Recent Awards
  • Divisional Recognition Awards @ Intel Labs for leading Algorithms Research in DARPA SpaceBACN - 2022,2024
  • Distinguished Invention Award @ Intel Labs for AI/ML based PHY techniques - 2021
  • Best Paper Award in IEEE International Conference on Communications (ICC) - 2021
Biography
Hi ! I'm currently a Senior Principal Architect at Marvell Technology, Santa Clara since Aug 2024. Previously I was a Staff Applied AI Research Scientist @ Intel Labs, Santa Clara working on Signal Processing and AI/ML algorithms design between Sep 2018 and Aug 2024. Prior to that, I was a Staff Systems Researcher at Qualcomm Research - Corporate R & D, San Diego between 2014 and 2018. Between 2003 and 2009, I held various positions at Intel Corporation and Samsung R&D Institute. I received my Ph.D. and M.S. in Electrical & Computer Engineering from University of California, Irvine (UCI) in 2014, advised by Prof. Syed Ali Jafar and B.E.(Hons.) in Electrical Engineering from BITS Pilani, India in 2003. At UCI, I worked on Information Theoretic Capacity characterization of Multi-user Interference networks with Non-generic channels. Over the last 2 decades, I've contributed to systems research for various product designs and features including but not limited to Qualcomm Snapdragon chipsets (with Apple, Samsung, LG.. clientele), Oneweb satellite systems, Samsung Kalmia Modems, Intel FlexRAN, and DARPA SpaceBACN.
US Citizen soon (Interview pending) 

University of California, Irvine
Ph.D. in Electrical & Computer Engineering      Sep 2011 - Sep 2014 
M.S. in Electrical & Computer Engineering       Sep 2009 - Sep 2011 
Thesis : Interference Alignment - Beyond Generic Channels
Advisor : Prof. Syed A. Jafar
PhD Committee: Prof. Hamid Jafarkhani & Prof. Ender Ayanoglu

Birla Institute of Technology & Science, Pilani (India)
B.E. (Hons.) in Electrical Engineering          Aug 1999 - Jul 2003

Marvell Technology, Santa Clara -- Connectivity products
Senior Principal Architect                        Aug 2024 - Present

  • Research & Design of new features for DSP & Custom AI Products towards Data Center and AI applications within Marvell Connectivity group
Intel Labs, Santa Clara -- Systems, Software and Architecture Org
Staff Applied AI Research Scientist            Sep 2018 - Aug 2024
  • 3GPP 5G - AI/ML for PHY: I was responsible for research and design of Probabilistic AI/ML techniques for FlexRAN PHY using Bayesian Inference for Computing and Communications, primarily using Variational Inference (VI) and Markov Chain Monte Carlo (MCMC) techniques. Co-author and inventor of Metropolis-Hastings Gradient Descent, which combined MCMC and Gradient descent techniques to enable superior performance at low complexity, winning ICC 2021 Best Paper Award in Machine Learning for Communications Track. Championed Expectation Propagation based MIMO demodulation and characterized its performance+complexity for Intel FlexRAN. Developed AVX512 vectorized implementation of EP algorithm for 3GPP 5GNR PUSCH and PUCCH receiver processing pipelines, evaluated in Skylake family and was adopted in FlexRAN. This improved BLER performance by upto ~2.5 dB, BER by 5 dB in various scenarios with 2.3x complexity of MMSE solution. Performed extensive analysis of correlation metrics considering MIMO channel propagation for dynamic switching of EP. Also, characterized performance of iterative bayesian receivers involving both channel estimation and MIMO demodulation. Proposed Softmax acceleration for improving EP and NLP architectures in Intel GNA accelerator roadmap. Liaison for various NSF and Intel sponsored university research programs such as MLWiNS and RINGS. Moderated several technical panel discussions and research updates for numerous NSF/Intel funded research projects.
  • 3GPP Positioning: Lead research efforts for ML based multi-modal positioning algorithms involving TDOA, AOA, RSSI and other metrics using SRS, TRS and other reference signals. Demonstrated improved sub-meter positioning in Non-line of sight (NLoS) environments with business group & external partners. Adopted in Integrated sensing & Communications research and product roadmap. Multipath model for NLoS tracking was expanded using various neural network architectures for fixed & mobile sensing.
  • 3GPP Massive MIMO CJT: Contributed to design of Coherent Joint Transmission (CJT) feature in 3GPP towards Rel-19 standardization. Performed Analysis of Inter-site synchronization aspects such as phase, timing and frequency offsets and characterized realistic system capacity improvements for both uplink and downlink CJT. Devised novel precoding algorithms for joint precoding from multiple TRPs, towards various split 7-2 DU-RU architectures. Filed 2 patents on joint processing needed for coherent joint transmission
  • DARPA program: Spearheaded Systems/Algorithms R & D efforts in various phases of DARPA SpaceBACN program to build a Reconfigurable Coherent Optical Modem (100Gbps, 100W, 100k$) and successfully delivered system model and algorithms design. Fully functional modem was recently demonstrated to Chief of Space Operations Gen. Chance Saltzman and US Space Force teams.
Qualcomm Research, San Diego -- Corporate R & D / Directly reported to Dr. Wanshi Chen
Staff Systems Researcher                        Oct 2014 - Aug 2018
  • 3GPP NTN: Research and System design for Oneweb Low Earth Orbit (LEO) Satellite Communication Systems - I worked on the design and analysis of Reverse link channels primarily on User Terminals. This was an adaptation of LTE air interface with single carrier waveforms enabling Satellite to Ground communication in Ka, Ku bands, which eventually got standardized as part of 3GPP Non-Terrestrial Networks (NTN) Standards.
  • 3GPP LTE+5GNR R&D: Contributed to Research & Design for Non-linear Interference Cancellation for Carrier aggregation usecases in LTE and 5G NR systems. Such carrier aggregation caused receiver desense due to harmonics and intermods within the receiver bands, which were digitally removed using digital self interference cancellation mechanisms. These designs were commercialized in various Qualcomm Snapdragon Modem chipsets between 2014 and 2017, used by Apple, Samsung, LG etc.. Closely partnered with Apple teams to deliver ~18 dB performance improvement and dynamic switching of NLIC feature. Resolved several complex system issues with TX carrier aggregation generating Inter-mod distortions.
Samsung R & D Institute, Bangalore -- Modem Algorithms Lab / Wireless Terminals Division
Lead Engineer                                    Apr 2006 - Jul 2009
Senior Engineer                                Dec 2004 - Mar 2006
  • 3GPP GSM-EDGE DSP: Responsible for Research, Design of Interference Cancellation algorithms for 2G/3G Communication systems. Lead the Single antenna interference cancellation (SAIC) design and development for GSM/GPRS/EDGE chipsets of Samsung using Ceva Teak and Teaklite 3 DSP. Contributed to various stages of product design from algorithm research, design and DSP assembly programming, Inter-Operability testing and Field testing. Optimized various algorithms to meet stringent cycle consumption and memory budgets, and was later commercialized in Kalmia line of Samsung Multi-mode chipsets between 2009 and 2011.
  • 3GPP2 EVDO Standards: I was also part of CDMA EVDO standardization group of Samsung focusing on Radio Link Protocol designs and Multi-flow architectures enabling diverse QoS guarantees, for which I also contributed to Modem designs for packet scheduling and QoS provisioning. Collaborated on various features for RLC, MFPA development in EVDO chipset family. Submitted several standard contributions towards new architecture of 1xEVDO Rel-B and Rel-C.
Intel Corporation, Bangalore -- Technology & Manufacturing Group
Software Engineer                            Aug 2003 - Dec 2004
  • Design of various manufacturing system (FAB/ATM) software enhancements for the mission-critical system called "Recipe Correlation Table" (RCT) that manages lot-processing recipes in all Intel sites. Also, I was a developer for Unit Level Traceability (ULT) handler software for an important traceability program adopted in Intel manufacturing sites.

AI/ML in Wireless & Sensing

Coherent Joint Transmission
  • P3: USPTO 20230198815 A1 Methods and devices for joint processing in Massive MIMO systems
    Intel : Lu Lu, Sundar Krishnamurthy, Clark Chen, Thushara Hewavithana
  • P4: USPTO Jun2024 Filing: System and methods for Clustering, User Association and Power Control to enable Coherent Joint Transmission
    Intel : Sundar Krishnamurthy, Lu Lu, Clark Chen, Thushara Hewavithana
DARPA SpaceBACN - DSP Architecture DARPA SpaceBACN - System Aspects MIMO Interference cancellation Radio Link Protocol (RLP)

Thesis - Interference Alignment : Beyond Generic Channels

The radical idea of Interference alignment enabled Capacity or Degrees of Freedom characterization (DoF, a first order approximation) for many multi-user interference networks. Various alignment schemes provided fundamental insights into the number of accessible signal dimensions in communication networks where the output signals are linear functions of the input signals.

Most of the prior art dealt with generic channels wherein the channel coefficients are assumed to be independent and drawn from a continuous distribution, continuous alphabet with infinite diversity, and the network is often single-hop. These assumptions were challenged due to the following reasons

  • Poor scattering in MIMO systems and network topology lead to spatial dependencies, manifested as rank deficient channels,
  • Multi-hop dependencies arise due to the presence of relays, and
  • Linear network coding applications (as in wired networks) act as finite field counterparts of wireless networks, with limited diversity.
Hence, my research focus had been on characterizing the Capacity / DoF of linear communication networks with Non-generic channels, through the following three categories of networks.
Single-hop Rank-deficient Interference channel: We found the DoF of the K-user MIMO rank deficient interference channel, with different ranks for the direct (D0) and the cross channels (D). For such rank deficient interference channels, it was shown that
  • The rank deficiency of direct channels does not help DoF and could hurt, while
  • The rank-deficiency of cross-channels does not hurt DoF and could improve.
The main challenge was to account for the spatial dependencies introduced by rank deficiencies in the interference alignment schemes that typically rely on the independence of channel coefficients.

Conferences - Globecom 2012 , ISIT 2014 Journal - Trans on. Info Theory, Jan 2015

Multi-hop Rank-deficient Interference channel: We found the DoF of 2-hop MIMO rank deficient interference channel with different channel ranks in the first (D1) and the second hops (D2), for which a rank-matching principle was identified reminiscent of impedance matching in circuit theory. For this channel, the DoF loss was shown to be the rank-mismatch (|D1-D2|) between the two hops under moderate rank deficiencies.

Conference - Globecom 2014            Journal - Trans. on Info Theory, Sep 2015

Finite field channels: Capacity results for the finite field counterparts of wireless networks were found, in order to explore the implications of channels being from a finite alphabet with limited diversity. By characterizing the capacity of constant finite field channels over Fp^n for 2-user X channel and 3-user interference channel, interesting parallels were drawn between p and SNR, and n and Channel Diversity.

Conference - ISIT 2013                Journal - Trans. on Info Theory, July 2014

Journal Publications

Conference Proceedings Thesis: Interference Alignment : Beyond Generic Channels
Copyright information for IEEE publications: Copyright is held by IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

Intel Labs
Qualcomm, Samsung

Courses
Stanford (Coursera): Machine Learning, Neural Networks and Deep Learning, Improving Deep Neural Networks, Convolutional Neural Networks, Structuring Machine Learning Projects, Sequence Models, AI for everyone

UC Irvine (Graduate): Information Theory, Estimation and Detection Theory, Design & Analysis of Algorithms, Advanced System Software, Digital Signal Processing, Digital Image Processing, Digital Communications, Random Processes, Space Time Coding, Wireless Communications

Competencies
Software: Matlab, C/C++, Python, CUDA, Simulink, SQL, Java
Tools/Libraries: Visual Studio, VTune, Intel AVX2, AVX512, OneAPI MKL, Nvidia Nsight, Boost
ML Frameworks: Tensorflow, PyTorch, Caffe, Keras, Stan, OneDNN, NumPy, JAX, OpenCV, Jupyter
DSP: Simulink-HDL, Qualcomm QDSP6, Ceva Teak, Teaklite-3


Reviewer: IEEE Transactions in Information Theory, IEEE Transactions in Signal Processing, IEEE Transactions in Vehicular Technology, IEEE Transactions in Wireless Communications, and several conferences

Member: IEEE Information Theory Society, IEEE Signal Processing Society, IEEE Communications Society