Arjuna Madanayake

Associate Professor, ECE Department at FIU


Lecture Information:
  • October 6, 2023
  • 2:00 PM
  • CASE 241

Speaker Bio

Arjuna Madanayake is an Associate Professor at the Dept. Electrical and Computer Engineering at FIU. He is the leader at the RF, Analog and Digital (RAND) Laboratory where he advises 8 Ph.D. students on multiple topics on wireless communications, analog and digital circuit design, FPGA systems, RF and microwave systems, AI/ML accelerators on hardware, mm-wave and 6G software defined radios, radar sensing, sub-THzradios, computer architecture and VLSI systems. Dr. Madanayake has been recently supported by more than 10 awards from NSF, and more than 12 additional awards from DARPA, ONR, CIA, NIH, NASA, Lockheed Martin, and others. He has graduated 8 Ph.Ds and more than 20 MS students in his career. He is interested in cybersecurity at the hardware level, computer architecture and processor design for AI/ML, and cryptographic/number theory based digital VLSI architectures. Dr. Madanayake served FIU has the Graduate Program Director in charge of the PhD program for Dept. of ECE for 1.5 years. He is a member of IEEE, and founding member of the Circuits and Systems Education and Outreach (CASEO) technical committee. For fun, Arjuna spends his spare time (a theoretical quantity) on music and HiFi systems. He has a deep love for vintage HiFi equipment.

Abstract

Analog computers have been winning wars for America well before the dawn of ENIAC, transistors or discovery of Moore’s Law. In fact, WW2 warships made extensive use of analog computers for targeting of its long range weapons. However, since the invention of digital logic, Von Neumann architecture, software, digital integrated circuits, and the exponential scaling of computation described by Moore’s Law, the world of analog computing has been all but forgotten. Nonetheless, the arcane world of analog computing has recently experienced a comeback. Believe it or not, analog computers are far from dead. In fact, they are thriving in niche areas of computation. Quantum computers are one type of analog computer making use of quantum physics principles. Another type of analog computer can be found in the world of machine learning (ML). Deep learning has driven digital computation to its limits. In fact, the development of larger and larger ML models are severely stunted by compute requirements, which for modern systems can require hundreds of megawatt hours of compute power. Analog computers are being pursued by companies, such as, IBM, Mythic, and Aspinity, towards achieving better energy efficiency for training large scale ML models. Advanced scientific computing (think fusion modeling, nuclear modeling, magnetohydrodynamics), much like AI and ML, requires extensive compute. This talk covers our recent DARPA STTR Phase-1 and -2 work on design of efficient and fast radio-frequency analog computers as custom CMOS integrated circuits. Our analog systolic arrays are for solving specialized problems in physics based simulation requiring spatio-temporal solution of linear and nonlinear partial differential equations. We address design, implementation, and test of analog computers using 180nm CMOS technology for applications in solving linear Maxwell’s Equations, nonlinear acoustics, and analog FFT. This talk covers a detailed view of analog computers for solving electromagnetic problems, including algorithm design, circuit mapping, IC design, fabrication, test and measurement and comparison with digital technologies such as NVIDIA GPUs and Xilinx FPGAs. Our example analog computer was measured to be 420x faster than GPUs, while consuming 200mW of power whereas the GPU consumed more than 100W for the same realization. The analog computer chip was measured to be 28x more efficient than FPGA based implementation of systolic-array based optimized realizations of the algorithm. We conclude that fast analog computing is viable for niche computing applications that are compute intensive for deep math operations.