arXiv subjects

Milos Popovic

Publications and source records attributed to Milos Popovic.

5 recordsLinked to original sources

arXiv2025

Systolic Array-based Accelerator for Structured State-Space Models

Sequence modeling is crucial for AI to understand temporal data and detect complex time-dependent patterns. While recurrent neural networks (RNNs), convolutional neural networks (CNNs), and Transformers have advanced in capturing long-range dependencies, they struggle with achieving high accuracy with very long sequences due to limited memory retention (fixed context window). State-Space Models (SSMs) leverage exponentially decaying memory enabling lengthy context window and so they process very long data sequences more efficiently than recurrent and Transformer-based models. Unlike traditional neural models like CNNs and RNNs, SSM-based models require solving differential equations through continuous integration, making training and inference both compute- and memory-intensive on conventional CPUs and GPUs. In this paper we introduce a specialized hardware accelerator, EpochCore, for accelerating SSMs. EpochCore is based on systolic arrays (SAs) and is designed to enhance the energy efficiency and throughput of inference of SSM-based models for long-range sequence tasks. Within the SA, we propose a versatile processing element (PE) called LIMA-PE to perform traditional and specialized MAC operations to support traditional DNNs and SSMs. To complement the EpochCore microarchitecture, we propose a novel dataflow, ProDF, which enables highly efficient execution of SSM-based models. By leveraging the LIMA-PE microarchitecture and ProDF, EpochCore achieves on average 2000x improvement in performance on LRA datasets compared to a GPU and 250x gains in performance and 45x improvement in energy efficiency, over traditional SA-based accelerators (TPU).

cs.LG↗

arXiv2023

Fine-Resolution Silicon Photonic Wavelength-Selective Switch Using Hybrid Multimode Racetrack Resonators

In this work, we describe a procedure for synthesizing racetrack resonators with large quality factors and apply it to realize a multi-channel wavelength-selective switch (WSS) on a silicon photonic chip. We first determine the contribution of each component primitive to propagation loss in a racetrack resonator and use this data to develop a model for the frequency response of arbitrary order, coupled-racetrack channel dropping filters. We design second-order racetrack filters based on this model and cascade multiple such filters to form a 1x7 WSS. We find good agreement between our model and device performance with second-order racetrack that have ~1 dB of drop-port loss, ~2 GHz FWHM linewidth, and low optical crosstalk due to the quick filter roll-off of ~ 5.3 dB/GHz. Using a control algorithm, we show three-channel operation of our WSS with a channel spacing of only 10 GHz. Owing to the high quality factor and quick roll-off of our filter design, adjacent channel crosstalk is measured to be <-25 dB for channels spaced on a 10 GHz grid. As a further demonstration, we use five of seven WSS channels to perform a demultiplexing operation on both an 8 GHz and a 10 GHz grid. These results suggest that a low-loss WSS with fine channel resolution can be realized in a scalable manner using the silicon photonics platform.

physics.optics↗

arXiv2020

A community-powered search of machine learning strategy space to find NMR property prediction models

The rise of machine learning (ML) has created an explosion in the potential strategies for using data to make scientific predictions. For physical scientists wishing to apply ML strategies to a particular domain, it can be difficult to assess in advance what strategy to adopt within a vast space of possibilities. Here we outline the results of an online community-powered effort to swarm search the space of ML strategies and develop algorithms for predicting atomic-pairwise nuclear magnetic resonance (NMR) properties in molecules. Using an open-source dataset, we worked with Kaggle to design and host a 3-month competition which received 47,800 ML model predictions from 2,700 teams in 84 countries. Within 3 weeks, the Kaggle community produced models with comparable accuracy to our best previously published "in-house" efforts. A meta-ensemble model constructed as a linear combination of the top predictions has a prediction accuracy which exceeds that of any individual model, 7-19x better than our previous state-of-the-art. The results highlight the potential of transformer architectures for predicting quantum mechanical (QM) molecular properties.

physics.comp-ph↗

arXiv2016

Identification of Intended Arm Movement Using Electrocorticographic Signals

A Brain Computer Interface (BCI) is a communication system that receives neurological signals from the brain and translates them into control commands for electrical (e.g., computer mouse) and electromechanical (e.g., Wheelchair) devices. The development of such systems was intended originally to aid individuals with a condition called locked-in syndrome. Individuals with this condition have lost all their voluntary muscle control but remain cognitively intact (i.e., mentally aware of their surroundings- can feel emotions, recognize objects/people but are unable to move). This means that they are trapped in their own bodies. The use of BCI may one day improve the independence and quality of life of people with this disability.

cs.HC↗

arXiv2015

Photonics design tool for advanced CMOS nodes

Recently, the authors have demonstrated large-scale integrated systems with several million transistors and hundreds of photonic elements. Yielding such large-scale integrated systems requires a design-for-manufacture rigour that is embodied in the 10 000 to 50 000 design rules that these designs must comply within advanced complementary metal-oxide semiconductor manufacturing. Here, the authors present a photonic design automation tool which allows automatic generation of layouts without design-rule violations. This tool is written in SKILL, the native language of the mainstream electric design automation software, Cadence. This allows seamless integration of photonic and electronic design in a single environment. The tool leverages intuitive photonic layer definitions, allowing the designer to focus on the physical properties rather than on technology-dependent details. For the first time the authors present an algorithm for removal of design-rule violations from photonic layouts based on Manhattan discretisation, Boolean and sizing operations. This algorithm is not limited to the implementation in SKILL, and can in principle be implemented in any scripting language. Connectivity is achieved with software-defined waveguide ports and low-level procedures that enable auto-routing of waveguide connections.

physics.optics↗