The Efficient circuits & systems for Machine Intelligence (EMI) research lab focuses on designing energy-efficient digital AI hardware (ASIC & FPGA) for edge computing, emphasizing real-time AI-based radio frequency and video/audio signal processing. 

We strongly believe that hardware-software co-design is the path to realizing the ultimate goal of intelligent edge solutions for complex real-world applications.

New master projects are available!

April 17, 2024

We're thrilled to announce our collaboration with GlobalFoundries in the GF 12LP+ University Partnership Program to support our research on Low-Power AI Hardware Accelerators using the cutting-edge 12 nm technology for enabling energy-efficient signal processing in transmitters for future wireless communications technologies and potentially useful for many other applications such as augmented reality and wearable healthcare.

Feb 2, 2024

The premier International Microwave Symposium (IMS) has accepted our paper "MP-DPD: Low-Complexity Mixed-Precision Neural Networks for Energy-Efficient Digital Pre-distortion of Wideband Power Amplifiers," mainly done by our PhD students Yizhuo Wu and Ang Li. The paper was ranked in the Top 50 and invited to the IEEE Microwave and Wireless Technology Letters. This paper proposes a novel DPD with Fixed-Point-Floating-Point Mixed-Precision parameters and intermediate states, which can reduce DPD inference power consumption by 2.6X on 160MHz 1024-QAM OFDM signals without losing linearization performance (-38dB EVM).

MP-DPD is trained using our recently released OpenDPD framework, available at https://github.com/lab-emi/OpenDPD.

Jan 19, 2024

We are thrilled to release OpenDPD: An Open-Source PyTorch-based End-to-End Learning & Benchmarking Framework for Wideband Power Amplifier Modeling and Digital Pre-Distortion (DPD). Authored by Yizhuo Wu, Gagan Deep Singh, Mohammad Reza Beikmirza, Leo de Vreede, Morteza Alavi, and Chang Gao (https://arxiv.org/abs/2401.08318). DPD enhances signal quality in wideband RF power amplifiers (PAs) and is a critical module in future 6G or Wi-Fi 7 wireless communication systems. OpenDPD comes with a free digital power amplifier I/Q dataset for you to train and benchmark machine learning (ML)/artificial intelligence (AI)-based DPDs and fairly compare them with other works. We will collaborate closely with our industrial partners to update this infrastructure periodically. OpenDPDv2 will come later this year with also free analog PA datasets.

OpenDPD code, datasets, and documentation are publicly available at https://github.com/lab-emi/OpenDPD. This work will be presented as a lecture at the 2024 IEEE International Symposium on Circuits and Systems (ISCAS), Singapore, in the Special Session: RFIC & AI: Pioneering New Wireless Communications.

Jan 5, 2024

One and a half years after the acceptance, our paper "Spartus: A 9.4 TOp/s FPGA-Based LSTM Accelerator Exploiting Spatio-Temporal Sparsity" finally appeared in the first issue of IEEE Transactions on Neural Networks and Learning Systems in 2024 (Volume: 35, Issue: 1, January 2024). This paper presents an ultra-low-latency Long Short-Term Memory (LSTM) RNN hardware accelerator for speech recognition by exploiting temporal-weight-hybrid sparsity through our proposed neuromorphic DeltaLSTM algorithm and a column-balanced structured pruning method to ensure workload balance. Spartus achieves down to just a 1-microsecond latency in a 14.2M-parameter LSTM network inference. This technique to combine temporal activation and spatial weight sparsity makes the artificial RNN run like a neuromorphic Spiking Neural Network (SNN) without losing accuracy; thus, it will be useful for realizing energy-efficient RNN edge inference for various latency-critical signal processing tasks.

August 24, 2023

We are glad to announce that our paper "3ET: Efficient Event-based Eye Tracking using a Change-Based ConvLSTM Network" has been accepted at the 2023 IEEE BioCAS conference. The paper introduces the Efficient Event-based Eye Tracking (3ET) system that employs a novel temporally sparse change-based ConvLSTM architecture to effectively extract sparse spatio-temporal features from event camera input streams in eye-tracking applications. Our network achieved over 30% better accuracy than CNN-based models and reduced computational operations by 4.7X compared to standard ConvLSTM models on event-based pupil datasets. This marks the first step of the EMI lab in building energy-efficient neuromorphic vision systems for emerging AIOT applications.