Georgia Institute of Technology
Energy-efficient On-chip Deep Neural Network (DNN) Inference and Training with Emerging Non-volatile Memory Technologies
Abstract
dc:description.abstractArtificial intelligence (AI) based applications are becoming pervasive in our daily life. The emerging non-volatile memory (eNVM) technologies are regarded as promising technological candidates for building energy-efficient AI hardware for edge devices. This thesis identifies and resolves several challenges related to AI hardware design for deep neural network (DNN) training and inference using eNVM-based technologies. For DNN inference, the existing compute-in-memory (CIM)-based AI accelerator suffers from area scalability and lacks reconfigurability for different DNN models. Besides, for large DNN models such as transformers, it is challenging to store the model parameters on-chip and execute various types of matrix multiplication workloads efficiently. For DNN training, the existing memory technologies, such as static random access memory (SRAM) and embedded dynamic random access memory (eDRAM), are not satisfactory in buffering a large amount of training data. Besides, the performance of eNVM-based in-memory training is limited by the high write energy and low write endurance of eNVM devices. In this thesis, I explored and presented cross-layer solutions to the abovementioned challenges involving device technology selection, circuit/architecture design, and neural network model optimization. For DNN inference, I proposed a monolithic 3D integration scheme based on the back-end-of-line (BEOL) compatible semiconducting oxide transistors. Together with a reconfigurable interconnect design using the ferroelectric field-effect transistor (FeFET) based routing switch, the scalability and reconfigurability of the CIM-based DNN accelerator are improved. A 3D heterogeneous computing platform using stacked FeFET CIM dies, and digital logic die is demonstrated for transformer models. It leverages the high memory density to store model parameters and the heterogeneity of computing paradigms to optimize the hardware performance for different matrix multiplication workloads. For DNN training, I designed an innovative dual-mode memory architecture based on ferroelectric random access memory (FeRAM). It demonstrates excellent dynamic and static performance for the DNN training accelerators by optimally switching between the volatile and non-volatile operation modes. To further improve the energy efficiency of DNN training, I proposed an in-memory training architecture using a novel hybrid weight cell design. It overcomes the high write energy and low write endurance of eNVM devices.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Electrical and Computer Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Luo, Yandong
- Advisor dc:contributor.advisor
-
- Yu, Shimeng
- Committee members dc:contributor.committeemember
-
- Hao, Cong (Callie)
- Lin, Yingyan (Celine)
- Krishna, Tushar
- Mukhopadhyay, Saibal
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/75066