Svein Anders Tunheim has followed the PhD Programme at Faculty of Engineering and Science, specialisation in ICT, Scientific field of Artificial Intelligence.
- Trial lecture starts at 10:15
- Public defence starts at 12:15
Title of trial lecture: "The energy cost of AI: Cost for training LLMs, their deployment and the need for specialised Hardware"
Read the thesis in the Norwegian Research Information Repository
Disputation chair: Andreas Prinz
Assessment committee:
- Professor Christos-Savvas Bouganis, Imperial College London, UK
- Professor Dag Trygve Eckhoff Wisland, UiO
- Professor Vladimir Oleshchuk, UiA
Supervisors in the doctoral work:
- Main supervisor: Professor Lei Jiao, University of Agder
- Co-supervisor: Professor Rishad Shafik, Newcastle University
Summary of thesis
Electronic devices that perform artificial intelligence (AI) tasks at the edge of communication networks are known as Edge AI devices. Such units typically have significant compute and hardware resource constraints and need to operate within limited energy budgets. For such systems, hardware accelerators are often used to offload the main system processor from various demanding tasks.
The aim of this thesis was to develop Tsetlin Machine (TM)-based, low-power hardware accelerators for image classification in Edge AI devices. Unlike deep neural networks, the logic-based structure of the TM algorithm does not require complex arithmetic circuits. The TM is therefore well-suited for implementation of low-power circuit solutions.
As a significant number of stochastic decisions are required during the training phase of a TM, it was studied how the length of repeating sequences of random numbers impacted the accuracy of the trained TM solutions. These results were taken into account during the design of the TM-based hardware accelerators.
Based on the convolutional TM (CTM), an accelerator for pattern recognition using the "2-Dimensional Noisy XOR" dataset was implemented on a field-programmable gate array (FPGA), which is a programmable integrated circuit. This solution supported both training and inference, and achieved a test accuracy of 99.9%, which was on par with the software version. It was the first reported hardware solution using the CTM.
An FPGA accelerator based on the convolutional coalesced version of the TM (ConvCoTM) with 128 clauses was developed. Full on-device training was supported, requiring minimal interaction with the system's main processor. Using the well-known MNIST dataset consisting of handwritten digits, a test accuracy of 97.6% was achieved. With a clock frequency of 50 MHz, the accelerator core performed 134 thousand image classifications per second, and a single training epoch of 60 thousand samples required a processing time of 1.5 seconds. This was the first reported hardware solution based on the ConvCoTM.
To demonstrate ultra-low-power operation, a ConvCoTM inference application-specific integrated circuit (ASIC) was designed. This was a synchronous, all-digital accelerator comprising 128 clauses. It was implemented in a 65 nm CMOS technology, and was the first reported ASIC based on the ConvCoTM. The accelerator achieved a test accuracy on MNIST of 97.42%, precisely aligning with the software model. The energy consumed by the ASIC core per classification was 8.6 nJ. This was the second lowest result among all reported solutions for manufactured chips operating on MNIST with comparable test accuracy, and the lowest reported result for an all-digital solution.
The hardware architectures developed for the CTM and ConvCoTM are well-suited to scaling and for operation on larger and more complex images. Distinct trade-offs between throughput and hardware resource utilization were identified. The FPGA-based hardware accelerators demonstrated full on-device training capabilities, enabling Edge AI devices that can learn continuously.
What to do as an online audience member
The disputation is open to the public. To follow the trial lecture and the public defence online, register on Zoom.
We ask online audience members to join no earlier than 10 minutes in advance. After these times, you can leave and rejoin the meeting at any time.
Opponent ex auditorio:
Deadline for the public to pose questions is during the break between the two opponents. Questions ex auditorio can be submitted to Kristine Reinfjord.
Contact person
- kristine.reinfjord@uia.no
- Phone
- +47 37 23 30 14