skip to content

Department of Computer Science and Technology


This internship opportunity is for a current undergraduate student. If you are interested in applying, please email your CV to Eiko Yoneki (

The Department of Computer Science and Technology actively supports equality, diversity and inclusion and particularly encourages applications from underrepresented groups.

Project Title:  Optimisation of DNN Accelerators using Bayesian Optimisation

Supervisor: Eiko Yoneki

Co-supervisors: Aaron Zhao


Essential knowledge, skills and attributes: Linux, basics of Machine Learning, basics of computer architecture, Python: Experience of using associated tools such as Pytorch, ideally basics of Bayesian Optimisation

Description:  A recent trend in deep neural network (DNN) development is to extend the reach of deep learning applications to platforms that requires different conditions such as energy constrained, where in order to reduce the DNN model size and improve the hardware processing efficiency, and have resulted in DNNs that are much more compact in their structures and/or have high data sparsity. These compact or sparse models are different from the traditional large models in that there is much more variation in their layer shapes and sizes, and often require specialised hardware to exploit sparsity for performance improvement. Therefore, many DNN accelerators designed for large DNNs do not perform well on these models. The recent evolution of NN architecture will eventually require transformer model, or GNN specific accelerators, where a complex trade-off between what should be implemented in hardware and what should be left to software needs to be considered.

The project aims at design-space exploration of new System-on-a-Chip (SoCs) especially DNN accelerators using Gem5-Aladdin Simulator, developed in Harvard University [1]. Gem5/Aladdin is a tool for end-to-end simulation of SoC workloads, including workloads with accelerated functions handled by fixed-function hardware blocks. Synthesising chips for experiments is very expensive, but the simulator itself is also expensive to run, and the large parameter space makes it prohibitive to use parameter sweepers (a search space here is ~264 combinations).

Bayesian Optimisation (BO) offers a data-efficient approach to global optimisation of black-box functions by building a model of the unknown function, including the uncertainty about its values we have yet to observe. Using simulation, various objective functions will be examined exploring a large parameter space for the model.

Thus, the task in this project is exploring the parameter space for various models and workload over Gem5/Aladdin. Moreover, such heterogeneous architectural design space consists of multiple design goals, e.g. performance and power and it will add multi objective function for BO.

If time allows, the project will be explored to deal with high dimensional parameter space using the structured BO described in [2] or [3].

[1] gem5/laddin

[2] BOAT: Building Auto-Tuners with Structured Bayesian Optimization.

[3] BoGraph: Structured Bayesian Optimization From Logs for Expensive Systems with Many Parameters

Length of project: 12 weeks

Remote or in person: Remote or in person participation is possible.