Heterogeneous computing platforms for resource-aware video and data analytics (HOPWARE)

Lead Research Organisation: University of Bristol
Department Name: Electrical and Electronic Engineering


HOPWARE will investigate how the multi-objective platform created in the CERBERO H2020 project (https://www.cerbero-h2020.eu/) can be applied to heterogeneous computing platforms that combine deep-learning compute units with different capabilities, performance and power characteristics. The proposed heterogeneous computing platform in HOPWARE combines devices like Google EdgeTPUs, Intel VPUs, ARM processors and Xilinx FPGAs programmable via a single machine-learning framework like TensorFlow so that once training completes the network can be run in an optimal configuration of multiple devices. Project partner UPM created in CERBERO an adaptive framework called ARTICO that can assess the status of the system at run-time and deploy different reconfiguration strategies to improve its energy efficiency, robustness and performance. In this research we are particularly interested in the ARTICO/CERBERO approach to run-time monitoring, sensing and estimation of capabilities to assist in the decision of how to distribute the available resources to achieve a functional goal within constraints in power and execution time.


10 25 50
Description A hardware accelerator working with sparse matrix can be used long a dense accelerator and different threshold points can be identified to switch between both modes depending on the classification task complexity resulting in significant energy savings. Dynamic function exchange can be used to extend this to precisions involving floating point and fixed point numbers.
Exploitation Route Further funding to explore how this can be combined with dynamic function exchange or partial reconfiguration as developed by academic partner UPM has been obtained from the Leverhulme trust.
Sectors Aerospace, Defence and Marine,Electronics

Description "High-performance video analytics with parallel heterogeneous neural networks
Amount £37,000 (GBP)
Funding ID IF-2021-003 
Organisation The Leverhulme Trust 
Sector Charity/Non Profit
Country United Kingdom
Start 01/2022 
End 06/2022
Description international collaboration Universidad Politecnica De Madrid 
Organisation Technical University of Madrid
Country Spain 
Sector Academic/University 
PI Contribution Edge processor optimized for sparse/dense tensor computing in Tensorflow Lite.
Collaborator Contribution Use of facilities including computer systems, office space and dynamic reconfiguration technology part of the CERBERO EU project.
Impact Nunez-Yanez, J. L., & Hosseinabady, M. (2021). Sparse and dense matrix multiplication hardware for heterogeneous multi-precision neural networks. Array, 12, [100101]. https://doi.org/10.1016/j.array.2021.100101
Start Year 2021