Dr. Michael Manzke

Dr. Michael Manzke is an Assistant Professor in Computer Science at TCD since 1999 and a member of the GV2 group. He received his PhD from TCD in 2006 and has published over 30 peer reviewed publications in journals and conference proceedings e.g. ACM Transactions on Graphics (TOG) and International Journal of Computer Vision (IJCV). Furthermore, he filed 4 patents. Dr. Manzke's focusses on high performance low power graphics architectures and algorithms. This work is equally applicable to computer vision problems. Interactive Computer Graphics and Computer Vision applications require systems that deliver high performance while consuming very small amounts power. This is increasingly important for ubiquitous mobile applications such as augmented reality. He and his team of researchers address these research questions by investigating microarchitectures and algorithms that are specifically designed for Visual Computing computation. He works closely with industry partners such as Xilinx, Dolphin Interconnect Solutions, Intel, Sony, Toshiba, IBM, and Movidius on these problems. Dr. Manzke is supervising and supervised in total 13 team members: Postdocs(3), PhD(8), and Research Engineers(2). He has received 8 grants from Science Foundation Ireland, IRCSET, Enterprise Ireland, and Irish Research Council with a total amount of €1,826,319 and a total amount allocated to the candidate of €1,155,159.

  Computer architecture   Computer graphics, Meta computing   Computer Science/Engineering   Digital systems, representation   Distributed systems   High performance computing   Image synthesis   Kalman Filter   Modelling, modelling tools, 3D modelling   Networks and telecommunications research   Peripherals technologies   Signal processing   Systems control, Modelling, Neural networks   Systems Engineering   Systems/Control

My PhD work on cluster performance estimation led to an interest into scalable reconfigurable architectures for large scale interactive graphics applications. I moved from the Computer Architecture Group to the Graphics Vision and Visualisation Group (GV2). I investigated with my PhD students and postdocs a scalable reconfigurable interface for Graphics Processing Units (GPUs) as a closely coupled alternative to high latency cluster solutions. We also investigated a hardware message passing protocol that interconnects clusters of field programmable gate array (FPGA) directly as a cost-effective alternative to a share memory solution. At that point I changed my focus from scalable rasterization hardware to scalable ray-tracing hardware because I saw this as a potential solution for future large scale interactive graphics applications. I also believe that scalable closely coupled systems need to execute parts of their algorithm on microprocessors and other aspects on fixed function hardware (microarchitectures). Consequently, we investigated hardware accelerated collision detection and hardware accelerated construction of SAH based bounding volume hierarchies for interactive ray-tracing with particular promising results for the BVH builder. Furthermore, we investigate joint acceleration data structures for collision detection and ray tracing. This work can be seen in Nvidia's latest GPUs. Another interest developed through my research activities as Co-Director of the SONY TOSHIBA IBM (STI) Cell Center of Competence Europe. We investigated some computer vision problems particular Skeleton Extraction and Multiple View Reconstruction. We also worked, in close collaboration with Sony Broadcast & Professional Research Laboratories, at real-time Cell processor algorithm for their professional stereo camera rigs. More recently my work addresses the problem of real-time interactive visualisation of 3D scientific data, namely volumetric datasets, which are ubiquitous in many application domains such as Medicine, Biology, Biomechanics, Neuroscience, Fluid Dynamics and Veterinary Medicine. Specifically, this work deals with the problem of time-variant and multivariate volume data sets which have become more readily available in recent years due to improved scanning technologies and simulation techniques. Such data sets are particularly difficult to deal with due to their computational complexity as well as the information overload that can arise from such visually challenging data. We employ graphics hardware and parallel computing strategies to address the performance issue, whilst information overload is dealt with using techniques from perceptually adaptive graphics, computational aesthetics and non-photorealistic rendering which apply smart techniques to improve usability and understandability of data. I did this research in collaboration with my colleague Dr. John Dingliana. I focus on the performance issue while Dr. Dingliana looked after the perceptually adaptive graphics. Furthermore, much of my recent work focuses on low power architectures for computer graphics, deep learning, and light fields for volume rendering.