Seminar 2015-11-09 G.G.
Computing in Space with OpenSPL
Speaker: Georgi Gaydadjiev, Mellanox and Imperial College, TBA
Georgi research interests include embedded systems design, computer architecture and micro-architecture, reconfigurable computing, hardware/ software co-design, VLSI design, and computer systems testing. Currently my research is on dynamic techniques to manage contemporary distributed memory systems and heterogeneous multicores, application specific acceleration, and low overhead reliability techniques. All the above topics are in the context of the expected technology trends in the future when the quality of the available transistors is deteriorating while their quantity on a single chip keeps scaling up. These trends combined with the severe power dissipation constraints require novel, holistic hardware-software approaches and place new challenges to computer systems designers. A possible oversimplification of the problem is how to transform the available transistor quantity into affordable quality in respect to power dissipation per unit area.
For a long time all atomic arithmetic and storage structures of computing systems were designed as two-dimensional (2D) structures on silicon. Currently processor vendors offer chips with steadily growing numbers of cores and recent circuits started to grow in the third dimension by integrating silicon dies on the top of each other. All of this results in severe increase of the programming complexity. To date, predominately the one-dimensional view of computing systems organization and behavior is used forming a severe obstacle in exploiting all the associated advantages. To enable this, a more natural, at least 2D view of computer systems is required to represent closer the physical reality in both space and time. This calls for radically novel approaches.
Computing in space allows designers to express complex mathematical operations in a more natural, space area aware way and map them on the underlying hardware resources. OpenSPL is one such approach that can be used to partition, lay out and optimize programs at all levels from high-level algorithmic transformations down to individual custom bit manipulations. In addition, the OpenSPL execution model enables highly
efficient scheduling (or better called choreography) of all basic computational actions with the guarantee of no side effects. It is clear that this approach requires a new generation of design tools and methods and a novel way to measure (or rate) performance as compared to all traditional practices. In this talk we will address all of the topics relevant to spatial computing and show its enormous capabilities to design power efficient computing systems. Examples and results based on real systems will emphasize the advantages of this approach but will also stress the difficulties along the road ahead.