Intel Collaborative Research Institute for Computational Intelligence
Accelerators for Massive Memory Parallel Machine Learning and Applications
Emerging 3D-stacking technologies promise large-scale integration of memory and compute elements in a single package, which will connect many gigabytes of memory and accelerators comprising hundreds or thousands of compute elements using interconnects that offer orders of magnitude higher bandwidth and lower latency than existing compute/memory interfaces. These hybrid accelerator/memory structures hold great potential for those machine learning algorithms that are both memory- and compute-bound.
The research team under Prof. Ran Ginosar (Technion) will explore the disruptive implications of such integration and study its impact on both accelerator architecture and machine learning. The goal is to provide insights into the computational and memory structures required for effective integration, what type of machine learning algorithms this tight integration will enable, and how will it affect the design of machine learning algorithms.
From machine learning perspective, the team plans to design new machine learning algorithms and paradigms matching the novel capability of memory intensive architecture. The goal is to work in the entire spectrum of scale, ranging from small devices that use only few such chips to server and cloud scales composed of several orders of magnitudes more units.
On the applications side, the team will develop robust algorithms that combine modern, machine-learning based image and video analysis techniques with novel efficient lower level image processing operations, in a manner suitable for execution on these novel accelerators.
The first year work consists of analysis of the amount of parallelism in existing code (both sequential and task-based) and developing algorithms that leverage large amounts of memory near the computing unit for various machine learning and image/video applications. In this year the team will analyze hardware/software computing and storage systems that make use of new accelerators so as to enable even the most demanding, I/O-intensive applications to enjoy native performance.
By the third year the team will focus on one or two test cases in machine learning and image/video analysis. The team will develop accelerator algorithms and architectures motivated by them.
Prof. Ran Ginosar, Technion EE
Prof. Yuval Cassuto, Technion EE
Prof. Koby Crammer, Technion EE
Prof. Avinoam Kolodny, Technion EE
J. Haj-Yihia, Y. Ben-Asher, E. Rotem, "Compiler Assessed CPU Power Management,", Compiler, Architecture and Tools Conference, sponsored by HiPeac, Haifa , Israel, November 2013