Ran Ginosar

Technion-Israel Institute of Technology, Israel

AI acceleration challenges on shared-memory manycores

Abstract

Shared-memory manycores integrate a large number of ‘standard’ cores with per-core vector units, shared memory, and special accelerators.
Cores and accelerators communicate only through shared memory, rather than directly with each other.
Multiple manycore chips are combined into large systems, together with a large memory.
Research questions related to manycore systems used for AI acceleration include: balance of number and size of cores vs. size of shared memory per chip, balance of vector units vs. matrix multiplication accelerator per chip, and memory access methods off-chip – RDMA vs. large shared address space vs. networked memories.

Biography

Ran Ginosar has earned his BSc in EE&CS from the Technion—Israel Institute of Technology in 1978, and his PhD in EECS from Princeton University in 1982. He has conducted research at Bell Laboratories, Intel and the University of Utah. He joined the Technion Faculty in 1983, where he is a Full Professor of Electrical Engineering and Head of the VLSI Systems Research Center. Ran co-founded several companies in areas of electronic imaging, medical devices, wireless communications and manycore architecture. Prof. Ginosar serves as the President of Ramon.Space, a company engaged in high-performance, low-power and highly reliable computing for Space applications in Satellites and Spacecraft.

Download