TALK TITLE: FUJITSU DEEP LEARNING UNIT
Fujitsu has more than three decades of experience developing AI and associated technologies, and its K Computer still stands as one of the most powerful commercial supercomputers in the world. But it is never enough as HPC wants to push power efficiency as high as possible and squeeze as much performance out of that power envelope. On the processor side the recipe is to have out of order execution and other challenging techniques as well as multiple execution units and large cache memory on chip, not to mention frequency tweaks. All of these demands are in direct conflict with any low-power processor, in particular specialized engines focusing on neural networks. Such low power device engines must be implemented with fewer transistors, less control logic and specialized highly efficient execution units operating at a lower frequency. The goal of the Fujitsu Deep Learning Unittm is to provide 10x the performance per watt compared to other available technologies, with high efficiency and unprecedented scalability.
I started to punch holes in cards in 1975, got a PhD in Distributed Computing some years after and signed for a first job at the Software Engineering Institute in Paris. Since, I worked at the National Space Agency (CNES) on the design of Satellite Control Centre, followed by first true HPC job as HPC Analyst at Alliant Computer in Boston in the late 80’s, and finally joined Fujitsu, after starting working in Japan on message passing technology and performance analysis tools, I came back to Toulouse where I still devote my days (and sometimes nights) to High Performance Computing.