The supercomputer “Fugaku” is the first exascale system with 1 EFLOPS(SP) peak performance, which is operated since March 2021 in R-CCS, Japan. Fugaku is an ultra-scale “general-purpose” manycore-based system with 158,976 nodes, 7.6M cores in total. Fugaku was designed in application-oriented and power-efficient approach to accommodate variety kinds of large-scale applications. Recently, as several supercomputers such as ORNL Frontier supercomputer in US are reaching to exascale, interests in the next of exascale systems are increased. Although it will be very difficult to reach to “Zetta” FLOPS, the system will be heterogenous and integrated with several different configurations for the next of exascale as “Zetta”-scale. Acceleration technologies including FPGAs and reconfigurable devices are promising and inevitable to challenge the “Zetta” scale computing. In this talk, the overview and features of “Fugaku” system will be presented, followed by the challenges for “Zetta”-scale computing.

Computing on edge is one of approaches to reach singular intelligence point with low carbon cost. The recent network architecture search (NAS) has provided a convenient means to find optimized deep learning networks for edge computing by searching bitwidth, sparsity and rank etc., which usually results in a mixed-complexity network at each layer. However, there is no hardware to support computing mixed-complexity networks on edge. This talk will show an agile hardware-software co-design for NAS-optimized deep learning networks. It mainly shows details on how to build a mixed- bitwidth/sparsity vector systolic accelerator with high energy efficiency. Recent works on tensor systolic and computing-in-memory architectures will be also mentioned briefly.

Abstract: Atom switch is a two-terminal resistive switch based on Cu precipitation in the polymer solid-electrolyte and is characterized by non-volatility, adjustable ON/OFF conductance ratio, radiation tolerance, and small footprint. Atom switch advantageously applies to both routing switch in FPGA and memory cell in non-volatile memory (NVM), resulting in high energy efficient electronics. The ON conductance and the cycling endurance, which are in a trade-off relationship, are controlled by programming current to meet the requirements of its application. Atom switch provides a large ON/OFF conductance ratio for FPGA. It also provides a 100k-cycle endurance for NVM applications. Nonvolatile SoC-FPGA solution with low power consumption is achieved by multi-usage atom switch.