English

We conduct research on computer architectures that achieve high performance processing at low power, taking into account technological trends in semiconductor manufacturing and packaging technologies. In particular, we explore computational models and microarchitectures of computing accelerators, system software such as their compilers, compute nodes and networks in general-purpose parallel computing systems, as well as domain-specific computing hardware.

As the semiconductor technology scaling, known as Moore's Law, slows down and the quality of the scaling itself changes from conventional planar gate structures to three-dimensional structures such as FinFETs and gate-all-around (GAA), it is becoming increasingly difficult to achieve the lower power and higher processing performance required for HPC (High-performance computing) and AI. In particular, more power is consumed for data movement on a semiconductor chip and reading/writing to off-chip memory such as DDR and HBM than for computing on a chip, and therefore it is required to explore a computer architecture that reduces unnecessary power consumption in achieving the desired processing by making data movement and memory access more efficient.

Also, in recent years, it is a mainstream to connect multiple chiplets, which are relatively small chips, to form a processor with a single socket. This is because semiconductor technology scaling alone does not significantly improve power consumption or cost per transistor while it increases the number of transistors per unit area; yields tend to decrease for large-area semiconductor chips along with semiconductor scaling; on-chip memory such as SRAM and analog devices are less likely to benefit from density improvements due to semiconductor scaling compared to logic that constitutes computing circuits. Chiplets allow processors, accelerators, memory, analog input/output (PHY), and optical-electro conversion devices to be manufactured using each of appropriate semiconductor technology nodes for them. Chiplets also have the advantage of improving yields by reducing the chip area. On the other hand, packaging technology, which is the technology to connect multiple chiplets integrated on a socket, is becoming more and more important. In addition to 2.5D Chiplet connection with silicon interposer that has been developed so far, 3D stacking technology is being developed especially for faster data transmission between memory and logic memory chips while reducing power consumption.

Considering the technological trends described above, our team researches computer architectures that achieve high performance processing at low power. Specifically, we are working on the following research topics:

• Exploration of next-generation computer architectures for high-performance and low-power HPC and AI:
  Targeting implementation with ASIC (application specific integrated circuit) and the next-generation memory technologies, in particular, we research coarse-grained reconfigurable arrays (CGRAs) that enable pipelined parallel execution of various arithmetic operations based on data-flow computing model.

• Exploration of application-specific computer architectures that use circuit structures optimized for each problem:
  For example, we research dedicated hardware for low latency and high throughput quantum error correction required for fault-tolerant general-purpose quantum computers, and dedicated hardware for high-throughput data processing near detection sensors in synchrotron radiation facilities such as RIKEN SPring-8.

• Research and development of systems for research on high-performance application-specific computer architectures:
  We develop an FPGA cluster system, in which state-of-the-art FPGAs (Field-Programmable Gate Array) are interconnected by a dedicated network, as a platform for research and development of the above dedicated hardware. In addition to the hardware system and the network between FPGAs, research is being conducted on the system software and on how to operate the system in conjunction with supercomputers such as "Fugaku".