Research team builds supercomputer using SiFive's RISC-V processor

A research team from the University of Bologna and CINECA (Italy's largest supercomputing center) has been researching and developing RISC-V supercomputers. Recently, the new ISA designed by the team has demonstrated the ability to run high-performance computing, which lays the foundation for building supercomputers. The team used SiFive's Freedom U740 SoC as the basis, and the researchers named their RISC-V cluster "Monte Cimone".

To create a supercomputer, you need hardware that looks like Lego bricks. These are called clusters and consist of motherboards, processors, memory, and storage. Italian researchers decided to try a different solution to the problem than Intel/AMD and use processors based on the RISC-V ISA.

Monte Cimone has four dual-board servers, each in a 1U form factor. Each board has a SiFive Freedom U740 SoC, which has four U74 cores running at up to 1.4 GHz and an S7 management core. There are eight nodes in total, for a total of 32 RISC-V cores.

Paired with 16GB of 64-bit DDR4 memory running at 1866s MT/s, a PCIe Gen 3 x8 bus running at 7.8 GB/s, a Gigabit Ethernet port, USB 3.2 Gen 1 interface, the system is powered by two 250-watt PSUs to support future expansion and the addition of accelerator cards.

The Italian team benchmarked the system using HPL and Stream to determine the machine's floating-point computing power and memory bandwidth. While the results aren't very impressive, they're a start for RISC-V.

Each node produces a sustained 1.86 GFLOPS of performance in HPL, for a total of 14.88 GFLOPS of compute power, with perfect linear scaling. However, the overall cluster is 85% efficient, with 12.65 GFLOPS of compute power. The nodes should achieve 14.928 GB/s of memory bandwidth; however, the actual result is 7760 MB/s.

These results show two things. First, the RISC-V HPC software stack has matured, but further optimization and faster chips are needed to achieve major tasks like weather simulations. Second, it shows that scaling in the HPC world is very tricky and requires careful optimization to make hardware and software coexist in a world where everything scales well.