Toward Open ISA Supercomputing: A Comparative Study of RISC-V, Arm, and x86 Architectures on HPC Workloads
AI Analysis Abstract This study evaluates the Sophgo SG2044 RISC-V processor against established Arm (NVIDIA Grace CPU Superchip) and x86 (Intel Xeon Platinum 8480+) architectures for High-Performance Computing (HPC) workloads. Using a consistent software stack and methodology,…
AI Analysis
Abstract
This study evaluates the Sophgo SG2044 RISC-V processor against established Arm (NVIDIA Grace CPU Superchip) and x86 (Intel Xeon Platinum 8480+) architectures for High-Performance Computing (HPC) workloads. Using a consistent software stack and methodology, the research characterizes the SG2044 across synthetic benchmarks (STREAM, HPL, HPL-MxP, HPCG) and scientific/AI applications (iPIC3D, PLUTO, OpenFOAM, llama.cpp). While the SG2044 exhibits 8-15x lower absolute performance due to narrower vector resources and a less advanced technology node, it demonstrates competitive scaling and energy efficiency on several workloads. Notably, in LLM inference decode, the 64-core SG2044 achieves superior relative scaling and energy efficiency compared to both Intel and NVIDIA, even surpassing Intel in throughput. The paper also highlights significant generational improvements within the Monte Cimone RISC-V project, with MCv3 (SG2044) showing up to 158x higher FP64 performance and 10x higher energy efficiency over MCv1. The findings suggest that RISC-V processors are progressively closing the performance gap in diverse HPC scenarios.
Key Highlights
- Extended characterization of the Sophgo SG2044 RISC-V processor across HPC benchmarks, scientific and industrial applications.
- SG2044 reaches 294 GFLOP/s FP64 and 107 GB/s memory bandwidth at 64 cores, with L2-aware pinning improving scalability.
- SG2044 matches or exceeds Sapphire Rapids energy efficiency across several workloads and outperforms NVIDIA Grace in LLM inference at 64 cores, reaching 1.20 tok/J (PP) and 0.209 tok/J (TG).
- Monte Cimone evolution in four years demonstrate up to 158x higher FP64 performance and 10x higher energy efficiency from MCv1 to MCv3.
- Performance analysis identifies vector width, memory bandwidth, and software-level affinity as key factors for improving RISC-V competitiveness in HPC.
Technical Details
- The study compares Sophgo SG2044 (RISC-V, 64 cores, RVV 1.0 128-bit vector, TSMC 12nm), NVIDIA Grace CPU Superchip (Armv9, 144 cores, SVE2 4x128-bit vector, TSMC 4N), and Intel Xeon Platinum 8480+ (x86-64, 112 cores, AVX-512 2x512-bit vector, Intel 7).
- A common software stack was used, including Linux kernel 6.12, GCC 14.2, OpenMPI 5.0, and OpenBLAS 0.3.34 with architecture-specific optimizations.
- The SG2044 achieved 294 GFLOP/s FP64 performance and 107 GB/s memory bandwidth at 64 cores, with L2-cache-aware process placement improving STREAM scaling up to 16 cores.
- In HPL, the SG2044 reached 8.02 GFLOP/s single-core and 294 GFLOP/s at 64 cores, while NVIDIA Grace achieved 2,818 GFLOP/s at 144 cores and Intel Sapphire Rapids 5,210 GFLOP/s at 112 cores.
- For LLM inference (llama.cpp, gpt-oss-20b-F16, batch size 512), the 64-core SG2044 reached 17.4 tok/s in Text Generation (TG), outperforming Intel Sapphire Rapids (14.0 tok/s) and achieving higher energy efficiency (0.209 tok/J vs. Grace's 0.159 tok/J).
- Power measurements were collected via IPMI telemetry (x86, Arm) and architecture-specific mechanisms (RAPL for Intel, hwmon for NVIDIA); for SG2044, IPMI reported board-level power, necessitating engineered combined metrics for a common power domain.
- The Monte Cimone project's evolution from MCv1 to MCv3 (SG2044) demonstrated a 158x increase in FP64 performance and a 10x improvement in energy efficiency.
Implications
- RISC-V has made substantial progress in HPC, demonstrating competitive system-level characteristics, particularly in scaling and energy efficiency for certain workloads.
- The performance of RISC-V platforms is not solely dependent on peak computational capability, with memory bandwidth and software affinity playing crucial roles.
- Future RISC-V HPC platforms can enhance competitiveness by improving vector capabilities, advancing memory subsystems for higher aggregated bandwidth, utilizing more advanced technology nodes, and continuing software stack and numerical library optimizations.
- The generational improvements observed within the Monte Cimone project suggest a viable path for RISC-V to progressively reduce the performance gap with established HPC architectures while retaining the benefits of an open ISA.
Limitations
- The Sophgo SG2044 processor remains 8-15x behind the evaluated Intel and NVIDIA platforms in absolute performance across classical HPC workloads.
- The performance deficit is attributed to the SG2044's narrower vector resources (128-bit RVV 1.0) and less advanced technology node (TSMC 12nm) compared to x86 (AVX-512 on Intel 7) and Arm (SVE2 on TSMC 4N).
- SG2044's memory bandwidth shows saturation around 16 cores, limiting further gains in STREAM, despite improvements over previous generations.
- HPL scaling efficiency on the SG2044 at higher core counts is less efficient than the previous-generation SG2042, even with higher absolute performance.
- Power measurement for the SG2044 relied on board-level IPMI data, requiring engineered combined metrics to account for components like fans, which may introduce estimation variability compared to direct processor-level measurements on other platforms.
- The OpenFOAM version used (v2512) employs a native LDU-COO sparse matrix format that is ill-suited for vectorization, limiting compiler auto-vectorization and resulting in strictly scalar execution within each core, which may not fully leverage vector-capable architectures.
- The HPL-MxP implementation was a custom build by the authors, linking against OpenBLAS, which might not reflect performance achievable with vendor-optimized HPL-MxP implementations on other platforms.
Source: https://arxiv.org/abs/2610.32632