oneDAL Arm SVE Enablement for Accelerated AI-ML Computing with FUJITSU-MONAKA

Fujitsu aims to revolutionise datacentre operations by offering a high-performance, energy-efficient Arm-based CPU, "FUJITSU-MONAKA," targeted for release in 2027. To achieve this goal, Fujitsu is focusing on application performance, ease of use, and open-source software. I will introduce the following contributions made to accelerate AI-ML algorithms by leveraging the Data Analytics Library (oneDAL) on the Arm architecture: 1) Fujitsu's efforts have led to the successful porting of oneDAL on Arm, leveraging SVE (Scalable Vector Extension) and porting enhancements with open-source optimised compute kernels from OpenBLAS. 2) This work resulted in notable performance gains of up to 40x speedup across multiple AI-ML algorithms, demonstrating the viability of Arm-based solutions for AI-ML workloads in datacentres. We acknowledge the NEDO Project “Technology Development of the Next Generation Green Data Centre” for the “Green Innovation Fund Project/Construction of Next Generation Digital Infrastructure." NEDO is the “New Energy and Industrial Technology Development Organisation," a national research and development agency in Japan. This presentation is based on results obtained from a project, JPNP21029, subsidised by the New Energy and Industrial Technology Development Organisation (NEDO).

MAD24-403 Optimising oneDAL and oneDNN for Arm CPUs to accelerate AI workloads

• Friday, 17 May 10:20 - 10:45
• Room: Session 1 | Las Palmas I

We at Fujitsu aim to deliver a high-performance, energy-efficient Arm-based CPU for data centres named FUJITSU-MONAKA. To achieve this goal, Fujitsu is focusing on application performance, ease of use, and open-source software. This abstract investigates the optimisation of AI workloads utilising the oneDNN and oneDAL libraries to accelerate AI workloads on Arm CPUs. Successfully ported oneDAL on Arm, leveraging SVE and incorporating enhancements with open-source optimized compute kernels of OpenBLAS, resulting in notable performance gains across multiple AI algorithms. Fujitsu has also contributed implementation of batch-reduced GEMM-based JIT kernels for MatMul and Convolution primitives, as well as expanding JIT kernels to support multiple ISAs, enabling them over more CPUs such as A64FX and Graviton3E. We acknowledge the NEDO Project “Technology Development of the Next Generation Green Data Centre” for the “Green Innovation Fund Project/Construction of Next Generation Digital Infrastructure." NEDO is the “New Energy and Industrial Technology Development Organisation," a national research and development agency in Japan. This presentation is based on results obtained from a project, JPNP21029, subsidised by the New Energy and Industrial Technology Development Organisation (NEDO)