Intel reports that Edinburgh, UK-based Codeplay has released an extension to its oneAPI ecosystem, the oneAPI Construction Kit. The open source project allows code written in Sycl to run on custom architectures for HPC and AI. The kit includes a framework for bringing oneAPI support to new hardware including AI accelerators.
Intel has opened up a ‘training portal’ for those wishing to migrate CUDA code to C++/Sycl and ‘unlock their code from the constraints of vendor-specific (read Nvidia) tools and accelerators’.
In a separate announcement Intel demonstrates how its Open HPC+AI portfolio powers ‘generative AI for science’. In a presentation made at International Super Computing 2023 Intel demoed ‘AI-accelerated HPC performance driven by oneAPI’. With help from ISV Ansys, the Intel Data Center GPU Max Series was claimed to outperform NVIDIA H100 by 50% on AI-accelerated HPC applications, in addition to an average improvement across diverse workloads of 30%. ‘The Habana Gaudi 2 deep learning accelerator delivers up to 2x faster AI performance over an Nvidia A100 for deep learning training and inference’.
Chinese oil country software developed GeoEast has reported use of the oneAPI dev kit to provide a ‘solid high-performance base’ to its geophysical exploration software, speeding pre-stack migration by a claimed 1,000x. Researchers at the Federal University of São Carlos, Brazil have also leveraged oneAPI to run seismic workloads on FPGAs. Here applications were accelerated using SYCL* with oneAPI for FPGA frameworks. However, Intel warns, your performance and mileage may vary ‘by use, configuration and other factors’. More from Intel.
© Oil IT Journal - all rights reserved.