48% of developers target heterogeneous systems using more than one type of processor. Therefore, more efficient multi-architecture programming is needed to address the growing scope and scale of real-world workloads. Code written for proprietary programming models like CUDA lacks portability to other hardware and can lead to isolated development environments that lock organizations into closed ecosystems.

Developers can use Intel's standards-based multi-architecture tools and oneAPI, an open, unified programming model, to achieve the freedom to choose hardware, performance, productivity, and code portability for CPUs and accelerators.

Intel unveiled its new Intel oneAPI 2023 tools on the 20th. They support 4th-generation Intel Xeon Scalable processors, as well as Intel data center GPUs, including the Xeon CPU Max series, Flex series, and the new Max series.

Intel oneAPI 2023 supports a new Codeplay plugin that allows developers to easily write SYCL code for non-Intel GPU architectures, along with improved performance. Intel's new standards-based tools will provide users with greater hardware choice and improve the ease of developing high-performance applications that run on multi-architecture systems.

The newly released Intel 2023 developer tools include the latest compilers, libraries, analysis and porting tools, and optimized artificial intelligence and machine learning frameworks required to build high-performance, multi-architecture applications for CPUs, GPUs, and FPGAs powered by oneAPI.
NVIDIA demonstrated 2.4x higher performance on AMD products in the MLPerf DeepCAM deep learning inference and training performance benchmarks, while Intel Xeon CPU Max with Intel AMX based on Intel oneAPI Deep Neural Network Library (oneDNN)2 achieved 3.6x higher performance. Additionally, the performance of the Large-Scale Atomic and Molecular Parallel Simulator (LAMMPS) workload, offloaded to six Max-series GPUs and optimized with a single oneAPI tool, running on the Intel Xeon Max CPU achieved up to 16x higher performance compared to 3rd Gen Intel Xeon or AMD Milan.

Additionally, the adoption of OneAPI within the ecosystem is increasing, with the establishment of several new centers of excellence. One such center is the University of Cambridge's Open Zettascale Lab, which focuses on porting key exascale candidate code to OneAPI, including CASTEP, FENiCS, and AREPO. The center offers workshops with experts who teach API methodologies and tools for compiling, porting, and optimizing code performance. Currently, a total of 30 oneAPI Centers of Excellence have been established.

“We’ve seen initial application performance improvements on development systems using Intel Max Series GPU accelerators,” said Timothy Williams, vice president of Argonne’s Computational Sciences division. “These applications are built on the Intel oneAPI compiler and libraries.”

“For leadership-grade computational science, we value the benefits of code portability across multi-vendor, multi-architecture programming standards, such as Python AI frameworks like SYCL and PyTorch, accelerated by Intel libraries,” he said. “Building on this technology, we look forward to achieving first-of-its-kind scientific discoveries on the Aurora system next year.”

Meanwhile, the tool will be available through the Intel Developer Cloud and official retail channels.