Over three days in early December, industry experts and attendees explored the latest advancements in cross-architecture programming, artificial intelligence (AI), high performance computing (HPC), and beyond at the oneAPI DevSummit for AI and HPC 2023.
Speakers from esteemed organizations presented engaging keynotes, tech talks and tutorials, covering AI and analytics using industry-standard frameworks and tools, finely tuned for performance by oneAPI, plus insights into the Intel® AI Analytics and HPC Toolkits, as well as comprehensive overviews of key performance analysis tools, complete with practical guidance on their utilization and where to obtain them, and hands-on workshops utilizing the Intel Developer Cloud.
Democratizing the use of AI: FUJITSU – MONAKA
Director of Software Engineering at Fujitsu Research of India Pvt Limited (FRIPL), Dr. Priyanka Sharma, addressed the challenge of accessibility in AI’s rapid evolution. Fujitsu aims to democratize AI through global collaboration for sustainable digital transformation.
Sharma discussed Fujitsu’s core technologies, spanning R&D in computing, HPC, networking, AI, deep learning, and data security. Fujitsu focuses on a computing workload broker supporting quantum simulators, AI accelerators, and Monaka—a oneAPI-enabled two-nanometer ARM-based CPU for efficient edge-level acceleration and green data center computing.
Highlighting Fujitsu’s supercomputing legacy with Fugaku, Sharma showcased its applications in disaster prevention, tsunami prediction, cancer genome analysis, and more. Fujitsu’s commitment to open-source software was emphasized, with contributions to projects like Linux kernel, KVM, OpenStack, and automotive-grade Linux.
Fujitsu’s collaboration with oneAPI, specifically with oneDNN, involved successful porting to Fugaku with ARM architecture, demonstrating performance improvements in models like ResNet.
As a key member of the UXL Foundation Steering Committee, Sharma concluded with Fujitsu’s commitment to coexisting with evolving technologies for sustainable digital transformation and a commitment to open source software.
Enhancing Aviation Predictive Maintenance with oneAPI: NASA’s Turbofan Engine Degradation Simulation
PhD student and Intel Certified Instructor in Machine Learning using oneAPI, Deepthi A J, delved into NASA’s Open Data Portal for Earth science, climate monitoring, and disaster management.
A J applied oneAPI’s unified programming model, combining sklearn, XGBoost, and daal4py libraries to optimize machine learning algorithms, focusing on predicting the remaining useful life of aircraft engines using NASA’s Turbofan Engine degradation simulation. The project’s broader applicability was emphasized, addressing challenges in obtaining suitable datasets for prognostics and health management.
A J discussed the Prognostics and Health Management Data Challenge (2019) and positioned oneAPI as a solution for power optimization crucial for space missions. The speaker tackled data-driven prognostics challenges, introducing the CMAP dataset for simulating realistic large commercial turbofan engines. A J detailed the CMAP system model, emphasizing fault and deterioration simulation in key components.
Transitioning to real prediction with oneAPI, A J demonstrated deployment using TensorFlow, scikit, and other libraries, showcasing performance improvements and CPU acceleration. The presentation concluded with project results, underscoring oneAPI’s advantages in speedup and power optimization.
A J encouraged exploration of oneAPI for enhanced performance, especially in critical applications like space missions, and shared a personal GitHub repository and references for further exploration.
Performant Portable HPC Applications with SYCL and oneMKL Interface
Hugh Bird, Staff Software Engineer at Codeplay Software Ltd, revealed how the oneMKL interface library optimizes GROMACS for heterogeneous computing. By utilizing SYCL implementation, oneMKL ensures peak performance of vendor-optimized DFT libraries in tasks like Discrete Fourier Transforms (DFTs). The talk emphasized oneMKL’s seamless synergy and its ability to minimize overhead through SYCL backend interoperability, achieving performance without sacrificing portability.
Bird discussed Codeplay’s commitment to expanding the open ecosystem for heterogeneous computing, particularly in its role with the UXL Foundation and oneAPI. The talk delved into programming heterogeneous devices using GROMACS, highlighting oneMKL as a solution for simplifying code maintenance and enhancing portability by unifying FFT backends across different hardware vendors.
Using GROMACS as a case study, Bird illustrated how SYCL and oneMKL streamline development and maintenance, reducing the need for multiple implementations for different hardware. The presentation emphasized the computational expense of molecular dynamics and the significance of FFTs in accelerating physics problems. oneMKL was positioned as a solution for addressing the challenges of maintaining code across various GPU programming interfaces and FFT libraries.
Performance comparisons on NVIDIA and AMD hardware showcased SYCL and oneMKL’s excellent portability without sacrificing speed. Bird recommended SYCL and oneMKL for individuals like physics PhD students who lack access to specialized hardware.
The presentation concluded by encouraging adoption of SYCL and oneMKL, highlighting advantages such as writing less code, reducing maintenance efforts, and enjoying portability and performance across different hardware vendors. The open-source nature of the code invites collaboration and feedback from the community.
SYCL Support for Continental-Scale Ecological Observations: Scalable and Portable Blending of Massive Image Mosaics
Utah State University Assistant Professor Steve Petruzza presented on the latest developments with The National Ecological Observatory Network (NEON), a continental-scale facility using Airborne Observation Platforms (AOPs) to study US ecosystem changes. Processed through OpenViSUS, this open-access data spans hundreds of terabytes. NEON overcame challenges in adapting proprietary GPU-designed components using SYCL, Intel MKL, and the DPC++ Compatibility Tool, achieving scalability, portability, and reduced development efforts.
Petruzza discussed the collaboration between the University of Utah and Lawrence Livermore National Lab in the oneAPI Centre of Excellence. NEON’s goal is to collect and distribute data, including sensor measurements and high-resolution images. An interactive viewer was introduced to explore the dataset through NEON’s production data portal.
The workflow involves a REST API, cloud services, data conversion, and integration with the NEON data portal. Petruzza addressed challenges in mosaics, adopting a gradient domain image processing approach to smooth out visible seams. A CUDA implementation was converted to run on oneAPI using the DPC++ Compatibility Toolkit, showing good portability across different platforms, including CPUs and GPUs. The conversion’s success enables algorithm use on machines with larger memories, even without GPUs.
In a tech talk titled “Development and Optimization of a SYCL Backend for libCEED,” Argonne National Laboratory and Intel engineers discussed adapting the libCEED library for efficient operation on the Aurora supercomputing platform. The Aurora system uses CPU Max and GPU Max architecture, and the presentation covered challenges and strategies in the adaptation process. The speakers introduced libCEED, emphasizing its capabilities, design philosophy, and runtime compilation aspect using the SYCL online compiler for dynamic code generation.
The talk focused on the optimization process, addressing experiences with SYCL programming, including specialization, constant work, group barriers, and SIMD size tuning. Challenges during optimization, such as using the oneAPI online compilation extension, were discussed. Key takeaways included the use of SYCL online compilation, optimization strategies, and ongoing profiling efforts using tools like Vtune and Advisor from the oneAPI toolkit.
In another presentation, “Building a Portable, Scalable, Performant ZFP Backend Using oneAPI and SYCL,” PhD student Alper Sahistan from the University of Utah explored developing a portable, scalable CFPB library using oneAPI and SYCL. Emphasizing the need for robust compression methods in memory-bound applications, Sahistan discussed experiences using oneAPI and SYCL to enable ZFP for a broader user base. The talk covered the DPCT migration tool’s role in converting CUDA code to SYCL, addressing challenges in code adaptation and reported issues with CUDA header files.
Sahistan delved into debugging using GDB via oneAPI, addressing challenges and limitations. The presentation discussed efforts to measure performance using kernel timing utilities and shared solutions to overcome inconsistencies and difficulties with timing measurements. Results from encoding and decoding throughput measurements on various devices highlighted observed behaviors and performance differences.
Sahistan concluded by summarizing the successful transformation of the CFP library, recommending improvements, and outlining future work on variable rates and optimizations for the SYCL port.
A Pivotal Moment that Shaped 2023 for the oneAPI Community
This past September, the Linux Foundation introduced the Unified Acceleration (UXL) Foundation, a collaborative initiative focused on establishing an open standard accelerator programming model.
The UXL Foundation aims to simplify the development of high-performance, cross-platform applications by bringing together industry participants to create a unified programming model. It builds on the success of the oneAPI initiative, which covers diverse architectures like CPU, GPU, FPGA, and accelerators.
Hosted by the Linux Foundation’s Joint Development Foundation, the UXL Foundation involves key industry players such as Arm, Fujitsu, Google Cloud, Imagination Technologies, Intel, Qualcomm Technologies, Inc., Samsung, and VMware. The foundation’s mission includes building a multi-architecture multi-vendor software ecosystem for all accelerators, unifying the heterogeneous compute ecosystem around open standards, and then building on and expanding open-source projects for accelerated computing.
The formation of the UXL Foundation is seen as a crucial step in advancing innovation and implementing the oneAPI specification across the industry.
Get Involved with oneAPI
You can watch all the enlightening sessions held at the oneAPI DevSummit for AI and HPC 2023 here to get a comprehensive and immersive learning experience that will enrich your professional growth and development.
The UXL Foundation is an evolution of the oneAPI initiative and focuses on the development of a specification and open-source projects through Working Groups (WGs) and Special Interest Groups (SIGs) on Hardware, Language, Math, Safety Critical, and also AI, which host discussions and presentations focused on the functions and interfaces needed to enable applications using machine learning and deep neural networks alongside more traditional algorithms for AI.
The oneAPI specification defines oneDNN (w/ Graph), oneCCL, and oneDAL API as building blocks for deep learning applications and frameworks.