Difficult to implement due to heterogeneous systems and programming fragmentation
Intel Delivers OneAPI Unified Environment to Prevent Fragmentation
ManycoreSoft supports HPC implementation from both hardware and software perspectives.


As the speed and volume of data collection rapidly increases, the demand for high-performance processors capable of quickly processing massive amounts of data continues to grow. IDC estimates that the total amount of data produced by humanity to date reached 33 zettabytes (ZB) in 2018, and this is projected to increase to 175 ZB by 2025.

The industry has been increasing the density per unit area to enhance processor performance. However, this approach already encountered limitations due to power consumption and heat generation in the early 2000s. This was overcome by multicore/many-core technology, which integrates multiple cores onto a single chip. The number of integrated cores continues to increase.

Beyond increasing core counts, other approaches are also being adopted. In fields requiring massive computational power, such as big data processing, deep learning, and machine learning, it's becoming increasingly common for general-purpose processors, efficient for complex, sequential tasks, to be supplemented by parallel processors, efficient for simple, repetitive tasks, in the form of accelerators.

"Heterogeneous systems" equipped with different types of processors, such as CPUs, GPUs, and FPGAs, are widely used in fields requiring high-performance computing (HPC), such as natural language processing, genome analysis, and financial forecasting. However, implementing them is not easy. Park Jeong-ho, CEO of Manicorsoft, explained that this was due to the fragmentation of the programming environment.

CEO Park Jeong-ho said, “Systems equipped with multiple processors require a variety of programming environments to coexist, making programming difficult and requiring significant effort for future program maintenance and management.”

If a system utilizes CPUs, GPUs, and FPGAs simultaneously, programmers must be able to write and manage code in pThreads and OpenMP for the CPU, CUDA and OpenCL for the GPU, and Verilog for the FPGA. Furthermore, if the system is divided across multiple nodes, they must also handle programming models like MPI, which require inter-node communication.

To combat programming fragmentation, Intel unveiled "OneAPI," a unified programming environment, later this year. Park explained, "OneAPI integrates Intel's previously released libraries and tools and provides the DPC++ programming model based on the standard heterogeneous parallel programming model, SYCL." He added, "This enables programming across a variety of heterogeneous processors."

OneAPI increases 'code portability' through the DPC++ programming model, which allows programs to be written using a single code to run on various types of accelerators. Additionally, performance portability has been improved by providing code optimized for the architecture of various accelerators in the form of a library.

CEO Park Jeong-ho emphasized, "Unlike SYCL, a heterogeneous programming model based on the C++ language, DPC++ includes the ability to write code that can run programs in parallel on accelerators." He also explained, "DPC++ provides unified memory, making data transfer between the accelerator and host memory easier to manage."

oneAPI is largely open-source software. CEO Park stated, "With future participation, it's possible and hopeful that hardware from other manufacturers, such as NVIDIA and AMD, will be supported."

◇ "Supporting HPC design capabilities in both hardware and software."

ManycoreSoft is a startup founded in 2012 in the Multicore Computing Lab at Seoul National University. It possesses hardware and software expertise in accelerators like GPUs and FPGAs, and provides solutions to address HPC challenges in diverse fields, including AI, finance, healthcare, manufacturing, and biotechnology. Currently, we design, manufacture and sell a water-cooled GPU system called 'DEEP Gadget'.

CEO Park Jeong-ho said, “We recently started a software project that automatically builds AI infrastructure that runs deep learning models efficiently by using all accelerators in the cluster,” and added, “The project supports various backend hardware architectures, and we expect OneAPI to play a key role in supporting various Intel hardware in this project.”

He also stated, “Manicorsoft plans to continue to work to ensure that its hardware and software know-how and capabilities in the HPC field can be utilized in various industries.”