Intel today provided greater detail around its plans to bring a full line of GPUs (Xe) and associated programming environment to market. The biggest news from an HPC perspective was introduction of oneAPI Gold, the first productized version of Intel’s programming platform for the Xe GPU line. On the hardware side, Intel added detail to its plans for offering distinct versions of its GPUs and introduced a video streaming GPU solution. Last month, Intel introduced Iris Max – the first discrete GPU in the Xe line.

The top of the Xe line, Ponte Vecchio, is still under development. It will be used in Aurora, one of the first U.S. exascale supercomputers.

The announcements came during the opening week of SC20 and coincide with Intel’s first oneAPI Summit. This year, SC is a virtual event spanning two weeks with the first week devoted to workshops and tutorials and the second week featuring keynotes, invited talks, panels, and awards announcements. HPCwire will announce the winners of its annual Readers’ and Editors’ Choice Awards next week as well.

For Intel, the plunge back into GPUs seems part of a broader embrace of heterogeneous computing going forward. Given the range of devices, home-grown and acquired, that are now part of the Intel portfolio (CPUs, FPGAs, GPUs, specialized DNN accelerators, etc.) Intel’s emerging ‘XPU’ strategy is to address functionalities across all of these device types. Intel says the need to conquer or at least tame the software Tower of Babel among them is critical. OneAPI, Intel hopes, can evolve to ease the programming and communications challenges across its full portfolio of device architectures.

Raja Koduri, Intel SVP, chief architect, and GM of Intel architecture, graphics, and software, said during a media pre-briefing, “We now live in a time where software is running the world. And we (Intel) are going through a massive culture and mindset change in order to shift to a ‘software first’ approach to hardware architecture and design. Today, you will see the first fruits of our CPU-to-XPU and software-first initiatives in product form. [The] Xe architecture is foundational to our XPU strategy and oneAPI is the essential software foundation to enable 20 million developers productively across our XPU architecture.”

OneAPI is presented by Intel as an open standard. It will be interesting to watch how other vendors respond since oneAPI was developed specifically to make optimum use of Intel’s underlying architecture. That said, many observers in the user community think the idea has merit if oneAPI could make their lives less complicated.

Intel said oneAPI will be available in December including being freely available on Intel’s Developer Cloud for use with Intel hardware including select versions of the Xe.

Jeff McVeigh, Intel VP, datacenter XPU products and solutions, said “We’re bringing the oneAPI product to market through a set of toolkits. The first one is really the most foundational one, we call it the base toolkit. This provides the core libraries and tools to allow you to do cross-architecture development of applications, frameworks and middleware. It provides that compatibility tool to go from CUDA to data, parallel C++ language support, building upon C++ for the data parallel C++ language and compiler, as well as optimized Python. [Also included is] a set of domain libraries covering math, parallel runtimes, artificial intelligence, video processing, that enable developers to quickly get the best performance and cross architecture support. There is also a familiar set of analysis and debug tools that have been enhanced to support multiple architectures.”

“The HPC toolkit, as the name implies, is for high performance computing, and includes the C++ compiler with OpenMP support Fortran compiler, cluster checker, MPI libraries, and other key capabilities necessary for deploying HPC applications at scale. The IoT toolkit really enables developers that are deploying on the edge and in network, the devices necessary for power efficient performance,” said McVeigh.

On the hardware side, Intel provided a brief, less detailed update. Its Iris Max GPU is one of the entry Xe LP (presumably low power) devices. The roadmap includes LP, HPG, HP, and HPC devices and Koduri described their overall targets.

“We’ve always said the first step to our product strategy was Xe LP,” Koduri said. “This is the foundation on which we built the rest of our roadmap. Last month we launched our first discrete GPU in more than two decades – the Intel Iris Xe Max graphics – with Deep Link technology focused on all important mobile creators segment. Xe HPG is for gamers and we are happy to report that the first GPU based on this architecture successfully [working] in our labs.

“Xe HP is our first datacenter scale GPU and is also our first multi-tiled scalable and high performance architecture. We demonstrated this GPU at our [recent] architecture day delivering over 40 teraflops of general purpose, FP32 performance and media streaming capabilities that will enable our customers to disrupt form factors for media racks in datacenters. Ponte Vecchio is our first Xe HPC-based XPU. This is our most ambitious project that combines many advanced technologies across all our six pillars. We have successfully released the first revisions of all chiplet designs to manufacturing and are eagerly awaiting their arrival to power on,” he said Koduri. (Intel’s six pillars: process and packaging; XPU architectures; memory and storage; interconnect; software; security)

Intel also drew attention to the XPU/oneAPI expanding ecosystem (excerpt from Intel literature):

• Argonne National Laboratory: Researchers at the U.S. Department of Energy’s Argonne National Laboratory are using Intel oneAPI Toolkits to test code performance and functionality using programming models that will be supported on Aurora. Aurora is set to be one of the nation’s first exascale systems and will be used to dramatically advance scientific research and discovery.
• Codeplay builds oneAPI support: Codeplay Software announced the first release of its Data Parallel C++ (DPC++) compiler for Nvidia GPUs.
• University of Illinois (UI): The Beckman Institute for Advanced Science and Technology at UI today announced a new oneAPI center of excellence (CoE). They are bringing the oneAPI programming model to life sciences and application NAMD to additional heterogeneous computing environments. NAMD, which simulates large biomolecular systems, is helping to tackle real-world challenges such as COVID 19.
• Heidelberg University Computing Center (URZ): URZ announced it is establishing a oneAPI CoE focused on bringing oneAPI support to AMD GPUs.
• Swedish e-Science Research Center (SeRC): Hosted at Stockholm University and the KTH Royal Institute of Technology, the SeRC’s oneAPI academic CoE is using oneAPI’s unified and heterogeneous programming model to accelerate research conducted with GROMACS, a widely used free and open-source application designed for molecular dynamics simulation.

McVeigh cited work at the Zuse Institute, Berlin: “ZIB took their commercial application, EasyWave, a tsunami simulation, utilizing the oneAPI product in particular – the compatibility tool – to port the implementation in CUDA to data parallel C++, and then run that same data parallel C++ code across a variety of architectures including Xeon CPUs XE GPUs, Stratix FPGAs, as well as back on the Nvidia GPU. [In] this example, the university was able to achieve 95 percent of the performance using DP C++ on Nvidia platforms relative to the original CUDA source. This really shows you the value and the capabilities of across architecture solution to provide productivity as well as performance.”

As part of the flurry of announcements Intel also launched “Server GPU” aimed at video streaming generally and Android gaming specifically.

“According to the Cisco VNI (Visual Networking Index), video will account for 82 percent of the global IP traffic by 2022. Live video will increase 15-fold from 2017 to 2022 to reach 17 percent of internet video traffic. Over that same time period, gaming traffic is expected to grow nine-fold. In response, new participants are coming to these market segments with companies such as phone manufacturers and OEMs joining along with cloud and communication service providers to vie for consumer attention. In fact, the game development ecosystem is increasingly focused on Android with more than 74 percent of the global mobile market share,” said Lynn Comp, Intel VP, data platforms group and GM, visual infrastructure division.

Intel’s new answer is the Server GPU solution.

“The Intel Server GPU is based on a low power discrete system on chip design, with a 128-bit wide pipeline, and eight gigabytes of dedicated onboard low power DDR4 memory. Four Server GPUs are packaged together in a three-quarter length full-height x 16 PCIe Gen three add-in card, with a target configuration of up to four cards per server,” said Comp.

She ran a demo with the popular video game Arena of Valor. “This is just one example. Typically, the HTC XG 310 card can support more than 100 simultaneous users and a two-card system includes Intel Xeon scalable processors and up to 160 simultaneous users. We have many software and services partners launching our solution, including Tencent games which is introducing their new Android cloud gaming service,” said Comp.

While not especially HPC-oriented it wouldn’t be at all surprising to see these cards put to varied use by inventive users.

Link to Intel announcement: https://newsroom.intel.com/news/intel-xpu-vision-oneapi-server-gpu/#gs.kni86m