Get the Latest on All Things CODE

Sign Up
author-image

By

/content/dam/www/central-libraries/us/en/images/2023-05/headshot-laura-reznikov-rwd.png.rendition.intel.web.480.480.png

Laura Reznikov

Senior Engineering Manager,  Intel Advanced Ray Tracing Technologies group
Twitter* | LinkedIn*

Intel Corporation

The computer graphics marketplace is even more exciting now with devices based on the Intel® Data Center GPU Flex Series becoming broadly available. Designed to make streaming video and media better and more efficient through its hardware-based AV1 encoder and accelerated AI visual processing and ray tracing with dedicated AI silicon, this new GPU powerhouse delivers outstanding compute density and energy efficiency for visual cloud workloads. What’s even more exciting for me beyond the GPU awesomeness—is how software will bring to life super-cool usages, and the platform’s open software stack brings flexibility and ease in driving these innovations.

A historical shortcoming of prior compute accelerators (GPUs) was that each new accelerator needed a software development toolchain to unlock benefits and deliver value–and importantly, those toolchains tended to lock developers into a single vendor and/or microarchitecture. Along with new hardware, these new GPUs pair with oneAPI developer libraries and languages that are open standards-based—and deliver performance on not only Intel architecture, but also those of other manufacturers. This approach enables greater choice for data center operators to choose new hardware—removing the barriers of proprietary programming models. It also provides software developers with flexibility to customize solutions and innovate freely for the performance and efficiency they need on the platform of their choice while gaining development productivity and reduced costs with oneAPI’s single-source codebase via SYCL for CPUs and GPUs.

What Comprises This Open Software Stack?

To effectively realize the hardware capabilities for virtual cloud workloads, Flex Series is paired with a comprehensive software stack to enable the ecosystem to deliver accelerated applications and services.

The Flex Series GPU open software stack for visual cloud workloads includes a mix of open software components and tools that developers can use to build, customize and differentiate their solutions.1,2

Learn more: Software for Intel® Data Center GPU Flex Series

The stack components are designed to work together with the built-in hardware capabilities to help developers build high-performance, optimized, multiarchitecture applications for visual cloud workloads. This includes:

  • Media Delivery – Developers can improve media performance at a lower cost by accessing the AV1 hardware encoder on the Flex Series GPU using open source technologies such as FFmpeg, GStreamer, Handbrake, and oneAPI Video Processing Library (oneVPL). Using the AV1 encoder can provide a compression improvement of up to 30%, potentially saving media processing and delivery providers significant costs.3
  • Cloud gaming4 – Cloud gaming developers must continually innovate to deliver first-rate playing experiences while operating an efficient infrastructure. They benefit from an ecosystem-enabled, flexible software stack that supports high-density streaming delivery, high-quality gameplay, and a breadth of gaming title accessibility. Combined with Flex Series, developers can deliver a high density of gaming instances per server. Workloads support Android* cloud gaming2 and Windows* cloud gaming.
  • AI visual inferencing – Taking advantage of Intel’s open software stack, visual inferencing and media analytics has never been easier. Using standard frameworks such as Gstreamer and FFmpeg combined with OpenVINO™ toolkit, developers can save time with Intel provided examples and models for workloads such as object detection or classification. Developers can also create their own visual pipelines using preconfigured containers for Tensorflow*, PyTorch* and others.
  • Media acceleration – I’m pleased to see VMware Horizon is already providing media acceleration utilizing Flex Series' AV1 encoding and decoding, optimized by oneVPL. It’s a solution with no virtualization license server setup, licensing or ongoing support costs providing VMware customers more choice, flexibility and cost options on an open platform, and super media experience.
  • Digital content creation – With the latest release, creators can get fast, real-time ray tracing via dedicated hardware acceleration through Intel® Embree5. And high-quality AI-based denoising can be completed in milliseconds using the latest Intel® Open Image Denoise5. SYCL and oneAPI enable a single codebase to take advantage of these libraries both on the CPU and GPU.

Why Open Development Is Better

Since its inception, oneAPI was developed in the open, accepting community input. Open development at oneapi.io enables community members to deliver their own hardware and software innovation, as well as improve and evolve the specification. oneAPI does not rely on a siloed and closed ecosystem.

Modern software developers are already familiar with the benefits of open, standards-based multiarchitecture programming methods that reduce the development complexity, cost, and time requirements to bring new solutions to market. These methods enable engineers and programmers to scale their solutions and spend more time innovating and building differentiating features rather than maintaining multiple codebases.

We also worked to make code portability easier through the open source SYCLomatic project that helps migrate proprietary CUDA* applications to the open-standard and cross-platform SYCL* language.

Building a Thriving Ecosystem

This is just the beginning. Extending the benefits of the Flex Series GPU optimized software stack untethers graphics programming from restrictive proprietary environments and draws on a broad ecosystem of service providers, independent software vendors (ISVs), original equipment manufacturers (OEMs), and others supporting the most popular workloads for a wide range of use cases. The openness and transparency of this approach encourages uptake by the open source community, creating a virtuous cycle to further enhance the software stack. Join the oneAPI community as the ecosystem grows and capabilities evolve.

Getting Started: Development Ease Through Pre-Configured Containers

The software stack is ready to go. Its common set of capabilities integrate easily into popular middleware and frameworks, and the stack is delivered in validated ready-to-run containers. The containers can be orchestrated with Kubernetes on bare metal, or in virtual machines (VMs) using SR-IOV virtualization with tools to assign and manage workloads. This includes Node Feature Discovery, that labels the nodes with GPU support and the Intel GPU device plugin for Kubernetes, which helps facilitate the assignment of full or fractional GPU resources to a workload. The toolset is designed to speed development and enable flexible deployment of multiple workloads on the same GPU. 

Get started: Download the containers now

New market entries bring new ideas that challenge the status quo and provide ecosystems with healthy diversity. These data center GPUs based on open development should provide just this kind of spark. I can’t wait to see more exciting solutions that developers will create using Flex Series and importantly through its open developer software.

Laura

Product and Performance Information

1oneDNN is the oneAPI Deep Neural Network Library. oneDAL is oneAPI Data Analytics Library. oneVPL is the oneAPI Video Processing Library. Intel created open source components are also included in Intel’s oneAPI toolkits and other tools: oneVPL, oneDNN, oneDAL, Intel VTune Profiler and the DPC++/C++ Compiler are part of Intel® oneAPI Base Toolkit; Intel-optimized TensorFlow and PyTorch are part of Intel® AI Analytics Toolkit; WebRTC is in Intel® Collaboration Suite for WebRTC.
2For Android cloud gaming, Intel licenses additional software components that work with the game streaming solution to support high server game stream density, game performance, and breadth of gaming title accessibility.
3Cost savings based on streaming x264 @ 4 Mbs (1.21 GB/hr) and 30 percent bit rate efficiency for AV1 @ 2.7 Mbs (0.81 GB/hr). Host hardware: Intel® Xeon® Platinum 8360Y processor (2.40 GHz base frequency, 3.50 max turbo frequency; 64 GB RAM (4x 16 GB DIMM DDR4-3200); 2x Intel® Data Center GPU Flex Series cards (Device 56c1 rev 05, model: 75W 128EU); Ubuntu 20.04.4 LTS (focal), Kernel: 5.10.54+prerelease2828 (agama-ci-prerelease-366); display driver version: 22.2.1.10510-main OneVPL: libvpl-dev version 2022.0.3.0.10510-main.
4Reflects capabilities of Intel Data Center GPU Flex Series that will be available when product is fully mature.
5Applies to Intel Embree and Intel Open Image Denoise open source versions available on GitHub. These versions will be added to the Intel® oneAPI Rendering Library in its next release.
6

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.