Intel’s Next Gen AI Mastery Turbocharges Everything

Intel’s Next Gen AI

Intel launched 5th Gen Intel Xeon and Intel Core Ultra CPUs for data center, cloud, and edge next gen AI at its AI Everywhere event on December 14.

Intel uses a software-defined, open ecosystem to make next gen AI hardware technologies accessible and easy to utilize. That includes incorporating acceleration into next gen AI frameworks like PyTorch and TensorFlow and providing core libraries (through oneAPI) to make software portable and performant across hardware.

The comprehensive set of enhanced compilers, libraries, analysis and debug tools, and optimized frameworks in Intel Software creation Tools 2024.0 simplifies the creation and deployment of accelerated solutions on these new platforms, maximizing performance and productivity.

AI accelerator engines

5th Gen Intel Xeon processors can handle demanding next gen AI workloads without discrete accelerators, building on 4th Gen’s built-in accelerator engines and providing a more efficient way to increase performance than increasing CPU cores or GPUs.

Intel Accelerator Engines:

• Intel AMX enhances deep learning training and inference. It excels at NLP, recommendation systems, and picture recognition.
• Intel QuickAssist Technology (Intel QAT) offloads encryption, decryption, and compression to free up CPU cores so computers may serve more customers or use less power. Fourth-generation Intel Xeon Scalable processors with Intel QAT are the fastest CPUs that can compress and encrypt simultaneously.
• Intel Data Streaming Accelerator (Intel DSA) improves streaming data transportation and transformation processes for storage, networking, and data-intensive workloads. Intel DSA speeds up data movement across the CPU, RAM, caches, all associated memory, storage, and network devices by offloading the most common data movement operations that generate overhead in data center-scale installations.
• In-Memory Analytics Accelerator (Intel IAA) speeds up database and analytics workloads and may save electricity. This built-in accelerator boosts query throughput and reduces memory footprint for in-memory databases and big data analytics. In-memory, open-source, and data stores like RocksDB and ClickHouse benefit from Intel IAA.

Through oneAPI performance libraries or popular next gen AI frameworks optimized by these libraries, Intel Software Development Tools are essential for maximizing accelerator engine performance. Consider Intel Advanced Matrix Extensions.

Intel AMX activation

Intel AMX accelerates next gen AI workload matrix multiplication with new x86 Instruction Set Architecture (ISA) additions. It has two parts:

• Two-dimensional registers (tiles) that can store submatrices from larger matrices.
• The Tile Matrix Multiply (TMUL) accelerator runs tile instructions.

Support for int8 and bfloat16 data formats boosts AI machine learning speed. Intel oneAPI performance libraries enable Intel AMX and int8/bfloat16 datatypes:

• Intel oneAPI Deep Neural Network Library (oneDNN) is a flexible, scalable deep learning library that performs well on many hardware platforms.
• Intel oneAPI Data Analytics Library (oneDAL) accelerates batch, online, and distributed big data analysis.
• Deep learning and other high-performance computing applications employ Intel oneAPI Collective Communications Library (oneCCL) for collective communication primitives like allreduce and broadcast.
• A popular C++ library for parallel programming, Intel oneAPI Threading Building Blocks (oneTBB) provides a higher-level interface for parallel algorithms and data structures.

Intel oneAPI Base Toolkit and Intel AI Tools enhance machine learning and data science pipelines. oneAPI performance libraries massively optimize TensorFlow and PyTorch, major deep learning AI frameworks.

PC AI Accelerator

Intel Core Ultra processors will power AI PCs with work and content creation apps. Intel’s software-defined and open ecosystem supports ISVs in developing the AI PC category and gives customers, developers, and data scientists flexibility and choice for scaling next gen AI innovation.​

ISVs, developers, and professional content creators can improve performance, power efficiency, and immersive experiences with Intel Core Ultra hybrid processors, Intel Software Development Tools, and optimized frameworks when building innovative gaming, content creation, AI, and media applications. The tools enable advanced CPU, GPU, and NPU functionalities.

• Intel oneAPI compilers and libraries use AVX-VNNI and other architectural features to boost speed. Intel VTune Profiler profiles and tunes applications for microarchitecture exploration, optimal task balancing/GPU offload, and memory access analysis. Get them in Intel oneAPI Base Toolkit. These solutions let developers utilize a single, portable codebase for CPU and GPU, lowering development expenses and code maintenance.
• Intel Graphics Performance Analyzers and Intel VTune Profiler help game creators eliminate bottlenecks for high-performance experiences. Intel Embree and Intel Open Image Denoise in the Intel Rendering Toolkit improve game engine rendering.
• To create content: Create hyper-realistic CPU and GPU renderings for content development and product design utilizing powerful ray tracing frameworks. Enable scalable, real-time GPU rendering with Intel Embree’s ray-traced hardware acceleration and next gen AI-based denoising in milliseconds with Intel Open Image Denoise from the Intel Rendering Toolkit.
• Media: Intel DeepLink Hyper Encode enabled by Intel Video Processing Library speeds up video converting by 1.6x. With a specific API, AV1 encode/decode, and Intel Deep Link Hyper Encode, Intel VPL can use multiple graphics accelerators to encode 60% quicker.
• To implement next gen AI at scale utilizing open source OpenVINO toolkit, use Intel accelerators CPU, GPU, and NPU to optimize inferencing and performance. Start with a TensorFlow or PyTorch-trained model and integrate with OpenVINO compression for easy deployment across hardware platforms. All with little code modifications. Enable Intel Advanced Vector Extensions (AVX-512) on CPU and Intel Xe Matrix Extensions (XMX) on GPU to accelerate deep learning frameworks using Intel oneAPI Base Toolkit’s oneDNN and oneDAL libraries. Optimize TensorFlow and PyTorch training and inference by orders of magnitude using Intel-optimized deep learning AI frameworks. Open source AI reference kits (34 available) accelerate model building and AI innovation across industries.