Intel's AI Revolution: Gaudi3, Core Ultra, 288-core Xeon

On September 19th (September 20th, Beijing time), Intel convened the "Intel Innovation" in San Jose, California, USA. The central theme of this summit was "Intel Client Hardware Roadmap and the Emergence of Artificial Intelligence." The event marked the debut of Intel's latest Core Ultral processors, designed to power the next generation of AI-driven PCs. Furthermore, Intel unveiled notable progress in advanced manufacturing processes, packaging technologies, and pioneering research in areas such as neuromorphic computing and quantum computing.

During the summit, a significant announcement was made regarding Stability AI, an open-source AI company that had successfully developed a large-scale AI supercomputer. This supercomputer was entirely constructed using Intel Xeon processors and 4000 Intel Gaudi2 AI accelerators.

The Gaudi 2 processor, introduced in May 2022, represents Intel's cutting-edge high-performance deep learning AI training processor. It is crafted using advanced 7nm process technology, featuring 24 programmable Tensor processor cores (TPCs), 96GB of HBM2e memory, and 24 100GbE ports. Performance comparisons showcased that the Habana Gaudi2 outperforms the NVIDIA A100 by approximately 2 times in various tasks, including ResNet50 Training Throughput and BERT Training Throughput, which pertain to video and natural language processing models.

Within this event, Intel also offered a glimpse into the forthcoming Gaudi 3, the next-generation AI chip. Gaudi 3 will leverage a more advanced 5nm process technology, promising substantial enhancements over Gaudi 2. Anticipated improvements include 4 times the BF16 performance, double the computational prowess, 1.5 times the network bandwidth, and 1.5 times the HBM capacity, resulting in significantly augmented AI performance.

Furthermore, Intel highlighted the capabilities of the fourth-generation Intel Xeon Scalable processors, equipped with built-in AI accelerators, which will contribute to generative AI training.

Despite the current slowdown in the PC market, Intel envisions a transformative convergence of AI and PCs. In pursuit of the "AI PC" era, Intel announced the official release date of the Intel Core Ultral processor, codenamed Meteor Lake, set to launch on December 14th this year.

The Intel Core Ultral processor will feature the first integrated Neural Network Processor (NPU) designed to deliver high-efficiency AI acceleration and local inference on PCs. Notably, it is the pioneering client chip design employing 3D Foveros packaging technology, alongside Intel's 4th-generation process technology, and integrates Intel Xe graphics, delivering performance comparable to standalone graphics cards.

The summit also included a roadmap presentation for Intel's Xeon processors. The fifth-generation Xeon processors, to be released on December 14th, promise to enhance performance and storage speed for global data centers while maintaining the same power efficiency.

Following this release, the Sierra Forest, a high-efficiency processor with 288 cores, is expected to significantly increase rack density by 2.5 times and enhance performance per watt by 2.4 times compared to the fourth-generation Xeon. Subsequently, the high-performance core (P-core) processor named Granite Rapids will further boost AI performance by 2 to 3 times compared to its predecessor.

Looking forward to 2025, the next-generation high-efficiency core processor, codenamed Clearwater Forest, will be manufactured using Intel's advanced 18A process node.

In the domain of advanced packaging technologies, Intel's commitment extends beyond EMIB and Foveros 3D packaging. Intel is actively developing novel packaging materials. Recently, Intel disclosed significant advancements in its semiconductor glass substrate technology. Compared to prevailing organic substrates, glass offers superior attributes in flatness, thermal stability, and mechanical stability. This empowers chip architects to accommodate more chiplets within a single package, leading to improved performance, density, flexibility, reduced costs, and power efficiency. Intel's glass substrate technology, by increasing chip area within a single package by 50%, holds the potential to extend Moore's Law beyond 2030.

Intel emphasized that glass substrates can complement traditional organic substrates to enhance structural integrity. Glass substrates are particularly well-suited for Intel's advanced packaging technologies like EMIB and Foveros.

To empower developers in harnessing the latest Intel software and hardware innovations for AI development, Intel unveiled the comprehensive "Intel Developer Cloud Platform." Developers can leverage this platform to access Intel's latest hardware offerings, including the Intel Gaudi2 accelerator for deep learning, the fifth-generation Intel Xeon Scalable processors, and Intel's Data Center GPU Max series 1100 and 1550. The Intel Developer Cloud Platform facilitates the construction, testing, and optimization of AI and HPC applications. Developers can run workloads ranging from small-scale to large-scale AI training, model optimization, and inference, all while achieving high performance and efficiency. This platform is founded on the open oneAPI programming model, which supports multiple architectures and hardware vendors, offering hardware choices and eliminating proprietary programming models. It promotes accelerated computing, code reusability, and meets portability requirements.

During the Intel Innovation, AI was a recurring and central theme, underscoring Intel's unwavering commitment to the field of artificial intelligence.