Every technology company has hopped on the AI bandwagon over the past two years. But the major players are now trying to position themselves as an end-to-end or top-to-bottom, depending on how you look at it, AI solutions provider. The same holds true for Intel, which unleashed a flurry of announcements around its “AI Everywhere” strategy at the company’s annual Vision event held in Phoenix, Arizona.

The Intel view

With the event focused on enterprise use cases, Intel promoted end-to-end solutions focused on three key areas: PCs, edge (IoT) solutions, and data center solutions. A key tenet to the Intel solutions is openness. Intel is promoting an open ecosystem to allow enterprises to build tailored, easy-to-use and secure AI solutions.

In Intel’s view, this means open applications, software, and infrastructure ecosystems to run on the selected hardware, including  Intel Core and Xeon processors, Arc GPUs, Gaudi AI accelerators, and infrastructure processing units (IPUs).

Intel’s strategy starts with AI PCs using the Intel Core Ultra PC processors, codenamed Meteor Lake, that were introduced in 2023. Intel expects more than 230 commercial AI PC designs using Core Ultra and is working with more than 100 independent software vendors (ISVs) supporting on-device AI applications and functionality. Intel CEO Pat Gelsinger held up a next-generation Core Ultra processor called Lunar Lake at Vision, hinting that the launch of the product is approaching.

For AI edge applications, such as industrial IoT, Intel discussed the new Core Ultra, Core, and Atom processors, Arc GPUs, and Altera FPGAs announced earlier in the week at embedded world 2024. Based on the PC processor architectures, the new Core Ultra and Core embedded processors offer more processing, graphics and AI processing capabilities for edge applications—especially for industrial applications like AI vision.

The Atom x7000C series processors are targeting networking and telecommunications applications with up to eight efficient CPU cores. The Atom x7000RE series processors are aimed at industrial classification systems. The Arc discrete GPUs offer additional AI performance, plus additional media and graphics processing capabilities for edge applications.

Additionally, the new Altera—an Intel company—Agilex 5 product family is now available. According to Altera’s CEO Sandra Rivera, this is “the first FPGA with AI-infused throughout the fabric.” These platforms will be supported by the Intel Edge Platform announced at Mobile World Congress 2024. The Intel Edge Platform is a developer environment that provides an easy way to develop, deploy, and run edge and AI applications.

For the data center, Intel provided a preview of the upcoming Xeon processors and introduced a new AI accelerator and networking solution.

First in line is the Xeon processors. Intel previewed the sixth-generation Xeon processors, to be called Xeon 6, that will offer more SKUs with new efficient CPU cores (E-cores) and performance CPU cores (P-cores) to better match the processor to the workload requirements. According to Intel, the Xeon 6 with E-cores, codenamed Sierra Forest, will offer a 2.4× increase in performance/watt and 2.7× increase in performance per rack over the previous generation.

Xeon 6 with P-Cores, codenamed Granite Rapids, incorporates support for the MXFP data format for a 6.5× latency reduction compared to using FP16 on the fourth-gen Xeon processors. It will also offer the ability to run larger large language models (LLMs) like the 70 billion parameter Llama-2.  Both products will be available later in 2024.

Next up, and the star of the event, is the new Gaudi 3 AI accelerator for training and inference. According to Intel, Gaudi 3 boasts 2× the AI performance using the 8-bit floating point data format (FP8), 4× the AI performance using the increasingly common bfloat-16 data format, 2× the networking bandwidth, and 1.5× the memory bandwidth when compared to the previous generation Gaudi 2.

Technical specs and details of Gaudi 3 are provided in a sperate article by my colleague and TIRIAS Research principal analyst Francis Sideco. The Gaudi solutions will be available later this quarter from OEMs in mezzanine card, universal board, and PCIe add-in card form factors and from the leading server OEMs Dell, HPE, Lenovo, and Supermicro. Intel claims price-performance value, but no pricing was available.

The final data center announcement is a new AI network interface card (NIC). According to Intel, the AI NIC supports the new open ethernet-based networking connectivity standards for AI and HPC workloads that are being developed by the Ultra Ethernet Consortium. The new standards will support the unique demands of AI workloads and the scale-out of platforms for large-scale AI deployments.

While not directly related to new product, technology or service announcements, Intel indicated that it will be delivering IPUs for infrastructure and management acceleration later this year, that all solutions will be supported by the company’s confidential computing solutions and services, and that the Intel Developer Cloud is available for access to the latest Intel hardware solutions and rapid development of AI solutions.

Intel promises to scale the Intel Developer Cloud with additional features and functions, as well as providing early access to upcoming products. Additionally, as part of Intel’s support for OneAPI and the Unified Acceleration Foundation (UXL), PyTorch 2.0 will include support for AI acceleration on Intel CPUs for inference and training, and is working on support for Intel accelerators—which they will be announcing later this year. Further supporting its call for an open ecosystem, Intel claims that acceptance of OpenVINO is accelerating with over one million downloads in the first quarter of 2024 alone.

In a separate announcement, Intel also introduced the Tiber portfolio of business solutions. According to Intel, Tiber combines some of the common software solutions the company has developed for AI, edge, cloud, and trust and security.

The AI future

Intel believes that we are currently in the age of AI co-pilots but are moving to the age of AI agents that can take over complete tasks. Intel also believes that in the future, we will move to AI functions that will combine AI agents to take over complete enterprise functions like finance and human resources.

Like many tech companies, Intel is seeing a shift from using a huge single model to using smaller and sometimes specialized models together, often called ensemble modeling or mixture of experts. Additionally, Intel sees the need to use proprietary enterprise data to generate more relevant and accurate models. Intel is working to address these needs though its open ecosystem activities and the company’s products, technologies and services.

 ----Form EE Times