Intel has unveiled a groundbreaking array of technologies and architectures at Computex 2024 today, which are set to have a major impact on the AI ecosystem across various domains, from data centres and the cloud to networks, the edge, and personal computers. These advancements promise to deliver enhanced processing power, leading-edge power efficiency, and lower total cost of ownership, providing customers with the tools to fully capitalize on the AI system’s potential.
Pat Gelsinger, CEO of Intel, remarked, “AI is propelling us into one of the most transformative periods of innovation in the industry’s history. The power of silicon is once again driving exponential leaps in computing capability, which will expand the boundaries of human potential and fuel global economic growth for years to come.”
Gelsinger also highlighted the company’s comprehensive approach to AI innovation, spanning semiconductor manufacturing, PC, network, edge, and data centre systems. At Computex, Gelsinger emphasised Intel’s latest platforms like Xeon, Gaudi, and Core Ultra, along with a robust hardware and software ecosystem, offering flexible, secure, sustainable, and cost-effective solutions for evolving customer needs.
At the keynote, Gelsinger also took some potshots. “Moore’s Law still applies, no matter what my friend Jensen (Jensen Huang, CEO, NVIDIA) says… As long as there are elements, we will have Moore’s Law.”
To put things in context, Huang had famously said, Moore’s Law is dead, because of which extracting better performance at a cheaper price is no longer possible.
Intel’s commitment to democratising AI was evident as Gelsinger stressed open standards and Intel’s ecosystem’s importance in accelerating AI opportunities. Industry leaders like Acer, ASUS, Microsoft, and Inventec echoed these sentiments, emphasising collaboration for AI advancement.
In edge and personal computing, Intel’s AI footprint expansion includes over 90,000 edge deployments and delivering 200 million CPUs to the ecosystem. Intel’s focus on AI PCs is transforming computing experiences, creating devices that learn and adapt in real-time, and enhancing productivity, efficiency, and creativity.
With AI PCs forecasted to dominate 60 per cent of new PCs by 2027, Intel has positioned itself as a frontrunner in this market. Supporting over 100 ISVs and 300 features, along with 500 AI models across its Core Ultra platform, Intel’s advancements in AI integration are poised to shape the future of computing.
Lunar Lake CPUs revealed
Intel has revealed the architectural details of Lunar Lake, its flagship processor set to power the next generation of AI PCs. Scheduled for release in the third quarter of 2024, Lunar Lake promises a substantial advancement in graphics and AI processing capabilities. Compared to its predecessors, it boasts up to 40 per cent more system-on-chip (SoC) power and over triple the AI compute performance.
Lunar Lake’s architecture introduces several key features to enhance performance and energy efficiency. It incorporates new Lion Cove Performance-cores (P-cores) and Skymont Efficient-cores (E-cores) for improved overall performance and energy efficiency. Additionally, the processor integrates a fourth-generation Intel neural processing unit (NPU), capable of delivering up to an impressive 48 TOPS of AI performance. The graphics processing unit (GPU) receives a significant upgrade with the introduction of the all-new Battlemage design, aimed at delivering enhanced graphical performance.
Moreover, Lunar Lake includes an advanced low-power island specifically engineered to optimize battery life, ensuring prolonged usage on portable devices. With these groundbreaking innovations, Intel aims to redefine the AI landscape, providing customers across various industries with the tools to unlock new levels of performance, efficiency, and innovation.
The new architecture
Yaron Alankry, Vice President of Client Computing Group and General Manager of Silicon Engineering, along with Arik Gihon, Senior Principal Engineer of SoC Architecture, delved into the intricacies of the Lunar Lake processors’ architecture, promising significant improvements in performance and power efficiency upon its anticipated launch later this year.
Intel has meticulously crafted Lunar Lake to deliver a host of cutting-edge features:
Firstly, it promises a breakthrough in x86 power efficiency, ensuring optimal performance while conserving energy.
Moreover, Lunar Lake boasts a substantial leap in graphics capabilities, enhancing the mobile gaming experience and delivering stunning visuals.
The processors also offer unparalleled AI compute capabilities, setting the stage for an exceptional user experience that anticipates and responds to user needs effectively.
Additionally, Lunar Lake excels in core performance, ensuring seamless operation across various tasks and applications.
Furthermore, the architecture of Lunar Lake is designed with future scalability in mind, laying the groundwork for advancements in subsequent generations.
A closer look at Lunar Lake’s microarchitecture reveals two distinct tiles interconnected through Intel’s renowned Foveros packaging technology, supplemented by memory on package:
The Compute Tile houses the latest-generation E-cores and P-cores, incorporating novel microarchitecture enhancements to achieve unprecedented x86 efficiency. Additionally, it accommodates the new Xe2 GPU, NPU 4, and an image processing unit (IPU).
The Xe2 GPU delivers over 80 per cent more gaming performance and more than five times the AI throughput compared to its predecessors, boasting over 60 TOPS. Moreover, advancements in display and media engine microarchitecture ensure an unparalleled visual experience.
The NPU 4, with up to 48 TOPs, offers three times more computational power than the previous generation, facilitating AI computations with remarkable efficiency. Meanwhile, the IPU enhances the camera experience while minimizing power consumption.
On the other hand, the Platform Controller Tile (PCD) integrates security and connectivity features. Its security components provide robust built-in security engines to safeguard system integrity, while the upgraded connectivity suite incorporates Wi-Fi 7.0, Bluetooth 5.4, PCIe Gen5, PCIe Gen4 ports, and Thunderbolt™ 4 ports.
The all-new NPU 4
Intel’s latest Neural Processing Unit (NPU 4) marks a significant leap in AI performance, offering a range of impressive capabilities:
NPU 4 delivers up to 48 TOPS, a four-fold increase compared to its predecessor, Meteor Lake. It also boasts twice the performance at ISO power compared to NPU 3, ensuring enhanced efficiency without compromising on power consumption.
NPU 4 is finely tuned for modern AI applications, including Large Language Models (LLMs), transformers, Vision models, and more. Its optimized software stack further enhances its adaptability to diverse AI tasks.
Built upon a proven foundation and drawing from three prior generations, NPU 4 offers higher compute capacity to accommodate a growing array of use cases while ensuring increased efficiency to support prolonged battery life, enabling all-day usage.
Compared to its predecessor, NPU 4 features an increased number of engines, improved clock speeds, and multiple architectural enhancements aimed at boosting both performance and efficiency.
One notable improvement lies in the significant enhancements to the frequency and voltage curve achieved through pipeline optimization, targeted voltage/frequency optimization, and the utilization of Machine Learning (ML) and Artificial Intelligence (AI) techniques, resulting in up to 20 per cent power reduction beyond process and design optimizations.
Additionally, NPU 4 doubles the Direct Memory Access (DMA) bandwidth, leading to improved network performance, particularly beneficial for LLM..
Moreover, it introduces specialized matrix and vector engines, an inference pipeline with MAC (Multiply-Accumulate) arrays, fixed function blocks, and programmable Digital Signal Processors (DSPs) to further augment its computational prowess.
Furthermore, NPU 4 extends its support with additional functions, including automatic data type conversion and fused operations tailored for quantized networks, enhancing its versatility across various AI applications.
The new Xe2 GPU
Intel’s commitment to advancing gaming and media experiences was outlined by Tom Petersen, who shed light on the company’s efforts with the Intel Arc GPU integrated into Lunar Lake. This includes the incorporation of the all-new Xe2 GPU architecture and updated media and display features, bolstered by support for the latest industry standards.
The freshly minted GPU is engineered to deliver a remarkable 1.5x performance boost over its predecessor, while also introducing novel AI capabilities to the Lunar Lake platform.
Key highlights of the Xe Media Engine include an efficiency-optimized design and low-power hardware decode tailored for the latest VVC codec technology, ensuring an enhanced media experience with reduced energy consumption.
Meanwhile, the Xe Display Engine boasts an efficiency-optimized design, facilitating support for up to three 4K60 HDR displays with HDMI 2.1 and DisplayPort 2.1, alongside compatibility with the latest eDP 1.5 standard for laptop panels.
The Xe2 GPU, optimised for improved efficiency and reduced software overhead, features 8x 2nd Gen Xe-cores, XMX AI acceleration capabilities with up to 67 TOPS INT8, and 8x enhanced Ray Tracing Units, promising heightened gaming and graphics performance.
Furthermore, platform and software optimisation efforts leverage Arc’s software stack, including XeSS AI-based upscaling with new optimised kernels, ensuring an optimised and immersive gaming experience for users leveraging the Lunar Lake platform.
