The Snapdragon Summit 2024 is underway as scheduled, and Lei Tech has arrived in Hawaii, USA, to provide on-site coverage. From the core content of the summit, it is evident that Qualcomm is eager to showcase what the true AI era looks like. Qualcomm devotes significant space to highlighting its technological breakthroughs in AI and demonstrating how its new Snapdragon chips leverage powerful technology to build an AI ecosystem.

Image source: Lei Tech

In the AI era, can Qualcomm maintain its dominant position in the mobile era? Follow Lei Tech as we explore the technologies and products released this time, and perhaps you'll find your own answers.

Snapdragon 8 Elite Edition released, performance comparable to desktop processors

The event kicked off with Qualcomm briefly reviewing Snapdragon's key achievements over the past year, including the recently launched Snapdragon X series commercial laptop products. Thanks to its NPU, which outperforms similar products, the Snapdragon X laptops have created a new category: Copilot+ AP PC. Additionally, Qualcomm's Oryon CPU, used in Snapdragon X, has redefined energy efficiency benchmarks for laptops.

Image source: Lei Tech

What if this powerful Oryon CPU was placed in a mobile chip?

At the Snapdragon Summit 2024, Qualcomm officially unveiled the all-new Snapdragon 8 Elite Edition mobile platform. The platform's "newness" extends beyond its naming convention to its architecture, performance, and energy consumption.

Image source: Lei Tech

Firstly, the Snapdragon 8 Elite Edition adopts the second-generation Qualcomm Oryon CPU (2+6). Notably, the (2+6) configuration here does not refer to the traditional P-cores and E-cores found in mobile processors but rather a similar "all-big-core" architecture as seen in the first-generation Oryon CPU.

Image source: Lei Tech

The second-generation Oryon CPU comprises two core clusters. The first cluster includes two super cores (Prime) with a clock speed of 4.32GHz, each equipped with 192KB of L1 cache. The second cluster has eight performance cores (Performance) clocked at 3.53GHz, each with 128KB of L1 cache. Regarding "efficiency cores," the Snapdragon 8 Elite Edition has completely eliminated this design.

Image source: Lei Tech

To meet the high-speed access demands of AI-era devices, the second-generation Oryon CPU supports LPDDR5X memory at 5.3GHz and boasts 24MB of L2 cache. Interestingly, to avoid potential latency issues with large L2 caches, Qualcomm has allocated the 24MB L2 cache evenly between the two core clusters (12MB+12MB).

Image source: Lei Tech

With this unique architecture, the second-generation Oryon delivers exceptional performance even in mobile devices. For instance, in GeekBench 6 tests, an engineering prototype equipped with the Snapdragon 8 Elite Edition scored 3209 (single-core) and 10205 (multi-core), outperforming an Intel Core Ultra 9 Gen 2 product by 134%, effectively packing desktop-level performance into a mobile chip.

Image source: Lei Tech

According to Qualcomm, the Snapdragon 8 Elite Edition offers a 45% boost in both single-core and multi-core performance compared to the Snapdragon 8 Gen3. Remarkably, even without using efficiency cores, energy consumption has decreased by 40%.

Image source: Lei Tech

For small mobile devices like smartphones, reduced energy consumption brings significant advantages across the board. Firstly, it significantly improves heat dissipation, enabling longer periods of high-performance operation while reducing the need for sophisticated cooling hardware. Additionally, lower energy consumption allows for either a slimmer device profile or extended battery life without sacrificing capacity.

Now, let's move on to the GPU. The Snapdragon 8 Elite Edition offers a 40% performance improvement over its predecessor. The GPU's tile-based rendering architecture, combined with a 1.1GHz clock speed and 12MB of graphics cache, ensures robust graphical performance.

Image source: Lei Tech

Furthermore, the Adreno GPU in the Snapdragon 8 Elite Edition supports various graphics and gaming features, including Snapdragon Game Super Resolution 2.0, FSR, VRR, ray tracing, a 100% framerate boost through intelligent frame interpolation, and even support for Unreal Engine 5's Nanite technology, making it the first mobile platform to do so.

Image source: Lei Tech

In the smartphone market, "gaming performance" has long been a key factor influencing flagship purchases. With the new GPU, the line between mobile and console gaming is blurring. In 2024, many gamers have already achieved "cross-platform play," with some handheld consoles even powered by Snapdragon chips, as highlighted in Qualcomm's keynote. The Snapdragon 8 Elite Edition is poised to further integrate mobile and console gaming, leveraging Qualcomm's signature 5G and Wi-Fi connectivity for ubiquitous gaming experiences.

Image source: Lei Tech

Apart from marking a new era for Snapdragon CPUs, the Snapdragon 8 Elite Edition also sets a new milestone for Snapdragon NPUs. It features an all-new Hexagon NPU, which incorporates tensor, scalar, and vector accelerators, all of which have been enhanced, with the latter two receiving additional cores. These accelerators are optimized for different AI model computations, enhancing the platform's AI model support.

Qualcomm has also refined the concurrent performance mode of the Snapdragon 8 Elite Edition NPU, allowing AI and computer vision (CV) computations to coexist and dynamically allocate resources within the NPU.

Image source: Lei Tech

Image source: Lei Tech

Thanks to the integration of multimodal models, the response speed of the Snapdragon 8 Elite Edition NPU to AI instructions has significantly improved. For example, with LMM (multimodal model) support, the NPU can directly recognize speech content without converting it to text first, significantly enhancing response times. Similarly, LMM enables the NPU to comprehend image content, accelerating text translation and deepening image understanding.

Image source: Lei Tech

Speaking of images, let's delve into the NPU's applications in photography. The AI ISP in the Snapdragon 8 Elite Edition now works in tandem with the NPU at the architectural level, enabling staggering data transfer capabilities—up to 4.3GP/s (gigapixels per second), allowing simultaneous support for three 48MP cameras.

Image source: Lei Tech

Furthermore, this design allows the NPU to directly access the AI ISP via memory-direct technology, enabling faster segmentation and element recognition for faster bokeh or object removal effects.

Image source: Lei Tech

This design opens up new possibilities for the Snapdragon 8 Elite Edition's imaging capabilities, such as adding real-time, adjustable simulated lighting effects to videos or applying AIGC object removal.

The focus of the AI era is not 'super apps' but 'connected apps'

At the Snapdragon Summit 2024, Lei Tech encountered an intriguing viewpoint:

In the AI era, the emphasis is not on standalone 'AI apps' but rather on leveraging universal AI capabilities to bridge different apps, integrating AI into various applications, expanding its capabilities through software, and reshaping app experiences.

Image source: Lei Tech

Put simply, future mobile AI use cases can be divided into four parts: AI developers, apps, mobile OS, and NPUs. Users make AI requests through apps, which access the NPU via frameworks like ONNX and DirectML, leveraging Qualcomm's AI software stack. The NPU processes these requests, and the OS schedules and presents the results through the app.

In Honor's demonstration, Dr. Guo Rui, Honor's CMO, showcased how Honor leverages Qualcomm's AI Agent for novel AI interactions:

After preloading training results, users can directly request AI assistance for tasks like ordering coffee or canceling automatic payments via voice commands. Upon understanding the instructions, the AI assistant simulates user interactions based on current UI elements to complete tasks, such as opening food delivery apps for coffee orders or payment apps to manage subscriptions.

Image source: Lei Tech

While AI-driven app interactions are still relatively rudimentary, Qualcomm envisions vast potential, including bypassing app interfaces to directly access app logic for "headless" computations, comprehensively enhancing interaction efficiency.