Shared todayAI seriesIn-depth Research Report:AI Topic: AI Hardware: The curtain of AI on the device side has opened, leading a new round of terminal innovation cycle
Report Producer: Southwest ** Research and Development Center).
Report total: 51 pages.
Featured Report**: The School of Artificial Intelligence
PC has gone through the ** period of the 90s of the last century. At that time, the Internet was open, computers began to penetrate rapidly into the mass market after miniaturization and low cost, and the Wintel alliance formed by Microsoft and Intel almost monopolized the operating system and CPU market, and the spiral upgrade of software and hardware led to the prosperity and development of the PC industry.
After the bursting of the Internet bubble in 2000, before the emergence of smart phones in 2007, the terminal market was dominated by feature phones, **, etc., but it had not yet formed a popular terminal form.
In 2007, Apple released the iPhone, which unified "phone + iPod + Internet" and opened the era of smart phones. The commercialization of 3G and 4G networks has promoted the great development of the mobile Internet, with various applications emerging in an endless stream, and the iOS and Android ecosystems are unprecedentedly prosperous. Smartphones are leading the innovation of terminals, driven by the needs of various application scenarios.
After 2018, as the penetration rate of mobile phones and PCs approached the ceiling, the overall innovation of terminals also showed a lack of innovation. The industry has also experienced a downward adjustment in the past two years.
The rapid development of generative AI, represented by GPT, has accelerated the transformation of work, study and life patterns and the intelligent transformation of various industries. As the ultimate carrier for AI to reach users, AI terminals will bring long-lost software, hardware, and ecological innovation to the terminal industry.
In 2023H2, major terminal manufacturers have successively launched or released corresponding AI products and future plans. 2024 will be the first year of the centralized release of AI PC, and the leading PC companies have a sales planTwo AI phones have been launched in 2023Q4, and more AI Android models will be launched next yearAI wearable devices represented by AI PIN have opened a new entrance to device-side AIIn addition, the combination of AI models and XR devices may bring about a new ecosystem, which can focus on the progress of Apple and Meta in this regard.
In terms of AI chips, AMD, Intel, Qualcomm, MediaTek, etc. have successively released terminal AI chip products. In terms of AI PC chips, given the dominance of the x86 architecture in the PC field and a strong ecosystem, Intel and AMD are still the main partners of PC manufacturersHowever, based on its advantages of low power consumption, small chip size and low cost, the ARM architecture is expected to gain more market share in the AI PC era with the continuous improvement of software compatibility, and the progress of mobile chip manufacturers such as Qualcomm in the PC field can be focused. In terms of AI mobile phone chips, we believe that Apple, Qualcomm, MediaTek, Samsung, and leading domestic communication manufacturers will continue to become major players in this market by relying on years of technology accumulation and patent barriers.
The penetration rate of smartphones is close to the ceiling, and the overall product lacks sufficient innovation, and mobile phone shipments have begun to show cyclical characteristics. Affected by the epidemic in 2020-2021, the demand for ** entertainment has increased, driving the demand for mobile phones. Since then, the smartphone industry has experienced a downward adjustment in the past two years.
The rapid development of generative AI and LLM has profoundly changed the way individuals live and work. As the ultimate carrier for artificial intelligence to reach end users, smart devices are becoming an important breakthrough in the future development and implementation of AI. AI mobile phones combine AI model applications with mobile phones, bringing innovation and change to the mobile phone industry, or bringing a new round of innovation cycle in the smartphone industry.
At present, the AI mobile phones vivo X100 and Google Pixel 8 are equipped with billions of parameters of lightweight AI models, and related AI applications are mainly focused on AI assistants, text generation, voice and image processing, etc.
Google unveiled the Pixel 8 phone on October 4, with the Pixel 8 starting at $699 and the Pixel 8 Pro starting at $999. The Pixel 8 Pro is the first device to be equipped with Google's AI-based model, which allows users to run and experience Google AI models locally, which is faster and more effective with the blessing of Google's self-developed Tensor G3 processor, and has further breakthroughs in natural language understanding. For example, the Magic Editor, which runs offline on the Pixel 8 Pro, allows users to freely select and move objects in the Pixel 8 Pro and reposition them, as well as replace the backgroundThe Audio Magic Editor removes noise from **;Gboard can automatically generate more natural and communication habits based on the conversation informationThe Pixel 8 Pro will have a built-in model dedicated to image processing, generating sharper details for zoomed-in images in the gallery* Enhancement function Video Boost can effectively adjust the color, light, noise, etc., and improve the quality;Implement functions such as generating web summaries, reading aloud and translating web pages.
In the coming months, Google will launch Assistant with Bard, powered by generative AI that combines the generative and inference capabilities of Google Bard. The tool will integrate with Google apps like Gmail and Docs, making it easy for users to talk to Google Assitent and let it help with a range of actions such as creating social posts, creating shopping lists, and finding information in emails. In addition, Google plans to add more intuitive generative AI features to Magic Editor in the future to boost the processing power of images** and more.
On December 6, Google released the Gemini multi-modal large model, which is divided into three versions: Ultra, Pro, and Nano. Of which the nano version has 1.8 billion 32500 million parameters, applicable to device-side devices. Gemini performed slightly better than GPT-4 in preliminary tests, with Gemini Ultra scoring 30 SOTA out of 32 benchmarks and being the first to reach the level of a human expert on the MMLU benchmark. The Gemini will be available on the Pixel 8 Pro.
Google has made a lot of attempts at end-to-end AI on Pixel. The Android team has demonstrated an application that integrates with the native keyboard, which not only learns the individual's typing habits, but also autocompletes the user experience based on the previous and lower levels. In addition, accessibility teams that primarily serve people with disabilities can now change the way visually impaired people interact with their phones through a multimodal large model, where users directly ask about the content of their phones and instruct them to tap on a specific area of the screen.
Apple has always considered AI and machine learning as foundational technologies and incorporated them into most of its products. Apple's R&D spending has steadily increased, with nearly $30 billion in R&D investment in fiscal 2023Capital expenditures exceed $10 billion annuallyFree cash flow reached 995 in fiscal year 2023$800 million. A good cash flow situation and huge investment in new technologies are the guarantee of Apple's in-depth deployment of AI.
Apple is working on generative AI. Some of the features in the previously released iOS 17, such as personal voice and real-time voicemail, are inseparable from the support of AI;Life-saving features on watches and phones, such as fall detection, crash detection, and electrocardiograms, also rely on AI technology. It is expected that Apple will add AI large model capabilities in iOS 18 and introduce new features based on AI technology. For example, improve Siri and Message app handling. A smarter Siri assistant is expected to arrive next year.
AI mobile phones not only have powerful computing power, but also bring innovative ways of interaction and visual experience. Running AI models locally on mobile phones will increase the demand for computing power and storage.
Qualcomm's next-generation Snapdragon platform, 8Gen3, injects high-performance AI into the entire system, enabling users to create unique content, boost productivity, and enable other breakthrough use cases to further scale on-device AI. The Snapdragon 8Gen3 is manufactured on a 4nm process. Compared with the previous generation Snapdragon 8 platform, the Snapdragon 8 Gen3 has a 98% improvement in NPU performance, a 30% increase in CPU performance, a 25% increase in GPU performance, and a 10% reduction in overall power consumption. Snapdragon 8Gen3 is the world's first chip platform to support devices to run 10 billion parameter models, generating 20 tokens per second for 7 billion parameter LLMs, and supporting the generation of images on mobile phones using Stable Diffusion in less than a second. A variety of AI applications will run stably on Snapdragon 8Gen3-powered phones, including GPT chatbots, local real-time translation, ultra-fast bilingual generation, various virtual assistants, and more.
MediaTek Dimensity 9300 is the first in the industry to adopt an all-large-core architecture design, with an octa-core CPU equipped with 4 Cortex-X4 ultra-large cores and 4 Cortex-A720 large cores, which is more than 15% higher than the previous generation of single-core performance and 40% higher multi-core performance. The Dimensity 9300 is powered by the 7th generation AI processor, the APU790, which improves AI performance by 200% and reduces power consumption by 45%. The GPU has 46% higher peak performance and 40% lower power consumption than its predecessor, and it supports the second-generation ray tracing engine, making it the first to achieve console-level global illumination. The Dimensity 9300 can support AI large models with up to 33 billion parameters. The Dimensity 9300 benchmark is class-leading in its class.
MediaTek provides a complete toolchain to help developers quickly and efficiently deploy multimodal AI applications on the device side, and provide an innovative AI experience for the device side, including text, images, and more. MediaTek has cooperated with vivo to achieve the implementation of AI large language models with 1 billion and 7 billion parameters, as well as AI visual models with 1 billion parameters.
Report total: 51 pages.
Featured Report**: The School of Artificial Intelligence