随着生成式AI技术的快速发展，终端侧AI的创新正在逐渐成为行业关注的焦点。高通技术公司作为行业领导者，正在通过其端到端系统理念，推动边缘侧生成式AI技术的发展。
生成式AI，尤其是Transformer架构，正在从传统的文本和语言处理扩展到更多模态，如汽车行业中的多摄像头和激光雷达协同，以及无线通信中的全球定位系统、摄像头和毫米波信号优化。这种技术进步正在为用户提供增强的体验、提高生产力和带来全新的娱乐形式。
在模态和用例方面，生成式AI的能力正在不断提升，包括语音UI、多模态大模型、智能体和视频/3D技术的提升。在能力和KPI方面，更长上下文窗口、个性化和高分辨率能力的提升正在成为可能。为了实现生成式AI的全部潜能，将这些趋势引入边缘侧终端对于实现时延改善、交互泛化和隐私增强至关重要。
高通技术公司正在通过知识蒸馏、量化、高效图像和视频架构，以及异构计算等技术优化生成式AI模型，使其能够在硬件上高效运行。例如，通过量化感知训练结合知识蒸馏，可以实现准确的4位大语言模型，而向量量化则可以帮助进一步压缩模型大小。
此外，高通还在开发高效的视频架构技术，如FAIRY视频生成方法，以提高视频到视频生成式AI技术的效率。这些技术进步不仅能够提升生成式AI的能力，还能够为具身机器人的实时交互提供支持，从而开辟通向具身机器人之路。
未来，随着高通技术公司在边缘侧生成式AI技术的持续创新，我们将看到更多的技术进步，这些进步将推动终端侧AI的发展，为用户带来更加智能和便捷的体验。
英语如下：
Title: “Qualcomm Leads the Revolution in Edge-Side Generative AI”
Keywords: Generative AI, Edge Computing, Qualcomm Innovation
Content: As generative AI technologies continue to advance rapidly, innovation at the edge side of AI is increasingly becoming a focal point of industry attention. As a leading industry player, Qualcomm Technologies is driving the development of edge-side generative AI technologies through its end-to-end system philosophy.
Generative AI, particularly the Transformer architecture, is expanding beyond traditional text and language processing to encompass more modalities, such as collaborative efforts between multiple cameras and lidar in the automotive industry, and optimizing global positioning systems, cameras, and millimeter-wave signals in wireless communications. This technological progress is providing users with enhanced experiences, boosting productivity, and introducing new forms of entertainment.
In terms of modalities and use cases, the capabilities of generative AI are continuously being enhanced, including improvements in voice user interfaces, multimodal large models, intelligent agents, and video/3D technologies. In terms of capabilities and Key Performance Indicators (KPIs), longer context windows, personalized experiences, and higher resolution capabilities are becoming feasible. To realize the full potential of generative AI, introducing these trends to edge-side terminals is crucial for improving latency, generalizing interactions, and enhancing privacy.
Qualcomm Technologies is optimizing generative AI models for efficient hardware operation through techniques such as knowledge distillation, quantization, efficient image and video architectures, and heterogeneous computing. For example, by combining quantization-aware training with knowledge distillation, accurate 4-bit large language models can be achieved, and vector quantization can help further compress model sizes.
Additionally, Qualcomm is developing efficient video architecture technologies, such as the FAIRY video generation method, to enhance the efficiency of video-to-video generative AI technologies. These technological advancements not only enhance the capabilities of generative AI but also support real-time interactions for embodied robots, opening the path to embodied robotics.
Looking ahead, as Qualcomm Technologies continues to innovate in edge-side generative AI technologies, we can expect to see more technological advancements that will drive the development of edge-side AI, bringing users smarter and more convenient experiences.
【来源】https://www.jiqizhixin.com/articles/2024-08-08-10
Views: 2