Workflow
AI模型压缩
icon
Search documents
速递|量子学家重构AI压缩算法,Multiverse已筹集2.15亿美元,打造出史上体积最小两款模型
Z Potentials· 2025-08-15 03:53
Core Viewpoint - Multiverse Computing has developed two of the smallest high-performance AI models, named after the sizes of animal brains, aimed at enhancing AI capabilities in IoT devices and enabling local operation on smartphones and personal computers [2][3]. Company Overview - Multiverse Computing is a European AI startup based in Donostia, Spain, founded by experts in quantum computing and AI, including Roman Orús and Samuel Muguel [4]. - The company has raised approximately €189 million (around $215 million) in funding, with a total of about $250 million since its inception in 2019 [4]. Technology and Innovation - The company utilizes a quantum-inspired compression algorithm called CompactifAI, which allows for significant model size reduction without sacrificing performance [4][5]. - Multiverse has released compressed versions of popular open-source models, including Llama 4 Scout and Mistral Small 3.1, and has also compressed large models like DeepSeek R1 Slim [4]. New Model Launch - The two new models, SuperFly and ChickBrain, are designed for IoT applications, with SuperFly being a compressed version of the SmolLM2-135 model, reduced from 135 million parameters to 94 million [6]. - ChickBrain, with 3.2 billion parameters, is a compressed version of Meta's Llama 3.1 8B model, capable of running offline on devices like MacBooks [6][7]. Performance Metrics - ChickBrain has outperformed its original model in several benchmark tests, including language and mathematical ability tests [7]. - Multiverse has not claimed that its models can surpass the performance of the most advanced large models, focusing instead on maintaining performance while reducing size [10]. Market Engagement - The company is in discussions with major device and appliance manufacturers, including Apple, Samsung, Sony, and HP, which has also invested in the company [10]. - Multiverse offers its compression technology for other forms of machine learning, such as image recognition, and has secured clients like BASF and Bosch [11].
计算机行业周报:AMD发布MI350系列GPU性能升级,中国科学院发布「启蒙」芯片设计系统-20250619
Huaxin Securities· 2025-06-19 06:35
Investment Rating - The report maintains a "Buy" rating for several companies in the computer and AI sectors, including 亿道信息 (Yidao Information), 科大讯飞 (iFlytek), 唯科科技 (Weike Technology), 泓淋电力 (Honglin Electric), 嘉和美康 (Jiahe Meikang), 寒武纪 (Cambricon), 鼎通科技 (Dingtong Technology), and 迈信林 (Maixinlin) [13][54]. Core Insights - AMD has launched the MI350 series GPUs, which offer a fourfold increase in computing power and a 35-fold increase in inference speed compared to the previous MI300 series. The MI350 series is designed to compete with NVIDIA's B200 GPUs, featuring 288GB HBM3E memory and 8TB/s bandwidth [4][20][24]. - The MI400 series, expected to be released in 2026, will be developed in collaboration with OpenAI and is projected to be ten times faster than the MI300 series, with significant enhancements in memory and processing capabilities [5][25]. - The "启蒙" (Enlightenment) chip design system developed by the Chinese Academy of Sciences aims to automate the entire chip design process, achieving or surpassing human expert levels in efficiency and performance [31][33][34]. - Multiverse Computing has completed a $217 million Series B funding round, focusing on AI model compression technology that can reduce model sizes by up to 95% without sacrificing performance [40][41]. Summary by Sections Computing Power Dynamics - AMD's MI350X and MI355X GPUs have been released, showcasing a significant performance upgrade over the MI300 series, with a fourfold increase in computing power and a 35-fold increase in inference speed [4][20]. - The MI350 series has a memory capacity 1.6 times that of NVIDIA's B200 and offers superior performance per dollar spent on token processing [4][24]. AI Application Dynamics - The average weekly traffic for Gemini has increased by 11.26%, indicating growing interest in AI applications [30]. - The "启蒙" system is designed to automate chip design processes, significantly improving efficiency and customization capabilities [31][33]. AI Financing Trends - Multiverse Computing's recent funding round highlights the increasing demand for AI model compression technologies, which can enhance performance while reducing operational costs [40][41]. Investment Recommendations - The report suggests a positive outlook for overseas computing power chains, particularly in light of Oracle's projected growth in cloud infrastructure revenue [52]. - Companies such as 嘉和美康 (Jiahe Meikang), 科大讯飞 (iFlytek), and 寒武纪 (Cambricon) are highlighted as key players to watch in the AI and chip technology sectors [53].