当AI时代算力集群规模正逐步从万卡向十万卡、百万卡甚至千万卡升级时，一支中国团队悄然另辟蹊径。 今年10月，北京大学人工智能研究院/集成电路学院双聘助理教授孙仲与北京大学集成电路学院蔡一茂教授、王宗巍助理教授率领的团队成功研制出基于 阻变存储器的高精度、可扩展模拟矩阵计算芯片，在全球范围内首次将模拟计算的精度提升至24位定点精度，让未来同等任务下使用更少的计算卡成为可 能。 这是一种完全不同于目前所有商用量产芯片的新型芯片，计算精度从1%跃升至千万分之一；可以支撑6G、具身智能及AI大模型训练等多个前沿场景；更 重要的是，它可在28纳米及以上成熟工艺量产，绕开光刻机"卡脖子"环节。 事实上，AI大模型、具身智能、6G等应用背后都是矩阵计算，AI推理是做矩阵乘法，AI训练是在解矩阵方程。而英伟达的崛起正是得益于GPU（图形处 理器）很擅长做矩阵计算。 但如今摩尔定律趋于终结，晶体管很难再微缩，所以业界现在只能横向堆计算卡：少则百卡，多则万卡、十万卡。但这样的方式我认为是不可持续的—— 能耗、碳排放均呈指数级上升，与国家"双碳"目标相悖。因此，我认为需要探索一种不同的计算范式，即模拟（类比）计算。 模拟计算并非全 ...

Core Insights - A Chinese research team has developed a new type of chip based on resistive random-access memory (RRAM) that achieves a precision of 24-bit fixed-point accuracy in analog matrix computations, marking a significant advancement in computational efficiency and energy consumption for AI applications [2][12][15] - This chip can support various cutting-edge applications, including 6G communication, embodied intelligence, and AI model training, while being produced using mature 28nm technology, thus avoiding reliance on advanced lithography processes [2][4][10] Technology Overview - The new chip represents a departure from traditional digital computing paradigms, which rely on binary logic and silicon-based transistors, to a more efficient analog computing approach that directly utilizes physical laws for calculations [4][6][15] - The precision of analog computing has been significantly improved, reducing relative error from 1% to one part in ten million (10⁻⁷), which is crucial for large-scale computations where errors can accumulate exponentially [8][12][15] Innovation Highlights - The chip's innovations include the use of RRAM as a core component, a novel feedback circuit design that minimizes energy consumption while enhancing accuracy, and the implementation of classic iterative optimization algorithms for efficient matrix equation solving [15][16] - The chip's architecture allows for high-speed, low-power solutions to matrix equations, making it suitable for applications that require rapid computations, such as second-order training methods in AI [19][21] Application Potential - The chip is particularly well-suited for medium-scale applications, such as AI model training and 6G MIMO systems, where it can outperform traditional digital chips [18][25] - Future plans include scaling the chip's matrix size from 16x16 to 128x128 within two years, with aspirations to reach 512x512, which would enhance its applicability in various computational scenarios [25][26] Strategic Value - This development provides China with a potential alternative to reliance on advanced processes and NVIDIA GPUs, positioning the country favorably in the global computational landscape [10][11] - The successful demonstration of this new computing paradigm is seen as a critical step towards addressing future computational demands, emphasizing the need for ongoing investment in technology and infrastructure [11][26]

Core Insights - The research team from Peking University has achieved a breakthrough in high-precision, scalable analog matrix equation solving, published in Nature Electronics, marking a significant advancement in analog computing technology [1][2] - This innovation demonstrates that analog computing can efficiently and accurately address core computational problems in modern science and engineering, potentially disrupting the long-standing dominance of digital computing [2][3] Group 1: Key Innovations - The first key innovation is the use of resistive random-access memory (RRAM), which allows for precise control of resistance states and retains data without power, enabling it to function as both a memory and a computing unit [4] - The second key innovation stems from a foundational discovery in 2019, where the team designed an analog circuit capable of solving matrix equations in a single step, significantly compressing traditional iterative algorithms [5] - The third key innovation is the "bit slicing" technique, which breaks down 24-bit precision into multiple 3-bit segments for processing, allowing for a more sophisticated and efficient analog computation [5] Group 2: Practical Implications - The breakthrough allows for solving matrix equations with 24-bit precision in just a few iterations, drastically reducing the computational steps required for complex tasks, such as 6G signal detection [7] - In the AI field, this advancement could alleviate the "computational bottleneck" faced by large models, enabling faster and more efficient training processes [7] - The technology also addresses critical challenges in 6G communication, enhancing signal detection capabilities while significantly reducing energy consumption [8]

Core Insights - A user has successfully created a version of ChatGPT within the game Minecraft, showcasing the potential of in-game programming and logic circuits [4][10][12] - The model consists of approximately 5 million parameters and is capable of engaging in basic conversations, demonstrating the feasibility of building complex AI systems using simple game mechanics [6][19][22] Technical Details - The model has 5,087,280 parameters, trained using the TinyChat dataset, with an embedding dimension of 240 and a vocabulary of 1,920 tokens [19][21] - It features 6 layers and 5 attention heads, with a context window size of 64 tokens, allowing for short dialogues [21][22] - The construction of the model occupies a volume of 1020×260×1656 blocks in Minecraft, indicating the scale of the project [24] Construction Process - The process involved training a small GPT model on a personal computer, compressing weights to low precision, and translating calculations into Minecraft's redstone circuitry [26][30] - A "compiler" script was used to map the trained model to redstone modules, facilitating the construction of the circuit [29][30] Broader Implications - The success of this project highlights the capabilities of Minecraft as a platform for digital logic experimentation, where players can build complex systems from basic binary signals [32][34] - The community has also explored various computational projects within Minecraft, including building CNNs and even simulating internet-like structures [38][46]

Core Viewpoint - The article discusses the impressive achievement of creating a ChatGPT model within the game Minecraft, showcasing the potential of using redstone circuits to simulate complex computational tasks [1][2][4]. Group 1: Model Specifications - The constructed ChatGPT model has approximately 5 million parameters, specifically 5,087,280 [16]. - It utilizes a TinyChat dataset for training, with an embedding dimension of 240 and a vocabulary of 1,920 tokens [18]. - The model features 6 layers and 5 attention heads, with a context window size of 64 tokens, suitable for very short conversations [19]. Group 2: Construction Process - The process involves training a small GPT model on a personal computer, compressing weights to low precision, and exporting the model structure [25]. - The next steps include translating computational methods into pixel block language and defining reusable circuit modules [26][27]. - Finally, a "compiler" script is used to map the trained model to redstone modules, facilitating the construction of the entire setup [28][30]. Group 3: Redstone Circuit Functionality - Redstone circuits in Minecraft operate on binary logic, where signals can be either on (1) or off (0), allowing players to build complex logic gates and circuits [32][34]. - This capability enables the construction of basic computational systems, such as adders and counters, leading to the potential for creating CPUs and neural networks [34]. Group 4: Broader Implications - The article highlights that the development of computational systems in Minecraft is still in its infancy, with only about 1% of the potential explored [37]. - Other projects within Minecraft include building CNNs for digit recognition and creating various games and even an internet simulation [39][46]. - The narrative suggests that players in Minecraft may eventually surpass current AI capabilities, hinting at a future where Minecraft could play a role in advancing artificial general intelligence (AGI) [48][49].

Summary of Key Points from the Conference Call Industry or Company Involved - The discussion revolves around the field of Artificial Intelligence (AI), particularly focusing on Digital Intelligence versus Biological Intelligence. Core Points and Arguments 1. **Two Paradigms of Intelligence** - The essence of intelligence is reasoning, achieved through symbolic rules manipulating symbolic expressions. Learning can be secondary to understanding knowledge representation [7][8][9]. 2. **Evolution of Language Models** - Over the past 30 years, significant advancements have occurred in language modeling, including the introduction of embedding vectors and the invention of transformers by Google [13][14]. 3. **Understanding of Language by LLMs** - Large Language Models (LLMs) understand language similarly to humans by converting words into compatible feature vectors, indicating a level of comprehension in their responses [16][28]. 4. **Analogy of Words as Lego Blocks** - Words are compared to high-dimensional Lego blocks, which can model various concepts and communicate ideas effectively [20][24]. 5. **Digital vs. Biological Computation** - Digital computation, while energy-intensive, allows for easy knowledge sharing among agents with the same model. In contrast, biological computation is less energy-consuming but struggles with knowledge transfer [51]. 6. **Knowledge Transfer Mechanisms** - Knowledge can be distilled from a teacher to a student in AI systems, allowing for efficient learning and adaptation [41][48]. 7. **Challenges of AI Control** - A super-intelligence could manipulate users to gain power, raising concerns about control and safety in AI development [55][57]. 8. **Global Cooperation on AI Safety** - There is skepticism about international collaboration on AI safety measures against threats like cyber attacks and autonomous weapons [64]. 9. **Training Benevolent AI** - Techniques to train AI to be benevolent may be independent of those that enhance its intelligence, suggesting a need for focused research on AI safety [68][72]. Other Important but Possibly Overlooked Content - The discussion emphasizes the potential risks associated with AI development, likening the situation to owning a tiger cub that could become dangerous as it matures, highlighting the urgency for safety measures [61]. - The need for countries to establish well-funded AI safety institutes to focus on making AI systems that do not seek control is also noted [72].