36Kr Exclusive | Peking University-incubated, China's first native robotics "brain chip" startup secures several hundred million yuan in financing 36氪首发 | 北大项目孵化，国内首家原生机器人“大脑芯片”企业获数亿元融资

1. Core Proposition: Solving the "Brain" Chip Trilemma

The article centers on WeFan Intelligence's mission to address a fundamental bottleneck in the robotics industry: the lack of a domestic, high-performance, and energy-efficient computing platform for a robot's central processing unit ("brain"). The core challenge is a three-way trade-off:

• High Computing Power: Required to run complex multi-modal perception, AI inference (like transformers and world models), and decision-making algorithms.
• High Energy Efficiency: Robots, especially mobile and autonomous ones, are power-constrained.
• Manageable Cost & Ecosystem: Solutions must be commercially viable and compatible with mainstream software development tools.

Currently, the market is dominated by NVIDIA's Jetson, which is powerful but expensive and faces localization and commercial deployment hurdles. Domestic alternatives are not yet mature. WeFan's strategy is to tackle this trilemma directly with its BiGPU architecture.

2. Technical Innovation: Homogeneous Fusion, Not Heterogeneous Glue

The key technical insight presented is WeFan's homogeneous fusion approach, which distinguishes it from previous attempts in the industry.

• The Problem with Past Approaches: Earlier efforts often used heterogeneous designs—essentially combining a separate Spiking Neural Network (SNN) accelerator module with a traditional NPU (Neural Processing Unit) or GPU module on the same chip. This creates two distinct systems with potentially separate instruction sets and software toolchains, increasing complexity and development costs.
• WeFan's Solution: The BiGPU: WeFan's innovation is to unify conventional GPU architectures (optimized for matrix-heavy AI workloads like GEMM) with brain-inspired computing principles within a single, coherent architecture. This is likened to integrating the two paradigms at the foundational level.
• Mechanism and Benefit: They leverage the fact that over 80% of neural network computation is matrix multiplication. By using a novel encoding method, they convert standard computations (ANN) into a form of cumulative calculation akin to SNNs. This dramatically reduces data movement and power consumption while preserving computational effectiveness. Crucially, this is achieved under a unified instruction set and software toolchain, ensuring compatibility with mainstream AI frameworks and lowering the barrier for developers.

3. Strategic Significance and Underlying Logic

The company's strategy and the investors' enthusiasm are built on several logical pillars:

• Future-Proofing with Dual Compatibility: WeFan's platform isn't just for today's dominant transformer-based models; it's designed to natively run both classical ANNs and next-generation SNNs, as well as hybrid models. This positions the chip as a long-term platform that can evolve with AI paradigms, potentially towards Artificial General Intelligence (AGI).
• "Using Next-Generation Tech to Solve Today's Problems": This quote from the founder captures the dual logic. The team uses mature insights from brain-inspired computing (a field with long-term AGI promise) to achieve concrete, near-term benefits in power efficiency for current robotic tasks. It's an evolutionary approach rather than a revolutionary bet.
• Leveraging Deep, Cross-Disciplinary Expertise: The technical difficulty is highlighted as understanding both brain-inspired computing and conventional GPU architecture deeply enough to fuse them at the silicon level. The team's pedigree—founded by a veteran with experience at IBM, GlobalFoundries, and leading domestic firms, and backed by members from top tech companies—lends credibility to their ability to execute this complex cross-disciplinary task.
• Alignment with National Strategic Goals: The investor commentary explicitly frames the investment within the context of technological self-reliance and security ("自主可控"). Developing a domestic, high-performance alternative to foreign AI chips is a strategic imperative. WeFan's work is thus positioned not just as a commercial opportunity, but as a contribution to securing China's critical infrastructure in the AI and robotics era.

4. Deeper Implications and Market Outlook

The article and the cluster of investors point to broader industry trends:

• The Shift to Efficient, Edge Intelligence: As noted by investor Han Nan, the global AI narrative is shifting from "high-power, large-scale computing" (data centers) to "ultra-low-power edge intelligence." Robotic brains are the ultimate edge devices, making energy efficiency paramount. WeFan's approach directly targets this macro-trend.
• Embodied Intelligence as the Next Frontier: Embodied AI—the

以下是对这篇文章的深度解读，从市场背景、技术路线、团队与资本逻辑等多个维度进行分析：

一、 市场背景与行业痛点：为何需要一颗新的“大脑”？

具身智能（即有物理实体、能与环境交互的AI，如机器人）的快速发展，对核心计算芯片提出了前所未有的要求。当前的机器人“大脑”芯片市场呈现出两个尖锐矛盾：

1. 外部依赖困境：市场高度依赖英伟达Jetson系列等产品，但面临价格高昂、本地化支持有限、商业化部署门槛高的问题。这对于中国机器人产业的规模化、自主化发展构成潜在风险。
2. 内部供给不足：尽管国产芯片发展迅速，但目前尚无成熟产品能真正满足机器人端侧对“大脑”芯片（需同时处理感知、决策与运动控制）在性能、能效和成本上的综合要求。

这正是维泛智能切入的市场机会：为国产机器人提供一颗性能达标、功耗可控、成本合理的“中国芯”。

二、 技术路线深度解析：BiGPU架构的创新与逻辑

维泛智能的核心解决方案是其自主研发的 “类脑启发式GPU”（BiGPU）架构。这并非简单的技术堆砌，而是一个深刻的架构创新。

• 核心理念：同构融合，而非异构拼接

  • 行业内以往的类脑计算尝试多为“异构”方案，即把类脑计算（SNN）模块和传统计算（NPU/GPU）模块像“拼乐高”一样拼接在一起，本质仍是两套独立系统，需维护两套软件工具链，开发复杂、生态割裂。
  • 维泛智能的 “同构融合” 是更底层的创新。它试图在同一个芯片架构内，将通用GPU的并行计算能力与类脑计算的核心机制深度融合。其最大优势在于可以共享同一套指令集和软件工具链，并兼容主流AI软件生态，从而极大降低了开发和迁移成本。

• 实现路径：从计算原点优化能效

  • 文章揭示了其关键技巧：神经网络超过80%的计算量是矩阵乘累加（GEMM） 操作。传统方式（ANN）需要处理大量高精度数据，功耗和带宽压力大。
  • 维泛智能通过技术手段，将这类计算转化为脉冲神经网络（SNN）的累加形式。SNN以离散的“脉冲”传递信息，更接近生物神经元的工作方式，天然具有事件驱动、低功耗的特性。
  • 通过这种转换，在保留GPU对多种算法框架通用适配能力（灵活性）的同时，引入了类脑计算机制来显著降低功耗与带宽压力（能效）。这直击了“算力—能效—成本”难以兼顾的行业难题。

• 技术前瞻性：立足当下，面向未来

  • 创始人殷积磊点明，这是一种 “用下一代技术解决当下问题” 的思路。BiGPU不仅兼容当前主流的Transformer、VLA（视觉-语言-动作模型）等大模型架构，也能原生支持类脑网络及未来两者融合的新模型。
  • 从更长远看，类脑计算被视为通向通用人工智能（AGI） 的重要路径之一。因此，BiGPU的研发也蕴含着为未来计算范式提前布局的战略眼光。

三、 团队与资本：技术落地的双重保障

• 团队基因：公司核心团队拥有 “产学研用”全链条经验。创始人殷积磊兼具