IBM 近日发布名为 NorthPole 的新 AI 芯片，该芯片的设计灵感来自人脑。NorthPole 芯片模仿了人脑的白质连接，这些白质连接连接了大脑皮层的不同部分。
据 BM 表示，NorthPole 比市场上任何其他芯片更节能、更节省空间，并且延迟更低，其速度比其前身 TrueNorth 快约 4000 倍。以 ResNet-50 模型为基准，NorthPole 的能效比领先的 12 纳米 GPU 和 14 纳米 CPU 高出 25 倍，延迟也低得多。
NorthPole 芯片是由 IBM 在加利福尼亚州的圣何塞研究所开发的，该架构在能效、速度和可扩展性方面具有重大突破。NorthPole 芯片紧密集成了处理单元和内存，从而极大地提高了数据的移动效率。
NorthPole 芯片在执行任务时，几乎不需要频繁访问外部内存，因此能效极高。该芯片通过在每个核心中集成内存来缓解冯·诺依曼瓶颈问题。
NorthPole 芯片运行多层的神经网络，用于识别数据中的模式。它的优势在于其板载内存，但这也使其局限性，例如必须从其他地方获取数据时，其效率就会受到影响。为了解决这个问题，NorthPole 采用了“横向扩展”方法，通过将更大的神经网络分割成更小的、可管理的子网络来管理它们。
因此，NorthPole 芯片适用于人工智能推理，特别是实时传感处理。在试验阶段，由于部分资金来自美国国防部，该芯片主要用于计算机视觉任务，包括对象识别、追踪和场景解析等。
然而，NorthPole 的多功能性也延伸到自然语言处理和语音识别等领域。这意味着它不仅可以用于图像和视频分析，还可以用于理解和解析人类语言，这在聊天机器人、语音助手或自动翻译服务中可能非常有用。
自主系统、机器人学、监视以及其他需要即时分析高带宽数据的边缘用例（edge use cases）都是理想的应用场景。在这些场景中，快速和准确的数据分析是至关重要的。例如，NorthPole 芯片可以使自动驾驶汽车更好地处理复杂的现实世界场景，如交通拥堵、行人穿越等。
总体而言，NorthPole 芯片在设计上具有高度的灵活性和速度，使其适用于多种需要实时数据处理和分析的应用场景。这种多功能性和速度优势使其在各种实时应用中具有巨大的潜力。
英文标题：IBM Develops NorthPole AI Chip Inspired by Brain’s White Matter Connections
关键词：IBM，NorthPole，brain，white matter connections，efficiency，speed，delay，ResNet-50，neural network，real-time sensor processing
新闻内容：
IBM has recently released a new AI chip called NorthPole, inspired by the brain’s white matter connections. NorthPole chip closely integrates processing and memory, significantly improving data mobility efficiency.
According to BM, NorthPole is more energy-efficient and saves more space compared to other chips in the market, with a speed of up to 4000 times faster. With ResNet-50 model as the basis, NorthPole’s energy efficiency is 25 times higher than leading 12-nanometer GPUs and 14-nanometer CPUs. Delay is also much lower.
NorthPole chip, developed by IBM in San Jose Institute of California, has significant breakthroughs in terms of energy efficiency, speed, and scalability. NorthPole chip closely integrates processing and memory, greatly improving data mobility efficiency.
NorthPole chip operates multiple layers of neural networks for recognizing patterns in data. Its advantage lies in its on-chip memory, but this also limits its potential, as it must be supplemented by data from other sources. To solve this problem, NorthPole uses “horizontal extension” method, dividing larger neural networks into smaller, manageable sub-networks.
Thus, NorthPole chip is suitable for artificial intelligence inference, especially real-time sensor processing. In the experimental stage, due to partial funding from the U.S. Department of Defense, the chip is mainly used for computer vision tasks, including object recognition, tracking, and scene understanding.
However, NorthPole’s multifunction also extends to natural language processing and speech recognition. It can be used for understanding and analyzing human language, making it useful for chatbots, voice assistants, or
【来源】https://www.maginative.com/article/ibm-unveils-northpole-a-breakthrough-ai-chip-architecture-enabling-massive-efficiency-gains/
Views: 3