北京大学：神经形态芯片性能超越英伟达A100
北京大学神经形态芯片在特定AI任务上以478倍性能超越Nvidia A100，通过消除内存瓶颈重塑AI行业影响。
原文链接详细分析
北京大学研究人员推出神经形态芯片，在实时神经表面重建任务上比Nvidia A100快达478倍，核心在于将存储与计算集成在同一内存阵列中。该架构消除了传统GPU在独立内存与处理单元间传输数据的瓶颈，直接缓解了AI训练成本高、能耗大的根源。尽管478倍优势仅限于特定工作负载，且Nvidia H100或B200在通用任务上仍占主导，但这一40纳米工艺成果表明，在出口限制环境下仍可实现前沿性能，印证了DeepSeek一月观点：更好架构即可匹配前沿表现并引发市场剧烈重估。
Bull Theory
@BullTheoryioResearch, Trades, onchain plays and all other crypto stuff simplified.Publishes institutional-grade cryptocurrency research and blockchain market intelligence. Delivers in-depth analysis of on-chain metrics, tokenomics, and decentralized finance (DeFi) ecosystems. Features proprietary data models, investment thesis breakdowns, and macro-level crypto trend forecasts. Provides strategic insights for sophisticated investors navigating digital asset markets. Maintains rigorous methodology in fundamental and technical analysis across crypto assets.
Premium Sponsors
.jpg)
