Sign in Subscribe
Artificial intelligence

Breakthrough in AI Computing: Peking University Team Develops Efficient Sorting Hardware Architecture

Strategic Insights

— 1 min read

A research team at Peking University has made a significant breakthrough in artificial intelligence (AI) computing by developing the world's first efficient sorting hardware architecture based on storage-computing integration technology. The team, led by Professor Yang Yuchao and Researcher Tao Yaoyu, has overcome the bottleneck problem of low computing efficiency in traditional computing architectures when facing complex nonlinear sorting problems. This achievement is expected to provide more efficient computing support for AI applications such as embodied intelligence, large language models, intelligent driving, smart transportation, and smart cities.

Key Takeaways:

  • The team has developed a highly parallel bit-reading mechanism based on a new in-memory array structure, which has achieved an order of magnitude improvement in sorting speed and energy efficiency.
  • The hardware solution can achieve a computing speed increase of more than 15 times in typical sorting tasks, but the power consumption is only 1/10 of that of traditional CPU or GPU processors.
  • The new architecture is particularly suitable for task environments that require extremely high real-time performance, such as AI reasoning scenarios where the reasoning response speed under dynamic sparsity can be increased by more than 70%.
  • The team has overcome multiple core technical problems in the storage and computing integration architecture, including complex logic, nonlinear operation, irregular data access, and high dependence on complex comparator networks.
  • The research has achieved remarkable results in numerical computing links with strong regularity such as matrix calculation, but the current mainstream international storage and computing integration architecture cannot solve the big data sorting problem.

Statistics:

  • 15 times: computing speed increase in typical sorting tasks
  • 1/10: power consumption compared to traditional CPU or GPU processors
  • 70%: increase in reasoning response speed under dynamic sparsity in AI reasoning scenarios
  • 10 times: improvement in energy efficiency
  • 2025:presumably the year the study was released

Sources:

  • CEEA (China Electronics Electronics Association) news release: "Peking University Team Develops Efficient Sorting Hardware Architecture"
  • China Electronics News interview with Researcher Tao Yaoyu: "Sorting is the Most Time-Consuming Basic Operation in Artificial Intelligence Systems"

Read more