The team from Moonshoot and Tsinghua University were handed the Erik Riedel Best Paper Award for their system named Mooncake at the USENIX FAST conference in California last month.
Their large language model (LLM) platform Mooncake helps to reduce dependency on graphics processing units (GPUs) and is up to five times more efficient than the previous baseline, helping to reduce the demand for computing power.
Mooncake powers the Kimi chatbot, processing 100 billion tokens – a basic unit of text that enables the model to process, understand and generate language – each day.
The measures were intended to limit China’s computational power in AI and affect the training and development of AI models.
