SenseTime SenseCore splits Prefill and Decode across heterogeneous GPU clusters
Overview
SenseTime SenseCore CTO Xuan Shanming says the company's 10,000-GPU heterogeneous clusters at its Lingang AIDC separate Prefill and Decode inference stages across mixed domestic and high-bandwidth GPUs, linked by KV Cache over 800G, 400G and 200G fabrics.
In one test configuration with the same number of flagship GPUs as a baseline, the company reports about five times higher token throughput after heterogeneous PD scheduling. Xuan presented these results at the 10th CSDI Summit in Shenzhen on October 16.
The company also says its platform serves about 6 trillion tokens per day and targets 10 trillion by Q4 2026. These throughput and capacity figures are company statements and have not been independently verified.
Written by AI from the articles below · updated Oct 19, 3:41 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- PandailyNewsSenseTime details Prefill/Decode split clusters for agentic inference
AISenseTime SenseCore CTO Xuan Shanming says its 10,000-GPU heterogeneous clusters split Prefill and Decode across mixed domestic and high-bandwidth GPUs, linked by KV Cache over 800G, 400G and 200G fabrics. The company reports about five times higher token throughput than a baseline with the same count of flagship GPUs, and says its platform serves about 6 trillion tokens per day, targeting 10 trillion by Q4 2026. These throughput and capacity figures are company statements.
Heat Trend
Not enough continuous observations to show a trend yet.