SenseTime SenseCore details Prefill/Decode split clusters for agentic inference

SenseTime SenseCore outlined Prefill/Decode separation inside 10,000-GPU heterogeneous clusters at its Lingang AIDC at the CSDI Summit in Shenzhen. In a test with the same number of flagship GPUs as a baseline, heterogeneous PD scheduling delivered about five times higher token throughput. The platform now serves roughly 6 trillion tokens per day and targets 10 trillion by Q4 2026.

• Prefill runs on domestic accelerators, Decode on high-memory-bandwidth GPUs
• State is exchanged via KV Cache links over mixed 800G, 400G and 200G fabrics
• One Decode node can be fed by 30–50 Prefill nodes, so a reserve of high-end cards is kept for failover
• Target is 10 trillion tokens per day by Q4 2026, up from about 6 trillion now