Baidu unveils tianchi 1024-card supernode capable of training 5 trillion parameter models, slated for next year's market debut

During the opening ceremony of Baidu World 2026, Baidu Group Executive Vice President Shen Dou introduced the Tianchi 1024-Card Supernode. This advanced computing cluster is scheduled for market release in 2027. The system incorporates industry-leading optical engine technology, boosting inter-card bandwidth by three times compared to its predecessor, resulting in a 2.5-fold increase in training performance and a threefold rise in inference capability. The supernode is equipped with a G3.5 KV Cache memory acceleration cabinet, which enhances cache reuse. This feature improves long-context inference performance by a factor of two, while simultaneously reducing the Time to First Token (TTFT) by 90% in multi-turn dialogue scenarios and increasing overall throughput by seven times. A single unit possesses the capacity to train models with 5 trillion parameters and supports horizontal scaling up to a ten-thousand-card cluster. This announcement marks an acceleration in Baidu's timeline, as the company had previously announced the Tianchi thousand-card supernode for 2028. Additionally, the Kunlunxin N series has also been moved up to a 2028 release date.