Dinotisia’s Seahorse Storage ranks first in global MLPerf Vector Database benchmark

Dinotisia, a company specializing in long-term memory AI and semiconductor solutions, announced that its AI storage system, ‘Seahorse Storage’, ranked first in throughput for the Vector Database category of the global ‘MLPerf® Storage v3.0’ benchmark.

The evaluation, conducted in collaboration with the Telecommunications Technology Association (TTA) as part of the ‘HPC Innovation Hub’ project supported by the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation (IITP), involved 19 institutions. Dinotisia was the only participant to complete all four workload evaluations—AI training, model storage, recovery, and LLM inference/vector search—using a single multi-node distributed storage configuration.

In the Vector Database assessment, the system recorded 57,619.8 QPS (queries per second) with a 4-node client configuration. Additionally, for the Llama3 70B model, it achieved a checkpointing read performance of 114.73 GiB/s, representing 82.1% of the theoretical bandwidth. The system utilizes a custom-implemented Ceph full-stack supporting Native RDMA to minimize CPU intervention and data copying in 400G InfiniBand environments.