Dynotisia's Seahorse Storage Tops MLPerf Storage Vector DB Performance

Dynotisia said it is accelerating its push into the AI infrastructure market after ranking No. 1 in vector DB processing performance in a global AI storage benchmark. On the 3rd, the company announced that its AI storage system, Seahorse Storage, took part in the vector DB evaluation in MLPerf Storage v3.0. In that evaluation, Seahorse Storage recorded 57,619.8 QPS, placing first by throughput.

The evaluation was carried out as part of the "HPC Innovation Hub" project supported by the Ministry of Science and ICT and the Institute for Information & Communications Technology Planning & Evaluation (IITP). It was conducted jointly with the Telecommunications Technology Association (TTA), which also participated as a submitting organization.

MLPerf's official results list the system name as "seahorse-storage." Dynotisia said the achievement demonstrates its AI storage system's vector DB performance in a global benchmark.

MLPerf Storage, a global AI storage performance benchmark overseen by MLCommons, added KV Cache and vector DB evaluations in v3.0 to its existing model storage and recovery tests focused on AI training. KV Cache is used for LLM inference, while vector DB serves as core infrastructure for RAG and AI search. As a result, the evaluation scope has expanded from AI training to include model inference and search.

This round drew 19 organizations, which submitted 150 results across 99 system configurations. The workloads consisted of four types: AI Training, model storage and recovery (Checkpointing), KV Cache, and vector DB. Only four organizations submitted results for all four workloads.

Dynotisia was one of the organizations that completed all four workloads, using a single multi-node distributed storage configuration. It was especially the only case that carried out all four workloads with one multi-node distributed storage configuration. This demonstrated the versatility of a single storage system in supporting AI training, model storage and recovery, LLM inference, and vector search.

In the vector DB evaluation, Dynotisia recorded 57,619.8 QPS with a four-node client configuration, taking first place in throughput. QPS is a metric for the number of search requests a storage system can process per second.

An increase in QPS means greater search request throughput over the same period. QPS is used as a key performance indicator for large-scale RAG and AI search services.

Dynotisia also proved its performance in the checkpointing evaluation, in addition to vector DB. Checkpointing is used to save the training state of large AI models and restore it when needed. In the checkpointing read evaluation for the Llama3 70B model, it recorded 114.73 GiB/s. That corresponds to 82.1% of the test environment's theoretical storage-side network bandwidth of 139.7 GiB/s.

Dynotisia said it secured high-speed read performance for large-scale model data by leveraging network performance, enabling recovery of training after failures and resumption of interrupted training.

The technology is based on the open source distributed storage Ceph and is designed to support a 400G InfiniBand network environment. It also directly implemented a Ceph full stack with Native RDMA support between server and client.

RDMA is a technology that transfers data directly to another server's memory over the network. It is considered an important technology for building high-performance AI infrastructure because it reduces CPU involvement during data transfer and minimizes unnecessary data copying.

Dynotisia said its Native RDMA-based Ceph full stack enables high-speed processing of large-scale AI data even in distributed storage built on general-purpose servers.

Seahorse Storage supports data processing for AI training, model storage and recovery, LLM inference, and vector search, integrating all of these in a single distributed storage system.

As the scope of AI infrastructure expansion has widened from large-scale GPU- and accelerator-centric systems to include data movement and storage systems, the role of storage is also expanding. In particular, as LLM inference and RAG services spread, high-speed access performance for KV Cache and vector data, in addition to the model itself, is drawing attention as a factor that determines the efficiency of AI service processing.

Against this backdrop, Dynotisia emphasized that Seahorse Storage makes it possible to run a variety of AI services from a single data infrastructure without building separate storage for each workload. It said this could simplify enterprise AI infrastructure deployment and operations while improving cost efficiency.

Han Joo-yeon, head of the TTA AI Infrastructure Verification Team, said MLPerf Storage v3.0 is meaningful in that it is the first round to expand the evaluation scope from AI training to inference and search. She also described it as a case of international benchmark validation for domestic storage technology and said TTA plans to continue supporting the objective verification of technology capabilities by domestic companies in a standards-based environment.

Jung Moo-kyung, CEO of Dynotisia, said the MLPerf Storage results confirmed Seahorse Storage's performance and versatility ahead of its release. MLPerf Storage is a global official benchmark for AI storage, and Seahorse Storage has demonstrated top-tier inference and search performance in MLPerf Storage, as well as versatility.

Jung said that as AI infrastructure bottlenecks are shifting toward memory systems and data processing, AI storage is emerging as core infrastructure. He added that the company plans to support a range of AI workloads through Seahorse Storage on a single data infrastructure, aiming to improve enterprises' AI deployment and operational efficiency.

Source: TECHWORLD · Park Gyu-chan
Original: https://www.epnc.co.kr/news/articleView.html?idxno=406467

References

• https://www.epnc.co.kr/news/articleView.html?idxno=406467