Denoticia’s Seahorse Storage Tops MLPerf Vector DB Performance
IT DAILY ·
✦ AI Summary
Denoticia said on the 3rd that its Seahorse Storage ranked first in throughput in the MLPerf Storage v3.0 vector DB evaluation.
The evaluation was jointly conducted by Denoticia and the Korea Telecommunications Technology Association (TTA) as part of the Ministry of Science and ICT- and Institute for Information & communications Technology Planning & Evaluation (IITP)-supported HPC Innovation Hub project.
Denoticia was the only submitting organization to carry out the full evaluation with a single, multi-node distributed storage configuration, posting 57,619 QPS in a 4-node client configuration in the vector DB performance category.
Denoticia announced on the 3rd that its Seahorse Storage (Seahorse Storage) recorded the top throughput in the vector DB evaluation of MLPerf Storage v3.0. MLPerf Storage v3.0 is an AI storage performance benchmark run by MLCommons.
The evaluation was jointly conducted by Denoticia and the Korea Telecommunications Technology Association (TTA). The joint effort was carried out as part of the Ministry of Science and ICT- and Institute for Information & communications Technology Planning & Evaluation (IITP)-supported HPC Innovation Hub project.
Version 3.0 added evaluation items for model storage and recovery for AI training, KV cache (Cashe), a factor affecting LLM inference performance, and vector DB for retrieval-augmented generation. Nineteen organizations took part in this evaluation, submitting 150 results based on 99 system configurations. Only 4 organizations submitted results for all 4 categories — AI training, model storage and recovery, KV cache, and vector DB — and Denoticia was among them.
Denoticia was the only submitting organization to carry out the full evaluation with a single, multi-node distributed storage configuration. As a result, Seahorse Storage ranked first in vector DB processing performance, posting 57,619 QPS in a 4-node client configuration. QPS refers to the number of search requests a system can process per second.
In the Chackpointing evaluation, it recorded a read performance of 114.73 GiB per second based on the Llama 3 70B model. Checkpointing is an item that evaluates the periodic saving and reloading of LLM training state, and it also measures model data read performance under a given network performance environment.
Seahorse Storage provides integrated support for data processing for LLM inference and vector search. Denoticia applied an open-source distributed storage foundation based on Ceph to strengthen its ability to respond to a network environment for high-speed, large-volume data transfer between servers. Seahorse Storage is scheduled for release in the second half of this year.
Han Joo-yeon, team leader of TTA’s AI Infrastructure Verification Team, said that MLPerf Storage v3.0 is the first case in which the evaluation scope has expanded beyond the existing focus on AI training to include inference and search, and linked it to a meaningful case of validating domestic storage technology against an international benchmark.
Jeong Moo-kyung, CEO of Denoticia, said the soon-to-be-released Seahorse Storage has demonstrated top-tier inference, search performance, and versatility in MLPerf Storage. He added that the company plans to support AI workloads on a single data infrastructure with Seahorse Storage, aiming to improve the efficiency of enterprise AI build and operations.
Source: IT DAILY · Kim Ho-jun
Original: https://www.itdaily.kr/news/articleView.html?idxno=241380
References
This article was produced with the help of an automated content generation algorithm.
Source: IT DAILY
View originalThis article was summarized and organized by BizCrush based on the original article from IT DAILY. For exact quotations and full details, please refer to the original article.