[Edaily Reporter Hyera Lee ] #Acryl Inc. announced on the 1st that it has completed Phase 1 (Horizontal Phase) of “K-Scale Evaluation,” a large-scale performance validation program for its artificial intelligence (AI) infrastructure operations platform, “GPUBase.”

The company validated AI training, inference, and operational performance using a total of 1,272 GPUs across seven types in the environments of the world’s top three cloud service providers (CSPs).
The validation results showed that environments using GPUBase achieved processing speeds up to 24 times faster than those without it during large language model (LLM) training, with training time reduced by approximately 96%. The company also explained that during the inference phase, it maintained stable response performance even under high-load conditions, and data throughput during periods of high data concentration between GPUs increased by approximately 24 times (2,375%) compared to environments without GPUBase.
In terms of operations, the solution achieved a GPU utilization rate of over 90% across seven different GPU environments, reducing task wait times by up to 93% and total task completion times by up to 34%. It also verified the operational stability of AI data centers by detecting and recovering from all 18 types of failure scenarios within five minutes.
GPUBase’s core technology is based on over 25 years of computer network research conducted by Dr. Yeom Ik-jun, Acryl Inc.’s Chief Technology Officer (CTO) and a professor in the Department of Software Engineering at Sungkyunkwan University.
Acryl Inc. was previously selected as the lead organization in the network sector for the Ministry of Science and ICT’s “Technology Development Project to Strengthen AI Cloud Competitiveness.” The company plans to target the market for AI data center infrastructure platforms that integrate and operate GPUs, servers, storage devices, and networks as a single system through GPUBase.
Dr. Yeom Ik-jun, CTO of Acryl Inc., stated, “As this validation has confirmed that GPUBase operates stably even in global cloud environments and GPU setups comprising 1,000 units, we will continue to verify its scalability and stability in larger clusters and long-term operational environments.”