"AX Infrastructure Innovation Accelerates…Acryl Strengthens Market Dominance with GPU-Based Cloudization and NPU Expansion"
Next-Generation AI Resource Optimization Solution Efficiently Manages and Utilizes Hundreds to Thousands of GPU Infrastructure
Acryl Accelerates Jonathan Platform's Transition to SaaS and FaaS, Integrating On-Premises and Cloud GPUs
Acryl (CEO Park Oe-jin), a Korean AI company, announced plans to launch a new version of "GPUBase" (product name), a solution designed to maximize the utilization of high-cost GPU-based AI infrastructure, early next year. GPUBase is a core solution of Acryl's integrated AI platform, "JONATHAN," and forms the Jonathan ecosystem alongside the MLOps solution, "FLIGHTBASE" and the LLMOps solution, "AGENTBASE."
GPUBase manages GPU resources across on-premise and cloud environments at the cluster level, stably providing GPU resources required by diverse AI learning and inference workloads. Based on Acryl's proprietary GPU virtualization technology, GPUs can be flexibly managed. It is designed to efficiently share tasks and partition GPUs, and its architecture enables stable resource operation even in large-scale GPU clusters with hundreds to thousands of GPUs.
Furthermore, GPUBase incorporates Acryl's proprietary multi-GPU communication optimization technology. This technology minimizes inter-GPU communication delays and inefficiencies that can occur when utilizing widely used distributed learning and inference frameworks such as NCCL, Horovod, DeepSpeed, and vLLM, thereby providing stable distributed processing performance even in large-scale LLM learning and inference environments.
The new version of GPUBase, scheduled for release early next year, will build on this foundation by adding integrated GPU operation capabilities that extend to multi-cloud. Acryl plans to implement a hybrid operating system that integrates on-premises GPUs and major cloud GPUs such as AWS, GCP, and Azure as a single GPU pool, automatically assigning the optimal GPU based on workload characteristics, cost, and latency.
The expansion of GPUBase is also linked to changes in the service structure of the entire Jonathan platform. Acryl is pursuing a plan to offer FlightBase and AgentBase as SaaS, enabling web-based AI without the need for separate infrastructure. We are establishing a system that utilizes learning, inference, and deployment functions. Furthermore, GPUBase plans to evolve into a FaaS (Function as a Service) architecture that utilizes GPU resources on a call-based basis. We aim for a serverless GPU utilization model where GPUs are "called like functions when needed."
Acryl CEO Park Oe-jin stated, "GPUs represent the largest resource in the overall cost of AI development and operation, so managing and utilizing them is key to a company's competitiveness." He added, "The new version of GPUBase, to be released next year, will set a new standard for AI infrastructure operation based on multi-cloud GPU integration and SaaS/FaaS-based scalability." He added, "Through the advancement of the Jonathan platform, we will create an environment where anyone can easily utilize high-performance AI infrastructure."
Acryl is accelerating the optimization of GPU-based infrastructure as well as securing scalability to NPU-based environments. This technology development is being carried out as part of the "PIM-NPU-based Large-Scale Artificial Neural Network Processing Platform SW Technology Development" project, promoted by the Ministry of Science and ICT and the Institute of Information and Communications Technology Planning and Evaluation (IITP). Going forward, Acrylic plans to advance its GPU/NPU hybrid operation technology and expand its demonstrations in industrial and public sectors, contributing to the strengthening of the domestic AI infrastructure ecosystem.