HyperAccel Joins Upstage Consortium, Collaborating on Independent AI Foundation Model Infrastructure

AI semiconductor startup HyperExcel is joining the government's independent AI foundation model cooperation ecosystem through the Upstage Consortium.

HyperExcel designs and develops AI semiconductors specifically for LLM inference, optimized for real-world application environments of generative AI. To address computational bottlenecks and power and operating cost issues arising during the service phase of ultra-large language models, HyperExcel has developed an LPU (LLM Processing Unit) based on its proprietary architecture.

Through its participation in this consortium, HyperExcel plans to collaborate on enhancing and streamlining the inference infrastructure for the widespread adoption of proprietary AI models. Based on LPU design and performance improvements specialized for LLM inference, and research and development of inference acceleration technology optimized for generative AI application environments, HyperExcel will move forward with implementing a high-efficiency, low-cost AI service infrastructure.

Additionally, the company will primarily develop and deliver a high-availability (HA) full-stack software stack based on vLLM and PyTorch. This will establish an infrastructure that simultaneously ensures inference performance and stability.

HyperExcel plans to comprehensively enhance the processing efficiency, power-to-performance ratio, and operational stability required for commercializing ultra-large models by building a structure that integrates hardware and serving software beyond simple semiconductor supply.

This is explained as supporting the operation of the reader AI foundation model in an economical and scalable manner in actual data centers and industrial service environments.

Kim Joo-young, CEO of HyperExcel, said that the competition in the ultra-large-scale AI depends not only on the size of the model, but also on how efficiently it can be serviced, and that he would contribute to securing technological independence of domestic AI infrastructure and global-level cost competitiveness through LPU-based inference acceleration technology.

• See more related articles