HyperAccel

Stand: 1239
HyperAccel is an AI semiconductor startup pioneering inference-first silicon for the generative AI era. As inference becomes the dominant AI workload, infrastructure designed for training is creating an unsustainable burden on data center power and economics. Founded in 2023 by Dr. Joo-Young Kim, HyperAccel was built to solve this problem at the silicon level with the LPU (LLM Processing Unit), a processor designed exclusively for LLM inference.

The HyperAccel LPU, fabricated on Samsung's 4nm process, adopts LPDDR5X memory in place of conventional HBM and combines it with a proprietary Streamlined Memory Access (SMA) architecture. This achieves approximately 90% memory bandwidth utilization, compared to 50 to 60% on standard GPU architectures, delivering significant gains in power efficiency and total cost of ownership over conventional GPU-based inference. Alongside its datacenter LPU, HyperAccel is developing an edge variant targeting robotics, kiosks, and on-device AI applications, scheduled for release in late 2026.

HyperAccel's technical leadership has been recognized on major international stages, including Best Paper awards at IEEE/ACM DAC and IEEE Micro, Best Exhibit at IEEE VLSI Design, and selection among the global top 8 in the AI Core Technologies & Chips category at IC Taiwan Grand Challenge. Backed by $45M in cumulative funding and partnered with AMD, Samsung, LG Electronics, NAVER Cloud, Advantech, and Inventec, HyperAccel is delivering the infrastructure layer that makes large-scale generative AI economically and operationally sustainable.

Visit website (opens in a new tab)
Loading

Contact Exhibitor