HyperAccel to Showcase Eight-LPU AI Inference Server at SC26, Claims 3.5 Times Higher Efficiency Than NVIDIA GPUs

16 November 2026 | NEWS

The South Korean fabless semiconductor company will demonstrate its LPU-based server technology at Supercomputing 2026 in Chicago, from 17 to 19 November. Internal benchmarks indicate more than 3.5 times higher token throughput per watt and per dollar compared with systems using NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, highlighting its focus on reducing AI inference costs and power consumption.

Eight-LPU server system and live remote demonstration of LLM inference running on operational servers in Korea

Internal benchmarks show more than 3.5x higher token throughput per watt and per dollar compared with systems powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs

HyperAccel, a South Korean fabless AI semiconductor company specializing in high-efficiency large language model (LLM) inference, will showcase its LPU (LLM Processing Unit) technology at SC26 (Supercomputing 2026), taking place November 17–19 at McCormick Place in Chicago. The company will present an eight-LPU server system and conduct a live remote demonstration of LLM inference running on operational servers deployed in Korea.

At the exhibition, HyperAccel will showcase how its LPU technology operates as part of a server-level system, extending beyond individual semiconductor components to practical AI infrastructure. Through a live remote demonstration from its server environment in Korea, visitors will be able to see LLM inference running on operational hardware and gain insight into the technology's potential for deployment in enterprise and data center environments.

According to HyperAccel's internal benchmarks, a server equipped with HyperAccel LPU cards (250W per card) delivers more than 3.5x higher token throughput per watt and per dollar than a comparable system powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs (600W per card) when running Qwen3-32B and Llama3-70B models. With support for even larger LLMs, including Kimi and GLM models, HyperAccel's LPU demonstrates the potential to significantly improve the energy efficiency and economics of LLM inference, offering a compelling solution for enterprise AI infrastructure and data centers.

These efficiency gains stem from the HyperAccel LPU architecture, designed specifically for LLM inference. It pairs power-efficient LPDDR5X memory with HyperAccel's Streamlined Dataflow architecture, which achieves approximately 90% memory bandwidth utilization, compared with 50 to 60% for GPUs, reducing power consumption and operating costs in large-scale inference environments.

In addition to its main exhibition, Joonyeon Kim, CFO of HyperAccel, will deliver a presentation titled "Redefining AI Infrastructure Through Token Efficiency" at Lablup Inc.'s booth (#4630) on November 18, from 2:30 p.m. to 2:45 p.m. Central Time.

Through its participation in SC26, HyperAccel aims to expand its global presence and accelerate the adoption of LPU-based inference infrastructure. The company is seeking partnerships with global server manufacturers, cloud service providers, and AI infrastructure solution providers, focusing on server platform integration, technical validation, and business development opportunities in global data center markets.

"As AI adoption accelerates, the industry is shifting its focus from raw processor performance to the efficiency and economics of AI inference at scale," said Joo-Young Kim, CEO of HyperAccel. "At SC26, we are excited to showcase how our purpose-built LPU architecture delivers superior token efficiency at the server level, reducing the cost and power consumption of AI inference. We also look forward to engaging with global infrastructure partners to bring more efficient and scalable AI solutions to market."

Visitors can meet the HyperAccel team at the AI Semiconductor Korea Pavilion (Booth #1411) to learn more about its LPU technology, explore the eight-LPU server system, and watch the live remote LLM inference demonstration.