[For more detailed corporate information on the startup featured in this article, visit the Unicorn Factory big data platform ‘DataLab’.]
HyperExcel, a fabless semiconductor company specializing in inference for large language models (LLMs), has begun full-scale proof of concept (PoC) testing for its first-generation semiconductor ‘Bertha.’ HyperExcel plans to expand the scale of PoC to rack units within this year and also conduct PoC tests overseas, including in Japan.
Kim Ju-young, CEO of HyperExcel, held the ‘K-Cloud Tech Day’ on the 27th at the Seoul AI Hub in Gangnam-gu, Seoul, and disclosed the status of the ‘K-Cloud Technology Development Project’ conducted with the Institute for Information & Communications Technology Planning & Evaluation (IITP). As a national R&D project involving Rebellions, Fanea, Mango Boost, Rableup, SqueezeBits, Naver Cloud, and KAIST, HyperExcel serves as the lead research institution.
CEO Kim stated, “We received Bertha in March this year, performed bring-up (semiconductor operation work), and began server-level PoC.” Bertha is an AI semiconductor specialized for LLM inference. It is manufactured at Samsung Electronics’ 4-nanometer factory. Unlike GPUs (graphics processing units) or NPUs (neural network processors) that use HBM (high bandwidth memory), Bertha uses LPDDR (low power DRAM), which is primarily used in smartphones, while delivering performance comparable to NVIDIA’s mid-to-high-end GPU, the RTX 6000.
Thanks to its LPDDR configuration, Bertha’s power consumption is reduced to one-third of that of comparable devices. Because LPDDR itself is inexpensive, the product price is also only one-quarter that of the RTX 6000. The slow speed of LPDDR was overcome through architecture. Unlike GPUs that continuously exchange data with memory, HyperExcel developed an architecture that streams data in one direction, achieving fast computing speeds even with slower memory.
HyperExcel also disclosed measured performance results on the same day. When running Alibaba’s ‘Qwen3-32B’ model using four Bertha chips, processing performance was approximately 40% higher than that of the NVIDIA RTX 6000. Performance per watt and performance per dollar were 3.7 times higher, respectively. CEO Kim noted, “Because the chip unit price is low, configurations with more cards within the same budget are also possible.”
HyperExcel plans to continuously expand its PoC efforts going forward. Recent strategic investment decisions by Krafton and validation tests with Japanese data center companies have also been announced.
Full-scale PoC testing is also expected to begin through the K-Cloud Technology Development Project. Now entering its second year, this R&D project aims to implement a ‘composable cluster’ starting next year, integrating technologies and components developed by participating companies into a single server and bundling inference and training AI semiconductors along with CXL (Compute Express Link) memory as separate resources.
CEO Kim said, “Our goal is to combine HyperExcel’s Bertha, Rebellions’ training NPU, Fanea’s CXL memory, Mango Boost’s DPU system, and Rableup’s infrastructure software into a single server for service delivery starting next year,” adding, “We will build K-Cloud infrastructure that effectively consolidates and services all domestic AI semiconductors.”
[MoneyToday Startup Media Platform Unicorn Factory]
in FUTURE
