- Furiosai’s new RNGD server offers 4 Petaflops computed to 3KW for an efficient AI
- Companies will be able to climb work loads without expensive infrastructure changes
- The RNGD server provides compatibility with the OpenAI API along with a growing set of SDK features
The South Korean chips startup furiosaai, which moved away from the acquisition offer of $ 800 million, continues to advance with new products as the efficient infrastructure demand is triggered.
The start is looking to provide hardware companies that can run LLM without the expensive data center updates and heavy energy costs often associated with GPUs.
Its latest product, the RNGD server, is an AI device ready for the company fed by the AI inference chips of Furiosaai (pronounced “renegade”).
More efficient scale
Each system offers 4 FP8 computing petaflops and 384 GB of HBM3 memory, while it operates only 3KW.
In comparison, NVIDIA DGX H100 servers can draw more than 10KW. This means that a 15KW Standard Data Center frame can contain five RNGD servers, while the same frame would fit only to a DGX H100.
Furiosai says that, since most of the data centers are limited to 8KW per rack or less, its design addresses a key barrier to companies.
Advanced AI models that are executed in such environments generally require new cooling and energy systems.
The company says that by adopting the RNGD server, companies can climb more efficiently, while maintaining compatibility with OpenAi API.
The startup recently closed a bridge round of the C series of $ 125 million and expanded its association with LG AI Research.
LG uses RNGD hardware to execute its Exaone models, and says it obtains more than twice the inference performance per watt compared to GPUs.
Furiosaai also recently collaborated with Openai, where the two companies demonstrated the chatbot in real time GPT-Oss of open weight that is executed in just two of the RNGD accelerators of Furiosaai.
The new RNGD server will receive continuous updates to the SDK of Furiosaai, which recently introduced tensioner’s parallelism between chip, new compiler optimizations and extended quantification formats.
The RNGD server is currently being tested with global clients and is expected to be available for order in early 2026.
