Furiosa RNGD Powers Daum's AI-Generated Search Overviews

Daum Overview has gone into commercial operation based on FuriosaAI's RNGD accelerator. Leveraging Upstage's Solar LLM, it provides real-time search summaries and agentic answers to millions of users, generating about 500 million tokens per day.

RNGD delivers the low latency and high output throughput required for search services while achieving NVIDIA H200-class performance. At the same time, it operates in a 180W air-cooled environment, lowering energy consumption per token and total cost of ownership (TCO) compared to high-power, liquid-cooled GPU clusters, with TCO reduced by more than 50%.

This deployment is a commercial example of a Sovereign AI stack built from domestic (Korean) technology.

• Application: Kakao/Daum's Daum Overview
• Model: Upstage Solar LLM, tuned for Korean and English context understanding
• Accelerator/software: RNGD and the furiosa-llm serving framework

Workloads currently running on Furiosa's Data Center Platform are set to be migrated in line with the expansion of Samsung SDS's commercial NPUaaS. The Solar model was compiled and optimized for RNGD's Tensor Contraction Processor (TCP) primitives within a few days via furiosa-llm, a drop-in replacement framework for vLLM, without modifying the architecture code.

FuriosaAI, Upstage, and AXZ are also considering migrating Daum's dense embedding model from existing CPU instances to RNGD NPUs going forward, and exploring the use of RNGD's multi-LoRA capability to serve specialized search tasks from a single base weight. RNGD entered TSMC mass production in January, with a target of shipping 20,000 units this year and expanding volume by 2-3x in 2027.