Daum Overview, the AI-powered search feature on South Korea’s leading web portal, is now running in production on FuriosaAI’s RNGD accelerators.
Powered by Upstage’s flagship Solar large language model, the service delivers real-time search summaries and agentic answers to millions of users. The deployment demonstrates the maturity of the RNGD hardware and software stack to run enterprise-grade, high-traffic AI services at production scale—delivering throughput on par with NVIDIA’s H200 while drastically lowering power consumption and token serving costs.
"When users enter a keyword into Daum’s search engine, the service provides both a concise summary and the supporting sources. As content changes, the AI automatically updates the results so users always receive the latest information,” said AXZ CEO Lee Kun-soo in a webcast with Furiosa Co-Founder and CEO June Paik and Upstage CEO Kim Sung-hoon.
AXZ, which operates the Daum portal in partnership with Upstage, plans to expand coverage into vertical services such as shopping and restaurant recommendations, while ultimately offering personalized AI agents for users.
H200-class performance at lower power and TCO
AI search summaries demand low latency, high output throughput, and strict Service Level Objectives (SLOs) under fluctuating user traffic.
By running Upstage’s Solar LLM on RNGD, Daum achieves:
- H200-class performance: Meets strict interactivity and time-to-first-token (TTFT) requirements for live search queries without throughput bottlenecks
- Superior efficiency: Operates within RNGD’s 180W air-cooled envelope, slashing energy consumption per generated token compared to high-wattage, liquid-cooled GPU clusters
- Lower operating costs: Delivers a significantly lower Total Cost of Ownership (TCO), proving that large-scale consumer AI search can be served sustainably and profitably
A fully integrated Sovereign AI stack
This deployment is one of the first commercial realizations of a complete domestic Sovereign AI stack operating at scale:
- Application: Daum Overview search interface (Kakao / Daum)
- Model: Upstage Solar LLM, tailored for high-accuracy Korean and English contextual understanding
- Silicon & Software: RNGD accelerator cards running the furiosa-llm serving framework
Currently hosted on Furiosa’s Data Center Platform infrastructure, the workload is scheduled to migrate to Samsung SDS as part of their commercial NPUaaS expansion. (Watch the recorded Korean-language deep dive here.)
Zero-code migration to RNGD
Porting the Solar model to RNGD required zero architectural code refactoring. Utilizing our furiosa-llm framework (a drop-in replacement for vLLM with native PyTorch integration), Upstage and Furiosa engineers compiled and optimized the model graph directly onto RNGD’s Tensor Contraction Processor (TCP) primitives in days.
What’s next
Building on the success of Daum Overview, Furiosa, Upstage, and AXZ are expanding their technical collaboration:
- Evaluating the migration of Daum’s dense embedding models from legacy CPU instances over to RNGD NPUs to further reduce infrastructure overhead
- Leveraging RNGD’s multi-LoRA capabilities to serve specialized downstream search tasks on shared base weights
RNGD is available for enterprise deployments across global cloud and on-premise data centers.
It entered mass production with TSMC in January, with 20,000 chips on track for delivery this year and 2-3x more volume in 2027. Earlier this month, Samsung SDS announced the launch of their RNGD-powered NPU-as-a-Service (NPUaaS).
To learn more about deploying LLM and agentic workloads on RNGD, contact sales@furiosa.ai or visit access.furiosa.ai.
Written by
The Furiosa Team
.png)
