AI BriefingLG AI Research Achieves 2.25x Improvement in LLM Inference Performance
Key point
LG AI Research adopted FuriosaAI's RNGD accelerator, boosting LLM inference efficiency for its EXAONE models by 2.25x compared to GPUs.
Details
LG AI Research has adopted FuriosaAI's AI accelerator, RNGD (Renegade), as the inference solution for its EXAONE models. This decision came after months of rigorous evaluation of performance, energy efficiency, and the software stack.
RNGD delivers 2.25x better LLM inference performance per watt compared to GPUs, meeting all of LG AI Research's demanding latency and throughput requirements. This gives the company an alternative that addresses the high power consumption and cost issues of existing GPUs.
Key performance and efficiency results are as follows:
- Energy Efficiency: 2.25x improvement in performance per watt compared to GPU-based solutions
- Compute Density: Generates 3.75x more tokens than GPU racks within the same power limit
- EXAONE 3.5 32B Performance: Using 4 RNGD cards, achieved 60 tokens/s at 4K context and 50 tokens/s at 32K context
Furiosa and LG AI Research plan to supply the RNGD Server to enterprise customers across various industries including electronics, finance, telecommunications, and biotech. This supports the building of Sovereign AI environments where enterprises can own and control their own AI stack.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.