Technology

Aim for The Most Sustainable and Efficient AI Inference

Optimized for power-efficient and cost-effective inference without compromising on performance by
integrating our LPU technology

Featured Technology

Streamlined Memory Access

Perfectly aligned bandwidth between memory and compute with minimal data buffering and reshaping for high performance and efficiency

Traditional Dataflow

Hierarchical with limited data reuse

LPU Dataflow

Streamlined with maximum data reuse

Featured Technology

Multi-Chip Scalability

Internal network controller, Expandable Synchronization Link, to leverage ever-increasing size of AI models with high scalability. Computation-communication overlapping to minimize communication overhead.

Traditional Compute-communication Sequence

LPU : Compute-communication Overlapping