YH002 Mezzanine Module 96GB

YH YH002 is the flagship next-generation Mezzanine module on RISC-V architecture with dual systolic arrays. Features 96 GB HBM3e (1200 GB/s).

Each server contains 8 YH002 modules and 4 switch interconnect chips, delivering 12.8 TB/s aggregate bandwidth via full interconnection of 8 TP16 cards.

This significantly exceeds the YH001 Server and is the key competitive advantage for distributed training with high-intensity data exchange.

Performance: 1024 TFLOPS FP32, 4096 TFLOPS FP/BF16, 8192 TOPS INT8. Blocked FP8 support.

Complete CUDA independence: native PyTorch and TensorFlow. 8U form factor.

Application scenarios

Large-scale LLM training (100B–500B+ parameters) with 12.8 TB/s interconnect.
Building AI super-clusters with TP16 topology, no CUDA.
High-load inference with maximum VRAM (768 GB/server).
Flagship AI data centers requiring complete independence from Western ecosystems.