Mathews02
Ideal for intelligent computing centers, general-purpose computing, data processing, large model training/inference, recommendation systems, automatic speech recognition, speech synthesis, image segmentation/detection, database acceleration, education/research, and cloud computing scenarios.
Detailed information
Processor Architecture: GPGPU architecture
Computing Power: Target FP32 compute performance of 15 TFLOPS, with up to 30 TFLOPS FP32 per card
Memory: 64GB HBM2E high-bandwidth memory with 1.8TB/s bandwidth
Interconnect Interface: Proprietary VelocityLink-M high-speed interface enabling seamless multi-GPU scaling; 4-card configuration supports 65B parameter model inference, 8-card for 130B models
Software Stack: Self-developed NebulaFlow software stack compatible with major GPU ecosystems
Ideal for intelligent computing centers, general-purpose computing, data processing, large model training/inference, recommendation systems, automatic speech recognition, speech synthesis, image segmentation/detection, database acceleration, education/research, and cloud computing scenarios.
