The Tianhe-150S accelerator card is a general-purpose GPU training and inference acceleration card based on self-developed chips. It features broad application coverage, high computational performance, low application cost, and strong deployment support. It is compatible with the CUDA ecosystem, enabling rapid adaptation and migration of ecosystem applications and new models with minimal cost and no need for redevelopment.

The Tianhe-150S chip adopts a general-purpose GPU architecture, 7nm process technology, and 2.5D CoWoS packaging technology. It supports various precision computations including FP32, FP16, and INT8. The Tianhe-150S delivers high-performance computing power, with a single card providing *** TFLOPS of INT8 computing power. Paired with 64 GB of large-capacity HBM2e memory and high transmission bandwidth, its performance is comparable to NVIDIA's A100. It excels in AI training and inference, particularly in large model training and inference scenarios. It can be deployed in various configurations such as single-card, 4-card, 8-card, or 16-card setups per machine, meeting diverse industry deployment needs with industry-leading overall cost efficiency.

The Tianhe-150S graphics card offers flexible programming capabilities, exceptional performance, and attractive cost-effectiveness. It is widely applicable in training, general computing, new algorithm research, and other scenarios. It serves various industries such as internet, smart finance, smart healthcare, computer vision, intelligent speech, natural language processing, and intelligent recommendation, empowering the AI-driven intelligent society.

DomesticGPUGraphics Card Tianhe150/150SProduct Specifications (Performance Comparable to NVIDIAA100):

GPU Architecture: General-Purpose GPU;

FP32 Computing Power: ** TFLOPS;

FP16 Computing Power: *** TFLOPS;

INT8 Computing Power: *** TFLOPS@INT8 (Supports INT32, INT16 Computation);

Memory: 64 GB HBM2E;

Interface: PCIe Gen4.0 x 16 lane;

Power Supply: 8-pin PCIe;

Power Consumption: Board-Level Power Consumption 350W (TDP);

Dimensions: Full-Length, Full-Height, Dual-Slot PCIe Card;

Cooling Method: Passive Cooling;

Packaging: 2.5D CoWoS Packaging;

Software Ecosystem: Compatible with mainstream general computing software frameworks, supports mainstream deep learning development frameworks;

Virtualization: Supported;