ZhiKai 100 is a general-purpose GPU inference acceleration card based on self-developed chips, featuring 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 ZhiKai 100 graphics card supports multiple precision computations including FP32, FP16, and INT8. It delivers high-performance computing power, providing *** TFLOPS INT8 computing power per card. Paired with large-capacity HBM2e memory and high transmission bandwidth, its performance is comparable to NVIDIA's RTX 4090. It excels in AI inference, particularly in large model inference scenarios, and can be deployed in various configurations such as 1-card, 4-card, 8-card, or 16-card per system to meet diverse industry needs, offering excellent overall cost efficiency.

The ZhiKai 100 acceleration card can be widely applied in areas such as smart finance, intelligent recommendation, fast search, content moderation, AI generation, and intelligent customer service.

The ZhiKai 100 graphics card comes with a comprehensive and efficient software stack, supporting architectures like x86 and ARM. The software stack is compatible with mainstream GPU general computing models, provides C/C++-based programming interface extensions, supports mainstream deep learning programming frameworks, high-performance inference engines, and various acceleration libraries. It also includes debugging and optimization tools, facilitating the development and deployment of high-performance computing and AI applications.

Domestic GPU graphics card ZhiKai 100 product specifications (performance comparable to NVIDIA RTX 4090):

GPU Architecture: General-purpose GPU;

FP32 Computing Power: ** TFLOPS;

FP16 Computing Power: ** TFLOPS;

INT8 Computing Power: *** TFLOPS@INT8 (supports INT32, INT16 computation);

Memory: 32 GB HBM2E;

Interface: PCIe Gen4.0 x 16 lane;

Power Supply: 8-pin PCIe;

Power Consumption: Board-level power consumption 150W (TDP);

Cooling Method: Passive cooling;

Dimensions: Full-length, full-height, single-slot PCIe card;

Video Decoding: Supports 128 concurrent streams (1080P @30fps), supports HEVC/AVC/VP9/AVS2;

Image Decoding/Encoding: 2000/500 fps (JPEG);

ECC: Supported;

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

Virtualization: Supported;