Xindong ‘Fenghua 3’: A New Benchmark for Domestic GPUs with 112GB Memory and Applications in Medical and Industrial Fields

A significant breakthrough has occurred in the field of domestic GPUs! Xindong Technology launched the “Fenghua 3” full-featured GPU on September 22, 2025, which, thanks to several globally pioneering technologies and industry-level applications, has become a milestone product in the process of domestic computing power autonomy. The following is an in-depth analysis from the perspectives of technical breakthroughs, application scenarios, and ecological layout:

1. Technical Architecture: Full-Function Design and Integration of Domestic Core

1. Hardware Architecture InnovationThe Fenghua 3 adopts a fully domestic underlying design, achieving the deep integration of open-source RISC-V CPUs and CUDA-compatible GPUs for the first time. Its integrated “Nanhuh V3” RISC-V processor IP core (performance comparable to ARM Cortex-A76) serves as the main control unit, handling data scheduling, power management, and other tasks, significantly improving system collaboration efficiency. The GPU core supports FP32 78 TFLOPs computing power, meeting the high-intensity demands of large model training and scientific computing.

2. Memory and Bandwidth BreakthroughEquipped with 112GB HBM3e high-bandwidth memory (bandwidth 4TB/s), a single card can support a 72B parameter large model, and a single machine with eight cards can seamlessly drive the DeepSeek 671B/685B full-version large model, breaking through the memory capacity and multi-card transport bottlenecks of domestic GPUs, reducing the deployment cost of ultra-large-scale models by over 60%. This design completely resolves the issue of model compression accuracy loss caused by insufficient memory in traditional domestic GPUs.

3. Rendering and Display CapabilitiesAs the first domestic GPU to support hardware ray tracing and 8K rendering, the Fenghua 3 is compatible with mainstream graphics interfaces such as DirectX12 and Vulkan1.2, allowing smooth operation of AAA games like “Tomb Raider” and enhancing the performance of industrial software like SolidWorks to international mainstream levels (e.g., assembly rendering speed increased by 5 times). Its globally pioneering DICOM medical-grade grayscale display technology achieves high-precision restoration of X-ray and CT images on ordinary monitors (grayscale error < 0.5%), meeting the first-level standards for medical equipment.

2. Application Scenarios: Disruptive Innovations in Medical and Industrial Fields

1. Medical Scenario Implementation

• Imaging Diagnosis RevolutionBy natively supporting DICOM grayscale display, the Fenghua 3 allows grassroots hospitals to accurately present subtle lesions without the need to purchase expensive professional medical monitors. Zhuhai People’s Hospital has launched an AI-assisted diagnosis project based on this GPU, achieving a lung cancer screening accuracy rate of 98.7% and tripling diagnostic efficiency.
• Surgical Navigation UpgradeCombined with 3D reconstruction algorithms, it can generate real-time three-dimensional models of patient organs, assisting doctors in planning complex surgical paths for tumor removal, with an error control within 0.2mm.

Industrial Intelligence Breakthrough

• CAD Software Performance LeapIn industrial design software such as SolidWorks and UG, the Fenghua 3 uses hardware acceleration technology to reduce the rendering time of large assemblies from 2 hours to 15 minutes, increasing engineering drawing generation efficiency by 8 times, completely resolving the previous pain points of domestic GPUs being unable to run professional software smoothly.
• Digital Twin and SimulationSupporting YUV444 lossless video encoding and 6-screen 8K30 multi-display, it can construct high-precision digital twin models of industrial equipment, achieving production line energy consumption optimization (reducing by 12%) and fault prediction (accuracy rate 95%).

Financial and Government EmpowermentLeading domestic private equity funds have seamlessly migrated quantitative trading models to the Fenghua 3, achieving a 10-fold increase in strategy backtesting speed, while being compatible with both Windows and Linux operating systems, meeting the stringent requirements for stability and compliance in the financial industry. The Nanjing Government Cloud has built an AI assistant based on the Fenghua 3, achieving a 92% accuracy rate in intelligent interpretation of policy documents, with a response time for the 12345 hotline reduced to seconds.

3. Ecological Layout: Parallel Compatibility of Mainstream and Domestic Alternatives

1. Software Ecosystem Breadth

• AI Framework CompatibilityFully supports mainstream frameworks such as PyTorch, TensorFlow, and Triton, reducing CUDA code migration costs by 90%. A certain internet company has migrated its recommendation system model from NVIDIA A100 to Fenghua 3, maintaining over 95% performance.
• Graphics and Computing InterfacesCovers interfaces such as DirectX12, OpenGL4.6, and OpenCL2.2, adapting to operating systems like Tongxin UOS, Kirin, Windows, and Android, meeting full-scene needs from cloud to edge.

Open Source Ecosystem CollaborationThe integrated RISC-V processor is deeply linked with the Xiangshan open-source community, allowing developers to conduct customized development based on open IP. Xindong Technology has launched the “RISC-V+AI” developer kit in collaboration with universities, attracting over 2000 companies to participate in ecological co-construction.

Industry Cooperation DeepeningAt the launch event, contracts were signed with nearly ten ecological partners, covering fields such as intelligent computing centers, medical care, and education. For example, a partnership with Xi’an Aerospace Intelligence has launched the “Space-Based Computing Power Plan,” extending GPU computing power to satellite remote sensing processing, promoting the application of spatiotemporal large models in disaster warning.

4. Industry Significance: From Single Point Breakthrough to Full-Scene Adaptation of Domestic GPUs

The launch of the Fenghua 3 marks the transformation of domestic GPUs from “single-function” to “full-scene empowerment”:

• Improved Technical AutonomyThe fully domestic underlying design (including IP cores, architecture, and drivers) breaks away from dependence on overseas technology, ensuring supply chain security in the context of semiconductor equipment restrictions.
• Release of Application ValueThrough collaborative innovation in memory, computing power, and ecology, the Fenghua 3 enables domestic computing power to compete with international high-end GPUs (such as NVIDIA H100) for the first time, promoting the large-scale implementation of AI technology in the real economy.
• Activation of Industrial EcologyIts RISC-V integrated design provides new ideas for innovation in domestic chip architecture, expected to drive rapid development in domestic EDA tools, basic software, and other industrial chain segments.

Currently, the Fenghua 3 has entered mass production, with the first batch of orders covering medical imaging equipment manufacturers, industrial software developers, and other fields. With ecological optimization and cost reduction, this full-featured GPU is expected to achieve a market share breakthrough of 15% by 2026, becoming the core engine of domestic computing power autonomy.