InnoSilicon Launches Fenghua No.3: A Breakthrough in Chinese GPU Development

InnoSilicon has introduced the Fenghua No.3 GPU, representing a major advancement in China’s graphics processing technology. Unlike its predecessors, which relied on Imagination Technologies’ PowerVR intellectual property, the Fenghua No.3 is built on an open-source RISC-V architecture. This new design is based on the OpenCore Institute’s Nanhu V3 project, signaling a strategic move towards greater independence and innovation in the Chinese semiconductor sector.

AI Acceleration and High-Bandwidth Memory

The Fenghua No.3 GPU is equipped with over 112 GB of high-bandwidth memory (HBM), positioning it as a strong contender in the AI accelerator market. According to InnoSilicon, a single Fenghua No.3 card can handle large-scale language models with 32 billion and 72 billion parameters. When eight cards are used together, the system can support models with up to 671 billion and 685 billion parameters, making it suitable for demanding AI and machine learning workloads.

Advanced Graphics and Display Capabilities

This GPU supports a wide range of graphics APIs, including DirectX 12, Vulkan 1.2, and OpenGL 4.6, ensuring compatibility with modern gaming and professional applications. The Fenghua No.3 also features hardware-accelerated ray tracing, a first for a Chinese-developed GPU. InnoSilicon has demonstrated the card running popular games such as Tomb Raider, Delta Force, and Valorant, though specific performance metrics and settings have not been disclosed.

For display output, the Fenghua No.3 can drive up to six 8K screens at 30 Hz and supports the YUV444 color format, enhancing color accuracy for applications in medical imaging, CAD, and scientific computing.

Virtualization and Software Ecosystem

The Fenghua No.3 features a vGPU architecture with hardware virtualization, enabling efficient resource sharing across multiple users or workloads. Notably, InnoSilicon claims that the GPU is CUDA compatible, which, if verified, could significantly broaden its software ecosystem. However, as CUDA is proprietary to NVIDIA, this compatibility claim warrants further independent validation.

Domestic Innovation and Industry Impact

Key highlights of the Fenghua No.3 include:

  • Over 112 GB of high-bandwidth memory for large-scale AI tasks
  • Support for 32B/72B parameter models on a single card, and up to 671B with eight cards
  • First Chinese GPU with hardware ray tracing support
  • vGPU design with hardware virtualization
  • Support for DirectX 12, Vulkan 1.2, OpenGL 4.6, and up to six 8K displays
  • Domestic RISC-V CPU design and comprehensive IP portfolio

China’s semiconductor industry continues to make significant progress, with a focus on reducing reliance on imported technology and building robust domestic supply chains for critical components. While the primary goal is to achieve self-sufficiency rather than directly competing with leading global firms, innovations like the Fenghua No.3 demonstrate the growing capabilities and ambitions of Chinese chipmakers in the GPU and AI accelerator markets.