China’s Innosilicon has released the Fenghua No.3 GPU, boasting RISC-V CPU integration and targeting AI, gaming, and professional applications. This new GPU aims to deliver strong performance with a focus on domestic Chinese technology.
Key Features of the Fenghua No.3 GPU
- RISC-V Architecture: A fully home-grown design, moving away from previous PowerVR IP reliance.
- High Memory Capacity: Equipped with over 112GB of HBM memory for demanding workloads.
- Broad Application Support: Targets AI, gaming, CAD, medical imaging, and 8K displays.
CUDA Compatibility Claims
Perhaps the most notable claim is CUDA compatibility, typically exclusive to NVIDIA GPUs. If validated, this could attract developers already invested in the NVIDIA ecosystem. However, the extent and nature of this compatibility remains to be seen, and could face legal challenges.
Performance and Application
- Gaming: Demonstrated running titles like Tomb Raider and Valorant, though concrete performance benchmarks are unavailable.
- Professional APIs: Supports DirectX 12, Vulkan 1.2, and OpenGL 4.6, along with hardware ray tracing.
- Operating System Compatibility: Compatible with Windows, Android, and Chinese operating systems like UOS and Kylin.
- AI Capabilities: Suited for AI applications, reportedly capable of handling large language models.
- Display Output: Claims support for 8K resolution across six monitors with YUV444 color format, catering to design and video professionals.
Target Applications
AI Workloads
The Fenghua No.3 GPU aims to support large language models, signifying Innosilicon’s ambition in the AI sector. Compatibility with Chinese-developed AI models is also emphasized.
Gaming and Entertainment
While demonstrated with popular games, specific performance metrics are still needed to assess its viability in the competitive gaming market.
Professional Use
Support for multiple 8K displays and professional APIs indicates a focus on CAD, medical imaging (with native DICOM support), and other visually intensive professional applications.