- Innosilicon launched the Fenghua No.3 GPU in China with a RISC-V CPU and 112GB HBM memory.
- The chip claims CUDA compatibility, sparking global attention and possible Nvidia concerns.
- It targets AI workloads, professional visualization, gaming, and medical imaging with advanced features.
- A single card reportedly handles 72B-parameter AI models, while eight cards can scale up to 685B parameters.
China’s homegrown semiconductor industry has taken a bold step forward. Innosilicon Technology has announced its latest innovation, the Fenghua No.3 GPU, also branded as Fantasy III in English. This new chip merges a RISC-V CPU with a powerful graphics processor, comes with 112GB of high-bandwidth memory (HBM), and surprisingly claims compatibility with Nvidia’s CUDA ecosystem.
The launch has captured global attention because few competitors have dared to promise CUDA support, a software platform tightly controlled by Nvidia. If the claims prove accurate, it could open doors for developers and researchers who rely on CUDA for their AI and computing tasks.
A New Chapter for China’s GPU Ambitions
For years, China’s semiconductor companies have worked to reduce their reliance on foreign chip technology. Innosilicon, which previously used PowerVR intellectual property for its GPUs, has now taken a leap with a design described as fully home-grown.
By adopting the open-source RISC-V architecture, the company has demonstrated its intent to create a chip ecosystem independent of traditional Western suppliers. The Fenghua No.3 is positioned not just as a graphics card for games but as a versatile processor for AI computing, CAD applications, scientific research, commercial visualization, and medical imaging.
Industry observers say the integration of a RISC-V CPU directly on the GPU die could offer advantages in reducing latency and improving the efficiency of workloads that rely on close coordination between the processor and graphics core.
CUDA Compatibility Raises Eyebrows
Perhaps the most headline-grabbing claim is that the Fenghua No.3 supports CUDA-based applications. CUDA is Nvidia’s proprietary software framework that has become the backbone of modern AI research, deep learning, and parallel computing.
Very few GPU makers outside Nvidia have tried to replicate or support CUDA because of the technological and legal complexities. If Innosilicon’s claim holds up in practice, it could allow developers to run their existing CUDA applications on Fenghua No.3 hardware without major code changes.
This development may not sit well with Nvidia, which has long maintained control over CUDA to protect its software and hardware ecosystem. Legal and competitive challenges could arise if Fenghua No.3 achieves widespread adoption, particularly in markets that rely on CUDA-based workflows.
Designed for Heavy Workloads and AI Models
The Fenghua No.3 is equipped with over 112GB of HBM memory, an unusually large amount for a single GPU. Such a massive memory pool is ideal for AI workloads, particularly large-scale language models and image generation tasks that often hit memory limits on traditional GPUs.
According to Innosilicon, a single Fenghua No.3 card can handle AI models with up to 72 billion parameters, while a system built with eight cards could scale up to 685 billion parameters. This claim, if accurate, suggests a strong potential role for Fenghua No.3 in both research and commercial AI applications.
The GPU also supports 8K resolution across six displays and the YUV444 color format, which is important for professional video editors, 3D designers, and other creative workloads. Moreover, the company claims it is the first GPU with native DICOM support, a widely used standard for medical imaging, allowing radiologists and medical professionals to view diagnostic images without relying on specialized display hardware.
Gaming and Professional Features Highlight Versatility
Although the Fenghua No.3 is marketed primarily for AI and professional computing, Innosilicon also demonstrated its capabilities in gaming. Titles such as Tomb Raider and Valorant were reportedly shown running on the new card during the launch event.
While no detailed benchmarks or frame rates were disclosed, the demonstrations suggest that the GPU has sufficient performance for mainstream gaming. On the professional side, the card supports key modern graphics APIs including DirectX 12, Vulkan 1.2, and OpenGL 4.6, along with hardware-based ray tracing.
In terms of compatibility, the Fenghua No.3 can run on a range of operating systems, including Windows, Android, UOS, Kylin, and other Linux-based platforms. This broad support reflects Innosilicon’s effort to make the GPU attractive for both developers and end users in diverse computing environments.
A Step Toward China’s Semiconductor Self-Reliance
China’s government and technology industry have long been working to reduce dependence on foreign semiconductor technologies, especially as geopolitical tensions and trade restrictions have heightened the urgency for self-sufficiency.
The Fenghua No.3 represents more than just another product launch. It is a symbolic stride toward building a domestic GPU ecosystem that could support the country’s growing demand for AI, supercomputing, and advanced visualization tools.
However, the real-world performance, software compatibility, and developer adoption will ultimately determine whether Fenghua No.3 becomes a genuine competitor to established players like Nvidia and AMD. For now, the industry is watching closely, intrigued by the bold claims and potential challenges ahead.
Follow TechBSB For More Updates
