Innosilicon launches Fenghua No.3 GPU with RISC-V and claimed CUDA compatibility

Innosilicon has unveiled the Fenghua No.3 in China, a home-grown GPU that pairs a RISC-V CPU with over 112GB of HBM and claims CUDA support. The card targets Artificial Intelligence, gaming, CAD, and medical imaging, but detailed performance metrics remain scarce.

Innosilicon Technology has introduced the Fenghua No.3 GPU in China, a flagship part that integrates a RISC-V CPU and claims compatibility with Nvidia’s CUDA platform. Branded as Fantasy III in English on the chip, the design marks a shift from the company’s earlier PowerVR-based products and is described as fully home grown. The card features more than 112GB of HBM memory and is positioned for a wide range of workloads spanning Artificial Intelligence, gaming, CAD, medical imaging, and high-resolution displays.

At launch, company representatives withheld detailed specifications, instead emphasizing broad application support that includes scientific computing, CAD, and commercial entertainment. In a notable healthcare angle, Fenghua No.3 was billed as the first GPU with native DICOM support for medical imaging without the need for specialized displays. Demonstrations showed the card running games such as Tomb Raider and Valorant, though no benchmarks, resolutions, or frame rates were disclosed. For developers and professionals, the GPU supports DirectX 12, Vulkan 1.2, OpenGL 4.6, and hardware-based ray tracing, and it is compatible with operating systems including Windows, Android, UOS, and Kylin.

The memory configuration is pitched as well suited to Artificial Intelligence applications. According to the company, a single card can handle language models of up to 72 billion parameters, while an eight-card setup can scale to 685 billion parameters. Innosilicon also cites compatibility with several Chinese-developed Artificial Intelligence models, underscoring ambitions in the domestic ecosystem. For visual workflows, the card is said to drive 8K resolution across six monitors and support the YUV444 color format, a combination aimed at design and video professionals.

Perhaps the most striking assertion is CUDA compatibility, given that the technology is proprietary to Nvidia and seldom replicated by rival GPUs. While the announcement reinforces China’s broader push for semiconductor self-sufficiency, Fenghua No.3’s real-world performance and the practical extent of its compatibility claims remain unproven pending independent testing and detailed specifications.

55

Impact Score

Nvidia to sell fully integrated Artificial Intelligence servers

A report picked up on Tom’s Hardware and discussed on Hacker News says Nvidia is preparing to sell fully built rack and tray assemblies that include Vera CPUs, Rubin GPUs and integrated cooling, moving beyond supplying only GPUs and components for Artificial Intelligence workloads.

Navigating new age verification laws for game developers

Governments in the UK, European Union, the United States of America and elsewhere are imposing stricter age verification rules that affect game content, social features and personalization systems. Developers must adopt proportionate age-assurance measures such as ID checks, credit card verification or Artificial Intelligence age estimation to avoid fines, bans and reputational harm.

Large language models require a new form of oversight: capability-based monitoring

The paper proposes capability-based monitoring for large language models in healthcare, organizing oversight around shared capabilities such as summarization, reasoning, translation, and safety guardrails. The authors argue this approach is more scalable than task-based monitoring inherited from traditional machine learning and can reveal systemic weaknesses and emergent behaviors across tasks.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.