Innosilicon introduces Fenghua No.3 GPU in China, but claims of CUDA compatibility will have Nvidia callings its lawyers

China’s Innosilicon launches Fenghua No.3 GPU with RISC-V CPU integration
Fenghua No.3 claims CUDA compatibility, 112GB HBM memory, and ray tracing
New GPU targets AI workloads, gaming, CAD, medical imaging, and 8K displays

Innosilicon Technology has introduced a new GPU in China which combines a RISC-V CPU with a graphics processor claiming compatibility with Nvidia’s CUDA platform.

The Fenghua No.3 (or Fantasy III as the English translation on the chip shows), the company’s latest flagship, comes equipped with over 112GB of HBM memory and targets a wide range of computing workloads, from AI to gaming.

Unlike its predecessors, which were based on PowerVR IP, the Fenghua No.3 draws on RISC-V architecture and is described by the company as a fully home-grown design.

CUDA compatibility

At launch, representatives avoided giving detailed specifications, but stressed the chip’s support for multiple applications, including scientific computing, CAD, and commercial entertainment.

Fenghua No.3 was also billed as the first GPU capable of native DICOM support for medical imaging without specialized displays.

CUDA compatibility is easily the most striking claim, as the technology is proprietary to Nvidia, and few rival GPUs have attempted to replicate compatibility.

If the claims are accurate, Fenghua No.3 could prove attractive to researchers and developers already relying on Nvidia’s ecosystem, but I can’t imagine Nvidia not taking some form of action to protect its platform and intellectual property.

In demonstrations, the card was shown running games such as Tomb Raider and Valorant, but no performance benchmarks, resolution details, or frame rates were provided.

For professionals, the card supports modern APIs including DirectX 12, Vulkan 1.2, and OpenGL 4.6, as well as hardware-based ray tracing. It’s compatible with Windows, Android, UOS, Kylin, and other operating systems.

The GPU’s memory configuration is well suited to AI applications and a single card is said to handle language models of up to 72B parameters, while systems with eight cards reportedly scale to 685B.

The company also claims compatibility with several Chinese-developed AI models, pointing to ambitions in the country’s domestic AI sector.

Innosilicon claimed support for 8K resolution across six monitors, along with the YUV444 color format, a feature aimed at design and video professionals.

China has been working toward self-sufficiency in semiconductor technology, and products like the Fenghua No.3 is another example of this, although the card’s real-world performance remains to be seen.