Article illustration 1

In a move that could reshape GPU competition, Chinese semiconductor firm Innosilicon has launched the Fenghua No. 3, its latest flagship graphics processor. Departing from its PowerVR-based predecessors, this GPU adopts the open-source RISC-V architecture—a strategic shift that underscores China's accelerating push for technological independence. With claims of Nvidia CUDA compatibility and specialized features for AI, gaming, and medical imaging, the Fenghua No. 3 isn't just another incremental update; it's a statement of intent in a high-stakes industry dominated by Western players.

Breaking from Tradition: RISC-V at the Core

Unlike earlier Fenghua models reliant on Imagination Technologies' PowerVR IP, the No. 3 leverages RISC-V—an open-standard instruction set architecture gaining traction for its flexibility and lack of licensing fees. This aligns with China's broader agenda to reduce reliance on foreign IP amid ongoing trade restrictions. Sources suggest the design draws inspiration from the Nanhu V3 project, an open-source RISC-V initiative, though Innosilicon emphasizes a "home-grown" approach. For developers, this shift could mean new opportunities in custom firmware and tooling, but it also raises questions about performance efficiency. As one industry observer noted in the source comments:

"Using a general-purpose ISA like RISC-V in a GPU brings firmware advantages, but it historically sacrifices performance density compared to dedicated architectures from Nvidia or AMD. This won't dethrone market leaders overnight, but it doesn't need to—it just has to be 'good enough' for China's domestic needs."

Bold Claims: CUDA, AI, and Gaming Capabilities

The Fenghua No. 3's most audacious assertion is compatibility with Nvidia's proprietary CUDA platform—a cornerstone of AI and scientific computing. If verified, this could allow seamless integration with existing CUDA-optimized applications, a potential game-changer for Chinese data centers and researchers. However, skepticism abounds. Nvidia tightly controls CUDA's core components, and past attempts to replicate it have faced legal and technical hurdles. As Tom's Hardware readers highlighted, unconditional CUDA support seems improbable without Nvidia's cooperation, making real-world validation essential.

For AI workloads, the specs are staggering: 112GB+ of HBM memory enables a single GPU to handle 32B-72B parameter LLMs, while an eight-GPU cluster manages models up to 685B parameters. Innosilicon touts support for popular frameworks like DeepSeek and Qwen, positioning this as an affordable alternative to Nvidia's H100 in inference tasks. But the GPU's versatility extends beyond AI—it's branded an "all-function" (likely meaning general-purpose) design targeting:
- Gaming: Support for DirectX 12, Vulkan 1.2, OpenGL 4.6, and ray tracing, with demos in Tomb Raider and Valorant. Yet, absent details on resolution or frame rates leave performance claims unverified.
- Professional Use: As China's first GPU with YUV444 format support, it promises superior color accuracy for CAD and video editing, plus six 8K display outputs.
- Medical Imaging: Native DICOM compatibility could revolutionize healthcare by rendering high-fidelity X-rays and MRIs on standard monitors, eliminating costly specialized hardware.

The Bigger Picture: China's Semiconductor Ambitions

The Fenghua No. 3 exemplifies China's methodical climb toward semiconductor self-reliance. While it won't rival U.S. tech giants soon, its focus on strategic sectors—AI, healthcare, and industrial applications—reflects a targeted approach to reducing vulnerabilities. Each domestic GPU announcement, like last year's Lisuan 7G106, accumulates into tangible progress. Yet challenges persist: manufacturing scalability, software ecosystem maturity, and global trust. As one commenter wisely noted, "These efforts will eventually bear fruit. When that happens, it'll be good to have historical context."

For developers and tech leaders, the Fenghua series signals a future where GPU diversity could democratize access and spur innovation—but only if claims withstand scrutiny. As China's tech sands shift, the world watches for the first genuine ripple in Nvidia's moat.

Source: Tom's Hardware, original reporting by Zhiye Liu.