China's latest GPU arrives with claims of CUDA compatibility and RT support — Fenghua No.3 also boasts 112GB+ of HBM memory for AI
While Innosilicon Technology's products may not be prominently featured on the list of the best graphics cards, the company has been hard at work developing its Fenghua (translated as Fantasy) series of graphics cards. As ITHome reported, Innosilicon recently unveiled the Fenghua No. 3, the company's latest flagship GPU. The company promises that its third GPU iteration is a significant advancement over its predecessors.
While previous Fenghua No.1 and Fenghua No.2 graphics cards were based on Imagination Technologies' PowerVR IP, the new Fenghua No.3 leverages the open-source RISC-V architecture instead. The graphics card reportedly borrows a page from OpenCore Institute's Nanhu V3 project.
The company representative didn't provide any more details on the Fenghua No.3 during the launch event, only that it features a home-grown design from the ground up. The Fenghua No.3 is also purportedly compatible with Nvidia's proprietary CUDA platform, which could open many doors for the graphics card if it holds true.
The Fenghua No.3 is designed for a bunch of different workloads, as Innosilicon describes it as an "all-function GPU" (translation). The company plans to deploy the graphics card in different sectors, including AI, scientific computing, CAD work, medical imaging, and gaming. Therefore, it's safe to assume there will be other variants of the Fenghua No.3.
The Fenghua No.3 is a jack of all trades
From a gaming perspective, the Fenghua No.3 claims support for the latest APIs, including DirectX 12, Vulkan 1.2, and OpenGL 4.6. The graphics card is also reportedly equipped to support ray tracing. The team demonstrated the Fenghua No.3 in titles such as Tomb Raider, Delta Force, and Valorant at the press conference, and reports claim that the gameplay was smooth. However, there was no available information on game settings, resolution, and actual frame rates, so take these claims with a grain of salt.
The Fenghua No.3 reportedly comes equipped with 112GB+ of HBM memory, making it an ideal product for AI. A single Fenghua No. 3 can handle 32B and 72B LLM models, while eight of them in unison work with 671B and 685B parameter models. Innosilicon claims unconditional support for the DeepSeek V3, R1, and V3.1 models, as well as the Qwen 2.5 and Qwen 3 model families.
Innosilicon also boasted that the Fenghua No.3 is China's first graphics card to support the YUV444 format, which offers the best color detail and fidelity—a feature particularly beneficial for users who perform extensive CAD industrial work or video editing. The manufacturer also highlighted the Fenghua No.3's support for 8K (7680 x 4320) displays. The graphics card can drive up to six 8K monitors at 30 Hz.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
The Fenghua No.3 is the world's first graphics card to offer native support for DICOM (Digital Imaging and Communication in Medicine). It enables the precise visualization of X-rays, MRIs, CT scans, and ultrasounds on standard monitors, eliminating the need for costly, specialized grayscale medical displays.
China's semiconductor industry is gradually improving. Although it is unlikely to rival that of the United States in the near future, it may not necessarily need to do so. China's primary goal is to achieve self-sufficiency in key areas. Announcements such as the Fenghua No.3 may seem insignificant individually. Collectively, they might amass into substantial progress, akin to accumulating grains of sand that eventually form a small beach.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button!

Zhiye Liu is a news editor, memory reviewer, and SSD tester at Tom’s Hardware. Although he loves everything that’s hardware, he has a soft spot for CPUs, GPUs, and RAM.
-
atomicWAR Reply
A fair amount of CUDA is proprietary already. The only open source parts I know of are the compiler, CV-CUDA library, CUDA Quantum development platform and some parts of the driver. The software stack itself is closed. So Nvidia already has a pretty tight grip on CUDA. I am not saying they can't lock it down further but I am not sure I'd call it open either as it stands now. Though I'd love to hear from someone that is more versed in CUDA though for sure.call101010 said:Easy , Nvidia ! Make "CUDA" License only and not open. problem solved. -
bit_user Thanks for the coverage, Zhiye Liu. Although we've seen a long series of announcements that haven't amounted to much, I do think that these efforts will eventually star bearing fruit. When that happens, it'll be good to have some historical context for those developments.Reply
I think that might be referring to this:The article said:The graphics card reportedly borrows a page from OpenCore Institute's Nanhu V3 project.
https://github.com/Siudya/Nanhu?tab=readme-ov-file#architecture
That appears to be a 4-way decode OoO RISC-V core with 160-entry RoB. I don't see anywhere it says how wide the vectors are. Of course, any or all of that could change in whatever incarnation they derive from it.
The main benefits of using RISC-V are ease of firmware development and available tooling. For embedded applications, it could also be useful to shift threads between CPU and GPU cores. However, in laptop, desktop, and server contexts, the downsides of using a general-purpose ISA in a GPU will tend to outweigh the upsides. That's why you don't see it done in GPUs by Nvidia, AMD, Intel, ARM, and Qualcomm. I think Imagination was dabbling in it, but I'm not sure if they ever got anything based on RISC-V to market.
Certainly, a GPU with 112 GB of HBM will be far too expensive for gaming or other desktop purposes.The article said:The company plans to deploy the graphics card in different sectors, including AI, scientific computing, CAD work, medical imaging, and gaming. Therefore, it's safe to assume there will be other variants of the Fenghua No.3.
Does anyone use YUV in contexts that don't involve video production or playback?The article said:Innosilicon also boasted that the Fenghua No.3 is China's first graphics card to support the YUV444 format, which offers the best color detail and fidelity—a feature particularly beneficial for users who perform extensive CAD industrial work or video editing. -
acadia11 Reply
I would do the exact opposite.call101010 said:Easy , Nvidia ! Make "CUDA" License only and not open. problem solved. -
IBM296 Bold claims but what is it equivalent to?? Like the Lisuan 7G106 announced in July was "claimed" to be roughly on par with RTX 4060.Reply
And I hope they show some 3rd party benchmarks when the GPU releases (helps in actually gauging performance). -
Amdlova Now it's wait to see if working 24hrs per day will last six monthsReply
Chinese tier hardware as no warranty -
-Fran- So, at a high level, this is the Larabee approach, but using RISC-V instead of X86. Got it.Reply
It works at least, but definitely won't reach the same performance or efficiency as the uArchs in nVidia's GPUs.
I know it's stating the obvious, but perhaps that simplification would help the discussion.
Regards. -
edzieba ReplyThe Fenghua No.3 is designed for a bunch of different workloads, as Innosilicon describes it as an "all-function GPU" (translation).
Probably "General Purpose GPU" (GPGPU) as the term was coined a few decades ago. -
gamerk316 Reply
It doesn't need to; it just needs to be "good enough" to act as an in-house replacement.-Fran- said:It works at least, but definitely won't reach the same performance or efficiency as the uArchs in nVidia's GPUs. -
araghur12 Reply
Who's gonna tell him where all PC hardware is made?Amdlova said:Now it's wait to see if working 24hrs per day will last six months
Chinese tier hardware as no warranty
Or the fact that this isn't going to retail on the western markets?