China's latest GPU arrives with claims of CUDA compatibility and RT support (www.tomshardware.com)

Innosilicon unveiled the Fenghua No.3, its new flagship GPU that shifts the company from PowerVR IP to a RISC‑V–based, home‑grown design reportedly inspired by OpenCore’s Nanhu V3. The card is billed as an “all‑function” GPU supporting DirectX 12, Vulkan 1.2, OpenGL 4.6, hardware ray tracing, YUV444 color, native DICOM imaging, and up to six 8K displays at 30 Hz. The most notable hardware claim is 112+ GB of HBM memory, with Innosilicon saying a single Fenghua No.3 can serve 32B–72B LLMs and an 8‑GPU cluster can handle ~671B–685B parameter models. The company also asserted compatibility with NVIDIA’s CUDA ecosystem and explicit model support (DeepSeek V3/R1/V3.1, Qwen 2.5/3), though launch materials omitted low‑level specs (clock, core counts, memory bandwidth, interconnects) and independent benchmarks. For the AI/ML community the card is potentially significant: very large HBM capacity and claimed multi‑GPU scaling target high‑parameter LLM workloads and medical/scientific visualization, and RISC‑V may simplify firmware/tooling. Crucial caveats remain—CUDA compatibility, efficiency versus established GPUs, software stack maturity, and inter‑GPU scaling are unverified and will determine real utility. If validated, Fenghua No.3 would be an important marker of China’s self‑sufficiency push in compute hardware; if not, it may be an ambitious but limited entrant until performance, software support, and third‑party benchmarks are published.