Innosilicon, a Chinese GPU maker, has announced the Fantasy 3 GPU, which features a massive 112 GB VRAM for AI tasks while supporting DX12, HW-RT & CUDA.
Innosilicon Goes All Out With Massive 112 GB Fantasy 3 GPU, Dual-Slot Design For AI Tasks While Supporting CUDA, DX12 & HW-RT
Innosilicon has released the third upgrade to its Fantasy lineup with the Fantasy 3 GPU. The previous Fantasy 1 and Fantasy 2 GPUs were launched a couple of years ago for the Chinese domestic market, but with the Fantasy 3, it looks like Innosilicon is going all out, tackling multiple segments at once, such as AI, Gaming & Content Creation.
During a recent event, the company showcased its latest domestic offering, the Fantasy 3, which is a multi-purpose GPU solution that will be coupled with an RISC-V CPU and also offer NVIDIA CUDA-compatibility alongside support for DX12, HW-RT, Vulkan 1.2, and OpenGL 4.6. Based on the new OpenCore architecture, the GPU is designed for AI Training, Large-Scale Science Workloads for HOPC, gaming (Cloud Computing), and more.
The company isn’t unveiling a whole lot of details for the Fantasy 3 GPU besides its massive 112 GB VRAM. The GPU also supports YUV444 color, precisely for video editing and content creation. The chip will be compatible across Windows, Linux, and Android platforms. According to Innosilicon, the Fantasy 3 GPU will be able to run local 32B/72B LLMs, while 8-card solutions in servers will be able to run up to 671B / 586B LLMs such as DeepSeek R1, Qwen 2.5, and Qwen 3.
Besides the Fantasy 3 GPU, Innosilicon is also showcasing its next-gen DDR5 and MRDIMM DDR5 memory solutions for servers alongside a 120-channel PCIe Gen5/4 server switch chip. The company expects to launch its Fantasy 3 GPUs soon. The company also demoed a game with DX12 & HW-RT support running on a single Fantasy 3 GPU. Innosilicon also hinted at over 112 GB variants based on Fantasy 3 chips, but that would depend on the market demand.
Innosilicon’s IP portfolio includes a wide range of advanced high-speed interface subsystems. The key IP Trio for Next Gen AI, such as LPDDR6/5X Combo, GDDR7/6X, MR DDR5, UCIe Chiplet, UALINK, PCIe 6.0/5.0,112G SerDes solutions, and more.
With one of the industry’s most advanced memory interface solutions, Innosilicon empowers customers to overcome the memory wall and accelerate leadership along the high-bandwidth AI innovation curve. It has supported over 300 top-tier global customers, contributed to more than 10 billion SoCs, and delivered silicon-proven IP across advanced process nodes at leading foundry TSMC, ranging from 28nm down to 3nm technologies including 28nm, 22nm,16/12nm, 7nm (including N6), 5nm (including N4), and 3nm processes.
To tackle key memory bottlenecks in AI workloads, Innosilicon introduced its cutting-edge LPDDR6/5X Combo PHY + Controller IP, fully compatible with TSMC’s N6 and N3 process technologies. At the TSMC North America OIP Ecosystem Forum, the company also will present its latest paper, “Shaping the Future of Memory Interfaces: LPDDR6/5X Combo Subsystem for AI, Security, Automotive, and Beyond,” highlighting its leadership in next-generation interface solutions.
The Innosilicon dual-protocol LPDDR6/5X Combo achieves peak performance of 14.4Gbps at core voltage in LPDDR6 mode. This high-speed capability suffices for numerous application scenarios with stringent data transmission speed requirements.