Tech

China’s Fenghua No.3 GPU: CUDA, RT & 112GB+ HBM for AI

China's Fenghua No.3 GPU: CUDA, RT & 112GB+ HBM for AI - News Directory 3

Summary of⁤ the Fenghua No.3 Graphics Card

Here’s a breakdown of the⁢ key facts about the Innosilicon Fenghua No.3 graphics card, based on the provided text:

* New Architecture: It moves away from Imagination Technologies’ PowerVR IP and utilizes the open-source RISC-V architecture, drawing inspiration from OpenCore⁢ Institute’s⁤ Nanhu V3 project.
* Homegrown Design: Innosilicon states⁤ it’s a ‍wholly original design.
* CUDA Compatibility: ⁢ Potentially compatible with Nvidia’s CUDA ⁣platform, which could substantially broaden its applications.
* Versatile Workloads: Designed as ⁤an “all-function GPU” for⁤ a wide⁢ range of applications⁣ including:
* ‍AI
* ⁢Scientific Computing
* ‍CAD
* Medical Imaging
* Gaming
* API⁤ Support: Supports DirectX 12, Vulkan 1.2, and ⁢opengl 4.6, and is reportedly capable of ray tracing.
* Gaming Demo: Demonstrated running Tomb Raider, delta ⁢Force, and Valuing smoothly, but performance ‍details (resolution, ⁢settings, frame rates) were not provided.
* Large Memory Capacity: Equipped with 112GB+ of HBM memory, making it well-suited for AI tasks.
* AI Capabilities: ⁣Can ⁢handle large language models (LLMs) – 32B/72B with⁣ a single card, and 671B/685B with eight cards. Supports DeepSeek and Qwen model families.
* YUV444 Support: China’s first graphics card to support the YUV444 ⁢format for superior color ⁣detail.

Overall: The Fenghua No.3 appears to be a powerful and versatile graphics card⁣ aiming to compete⁣ in multiple sectors, particularly AI and high-performance computing, while also offering ⁤gaming capabilities. however, concrete performance data is currently lacking.