China's first HPC AI chipset, the "HaloDrive H30" from Houmo Intelligent, offers 256 TOPS computation and 35W of power consumption.
The new HaloDrive H30 uses a 12nm process node in comparison to the Journey 5 SoC, which was designed on a 16nm process technology and has 128 TOPS while using 35 W of power. SRAM memory, a dedicated IPU, and INT8 processing are provided. The Houmo.AI HaloDrive H30 SoC was created for the smart car market, where firms like Journey and NVIDIA are now fierce competitors. The first-generation IPU architecture is used by the SoC. It was wholly created by the manufacturer.
A proprietary software development kit (SDK) based on the Houmo Avenue microprocessor was also made public by the business. PyTorch, TensorFlow, and ONNX will all be supported by this new SDK. In order to achieve operational and development efficiency, Houmo Avenue will be compatible with a "CUDA front-end syntax" that will support the programming models SIMD and SIMT.
According to Wu Qiang, founder and CEO of Houmo, Houmo Intelligent was founded two years ago with the goal of improving AI computing efficiency by innovating the architecture of integrated storage and computing. This approach combines storage and computing, utilizing function fusion, a more human-like computing method that is significantly more efficient than traditional methods. As large models like GPT have gained prominence, integrated storage and computing chips are receiving increasing attention. They are pleased to see a growing number of startup companies joining them in promoting the innovation and practical application of hard technology.
The new H30 will start shipping to initial partners for internal testing in the coming month, according to Houmo.AI co-founder and vice president of products Xin Xiaoxu. The business is also working on the next-generation SoC, the H50, which will debut some time next year and start mass production in 2025.