Houmo.AI Claims To Have Created The First Chinese HPC AI Chip With Computing & Memory In A 35W Package

wccftech.com:

Houmo.AI has revealed the new HaloDrive H30 auto System-on-Chip (SoC) — the first Chinese HPC AI chip.

Compared to the Journey 5 SoC, developed on the 16nm process technology and featuring 128 TOPS while consuming 35 W of power, the new HaloDrive H30 uses a 12nm process node. SRAM memory and a dedicated IPU while offering INT8 processing.  The Houmo.AI HaloDrive H30 SoC was developed for the smart car segment, with companies such as Journey and NVIDIA becoming close competitors in the marketplace. The SoC utilizes the first-generation IPU architecture which has been entirely developed by the manufacturer.

Houmo.AI developed the chip in-house that can offer an efficiency ratio of 7.3 TOPS per watt. In recent Resnet 50 model benchmark tests, the new H30 reached 8700 fos and 10300 fps in both Batch Sizes, equal to 1 and 8, respectively.

The company also revealed a proprietary software development kit (SDK) that is based on the chip called Houmo Avenue. This new SDK will support PyTorch, TensorFlow, and ONNX. Houmo Avenue will be compatible with a "CUDA front-end syntax" that will support the programming models SIMD and SIMT to achieve operating and development efficiency.

Two years ago, Houmo Intelligent was established. We decided to innovate the underlying architecture of integrated storage and computing to achieve the ultimate breakthrough in AI computing efficiency. The integrated storage and computing architecture combines storage and computing. Function fusion is a computing method closer to the human brain than the traditional architecture and has much higher computing efficiency than the traditional method. With the emergence of large models such as GPT, storage, and computing integrated chips are getting more and more