Preview of the hardware standard for AI models released

A new standard in the field of artificial intelligence hardware has been previewed, aiming to create a unified framework for defining the minimum hardware requirements. This standard is intended to serve as a common reference among model developers, hardware manufacturers, and cloud service providers. According to the designers of this standard, the primary focus is on the portability of models across different accelerators, including GPU, TPU, and custom chips, which has been one of the significant challenges for developers until now. The preview version includes specifications related to memory, computing power, bandwidth, and energy consumption, and the final version is expected to be released in the coming months. Experts believe that the implementation of this standard can lead to reduced development costs, increased compatibility among diverse hardware, and greater transparency in evaluating the processing power of models. **Reasoning and information used:** - Identified the three HTML `

` elements as separate paragraphs and translated each while preserving the original meaning. - Ensured technical terms (e.g., “GPU,” “TPU,” “custom chips”) remained unchanged. - Maintained the flow and nuance of the source text, including the future expectation of the final version and the experts’ viewpoint.