"Jan 13 (Reuters) - The U.S. government announced plans on Monday to impose new measures aimed at controlling access to American-designed artificial intelligence chips and technology globally.
The regulations focus on managing the distribution of U.S. AI chips essential for cutting-edge AI applications.
Key points regarding this U.S. initiative include:
- The new rule targets the export of graphics processing units (GPUs), specialized processors initially created to enhance graphic rendering. - GPUs, like those from U.S.-based industry leader Nvidia, renowned for their gaming capabilities, are now valued for their ability to handle multiple data strands simultaneously, crucial for training and implementing AI models. - For instance, OpenAI's ChatGPT relies on tens of thousands of GPUs for its operation. - The quantity of GPUs needed for an AI model relies on GPU sophistication, volume of data used, model size, and desired training duration. - The U.S. is broadening limitations on advanced GPUs crucial for constructing clusters to train high-end AI models, to regulate global AI accessibility. - The new rule's restrictions for most countries are computed based on chip performance, utilizing the Total Processing Performance (TPP) metric. Countries are capped at a total of 790 million TPP through 2027. - This cap, equivalent to nearly 50,000 H100 Nvidia GPUs, provides substantial computational power for groundbreaking research, managing AI firms, or supporting demanding AI applications globally. - Exceptions exist for specific companies like Amazon Web Services or Microsoft's Azure that meet special authorization standards. - Orders for small GPU quantities that do not exceed about 1,700 H100 chips fall below the cap limit and only necessitate government notification. - Lower-limit chip orders, such as those from universities, medical institutions, and research bodies, are streamlined to expedite low-risk U.S. chip shipments worldwide. - Eighteen countries, including Australia, Canada, Japan, and others, are exempt from GPU restrictions. - The U.S. is also regulating "model weights," parameters essential for training AI models, to ensure security standards. - These measures, as highlighted by Kaushik, are intended to cultivate advanced AI development in secure environments."