Newsnews

SambaNova Introduces New Chip Designed For 5 Trillion Parameter Models

sambanova-introduces-new-chip-designed-for-5-trillion-parameter-models

SambaNova, the AI hardware and software company, has unveiled its latest chip, the SN40L. This fourth-generation custom AI chip is specifically designed to handle large language models with up to 5 trillion parameters. The chip promises to significantly reduce the cost of running these models while delivering state-of-the-art accuracy.

Key Takeaway

SambaNova’s new SN40L chip aims to revolutionize the processing of large language models. By reducing the number of chips required and delivering optimal efficiency, SambaNova promises to significantly lower the cost of running these models while maintaining state-of-the-art accuracy. With its full-stack hardware and software solution, the company offers customers greater cost certainty and ownership of their AI assets.

Controlling Hardware for Maximum Efficiency

Unlike tech giants like Google, Microsoft, or Amazon, SambaNova may not be a household name, but it has been steadily building its full-stack AI solution that includes both hardware and software. By developing its own chips, SambaNova aims to have full control over the underlying hardware, enabling maximum efficiency in processing these resource-intensive language models.

Founder and CEO Rodrigo Liang emphasized the need to move away from the brute-force approach of using more and more chips to handle large language models. With the SN40L, SambaNova claims to achieve 30 times more efficiency, reducing the number of chips required to power these models. According to Liang, running the same trillion parameter model on competitor chips would require 50-200 chips, while SambaNova has managed to achieve the same results using just 8 chips.

Full-Stack Hardware and Software Solution

SambaNova offers a complete hardware and software solution for AI applications. The company provides a multi-year subscription that includes everything needed to build and train AI models based on private data. Liang assures customers that even though SambaNova assists in training the models, ownership remains with the customer once completed. This allows companies to retain full control over their data and models, turning them into valuable assets.

In addition to cost efficiency, the SN40L chip is fully backward compatible with previous generations, making it easy for existing SambaNova customers to upgrade their hardware while maintaining compatibility with their existing infrastructure.

Leave a Reply

Your email address will not be published. Required fields are marked *