Intel launches Habana Labs Gaudi 3 AI accelerator

Intel today unveiled its next-generation AI solutions with the launch of the Xeon 6 and Habana Labs Gaudi 3 AI processors, in another bid to capture a significant share of the AI ​​chip market. The Xeon 6 is the last word in the server and supercomputer market, and is a core processor (CPU) for servers that challenges rival AMD. The Gaudi 3 graphics processor was developed in Israel by Habana Labs, an Israeli subsidiary of Intel. According to Intel, Gaudi offers a performance-to-price ratio 1.8 times higher than Nvidia’s H100 processor, works with language models and is estimated to be about a third of Nvidia’s new Blackwell platform.

Intel says the Gaudi 3 has been upgraded to offer twice the computing power and network bandwidth of the Gaudi 2. The acquisition of Caesarea-based Habana Labs was Intel’s big hope in AI. It was acquired in what some say was a hasty $2 billion deal about five years ago, and the company has undergone a number of changes since then. In fact, the Gaudí 3 is the swan song for the Israeli company, which until now had operated independently and in recent months dissolved as an independent company and merged completely with Intel. Habana founders Dudi Dahan and Ran Halutz left the company to join co-founder and serial chip entrepreneur Avigdor Willenz in founding a new company called Touch, Globes reported last month.

Intel has announced that Gaudi 3 will be integrated into Dell, Super Micro, and Lenovo servers.

This will be the last version that Gaudi produces in its current form as an independent brand, having failed to capture market share compared to competing graphics processors from Nvidia and, to some extent, AMD. The next version will combine parts of the Israeli processor’s own graphics accelerator with components previously developed in-house at Intel, and its Falcon Shores software.

The future chip is scheduled to launch next year as a processor focused on running AI models (inference), which according to the company, will outperform competitors at a lower cost and will practically compete with Nvidia’s graphics processor on servers. Intel is aware of the fact that the main difficulty in gaining market share is that software developers have adapted to the Nvidia development environment and therefore promises to provide tools to help developers move from current Nvidia solutions to Intel solutions, such as dedicated sites like GitHub and support platforms like the Habana Labs developer forum.

How much will the new chip cost?

According to Forbes, a server based on Gaudi 3 is expected to cost about $65,000, a third of a competing product from Nvidia. Analysts say the new chip is expected to generate $500 million for Intel by the end of the year. The company hopes to break through the glass ceiling and see a big boost in revenue from AI products this year, including its Lunar Lake processors that bring AI capabilities to computers. Gaudi has an advantage that Nvidia doesn’t: the ability to work with popular Ethernet communications standards. Intel is trying to capitalize on this advantage through alliances with Samsung, Google, and HPE. Nvidia, for its part, is trying to move the market to the Infiniband standard from its Israeli company Mellanox.

The Gaudi 3 processor is made by TSMC, which, while competing with Intel, is increasingly producing its own chipsets. Just a few weeks ago, Intel announced that its Lunar Lake AI chips will also be manufactured outside its own factories. The exception to this is the new Xeon, whose cores are manufactured at Intel’s factories in Ireland and many of its other components, such as input and output components, are manufactured at Intel’s factories in Kiryat Gat, Arizona.

This article was published in Globes, Israeli Business News – en.globes.co.il – on September 24, 2024.

© Copyright Globes Publisher Itonut (1983) Ltd., 2024.