Sources say Intel Gaudi3 is based on TSMC 5nm, World Core Electronic Assistance: AI performance surpasses Nvidia H100

CTOnews.com, December 18 (Xinhua)-- Intel officially unveiled the new Core Ultra mobile processor and the fifth-generation Xeon scalable processor at an event held on Friday, and demonstrated the Gaudi3 accelerator for deep learning and large-scale generation of AI models for the first time.

Intel CEO Pat Kissinger (Pat Gelsinger) said the Intel Gaudi 3 AI accelerator, which will be released next year, has been prototyped and packaged, and he has announced the integration of Falcon Shores GPU and Gaudi into one product by 2025.

Taiwan's "Industrial and Commercial Times" reported that Intel Gaudi3 will be built by TSMC 5nm process, and will be assisted by Shixin, and in the future Intel reasoning chip Goya will also be won by Shixin.

According to previous reports from CTOnews.com, Intel Gaudi3 bandwidth is 1.5 times that of the previous Gaudi 2 (7nm process), BFloat16 performance is tripled, network computing power is doubled, and the amount of computation is increased by 50%. HBM3e storage with the highest 128GB is expected, and AI learning and training performance surpasses Nvidia H100 and challenges H200.

Related readings:

"for the standard Nvidia H200, Intel predicts Gaudi 3 accelerator card: with 128GB HBM3e RAM"