Intel Unveils Gaudi 3 AI Accelerator, An All Round Nvidia's H100 Competitor

Intel's Gaudi 3 AI Accelerator is the latest breakthrough in AI technology, introduced at the Intel Vision 2024 event. This powerful chip is designed to revolutionize enterprise generative AI applications with its advanced technical specifications and performance capabilities.

The Gaudi 3 features a dual-block design with 48 MB of SRAM shared memory per tile, 16 Tensor Processing Cores (TPCs), and two Matrix Multiply Engines (MMEs). This configuration results in a total of 96 MB of SRAM and an impressive 12.8 TB/sec of aggregate bandwidth within the Gaudi 3 complex. Additionally, the accelerator is equipped with eight HBM2E memory stacks, providing a substantial 128 GB of capacity and 3.7 TB/sec of bandwidth, which is crucial for training AI systems and running complex software.

Intel's Gaudi 3 is set to be released in the third quarter of 2024 and is expected to be a game-changer in the AI chip market, challenging Nvidia's dominance. The chip's design focuses on enhancing the training of AI systems and the execution of the resulting software, which are key areas for improving AI performance.

The release of Gaudi 3 is part of Intel's comprehensive AI strategy for enterprises, which includes open, scalable systems that work across all AI segments. Intel's approach emphasizes open community-based software and industry-standard Ethernet networking, providing customers with more flexible system scaling options. This strategy aims to address the challenges businesses face in scaling AI initiatives and to promote the adoption of enterprise AI solutions.