Actual photos of 'Gaudi 3', the AI accelerator announced by Intel to compete with NVIDIA, are now available
Intel announced the AI accelerator ' Gaudi 3 ' in April 2024. The hardware news site ServeTheHome has published a report on Gaudi 3, which is said to have higher performance than the previous generation Gaudi 2 and NVIDIA's H100, after actually seeing it at the venue.
This is Intel Gaudi 3, the new 128GB HBM2e AI chip in the wild
Intel says that Gaudi 3 will outperform NVIDIA's AI-specialized GPU 'H100' in terms of both AI learning performance and AI inference performance, and has revealed that it plans to start mass production in the second half of 2024.
Intel announces AI accelerator 'Gaudi 3', which is faster and consumes less power than NVIDIA's H100 and even beats the H200 in some tests - GIGAZINE
Kennedy participated in the event 'Intel Vision 2024' where Gaudi 3 was announced, and released a movie showing him picking up the actual Gaudi 3 at the venue.
Intel Gaudi 3 AI Accelerator with 24x 200GbE #ai - YouTube
Gaudi was originally a chip developed by a company called Habana and released in the second half of 2019. Habana was acquired by Intel at the same time, and Intel announced that it would discontinue the 'Nervana' series of AI chips that it was developing at the same time, and showed its intention to focus on Habana's products.
Intel decides to discontinue development of AI chip 'Nervana' just two months after its official announcement - GIGAZINE
The previous generation Gaudi 2 , announced in 2020, is designed with a 7nm process, while Gaudi 3 is designed with a 5nm process. Compared with the previous generation, Gaudi 3 has a significant upgrade in computing power and bandwidth.
In the photo below, the left is Gaudi 3 and the right is Gaudi 2. Comparing the two, you can clearly see that Gaudi 3 has a larger silicon package.
The silicon package consists of two dies, each with 48MB of SRAM, 16 TPCs (tensor processing cores), and a media engine.
The actual Gaudi 3 silicon package looks like this.
A board equipped with eight Gaudi 3. The TDP of the OCP Accelerator Module (OAM) version of Gaudi 3 is 900W, which means that even if you ignore other components, the total power is about 7200W.
The bottom of the OAM version of Gaudi 3 is shown in the photo below.
Gaudi was originally conceived as a key concept from Habana, the concept of 'scaling using Ethernet'. As of 2024, network switch speeds are much faster. In the days of Gaudi 1, 32-port 100GbE was still quite high-end, limiting network bandwidth. The OAM version of Gaudi 3 uses 21 of the 24 lanes and connects to the other seven Gaudi 3s using 3x 200GbE. The remaining 3x 200GbE connections are fed to
A Supermicro X14 server equipped with eight Gaudi 3s was also unveiled at Intel Vision 2024. There are six OSFP connectors in the center of the back.
The Gaudi 3 will be available in both air-cooled and water-cooled versions, with the air-cooled version scheduled for release in the third quarter of 2024 (July to September) and the water-cooled version scheduled for release in the fourth quarter of 2024 (October to December).
Related Posts:
in Hardware, Posted by log1i_yk