Intel launches new Gaudi2 processors for the Chinese market: total memory bandwidth 2.4TB / s, used to train AI

Thanks to CTOnews.com netizens for the delivery of clues about the past. CTOnews.com July 11 news, Intel today launched a new Gaudi2 processor for the Chinese market, mainly used to accelerate AI training and reasoning.

CTOnews.com summarizes the parameter information of this processor:

24 programmable Tensor processor cores (TPCs)

21 100Gbps (RoCEv2) Ethernet interfac

96GB HBM2E memory capacity

Total memory bandwidth of 2.4TB per second

SRAM on 48MB chip

Integrated multimedia processing engine

The architecture of the Intel Gaudi2 processor focuses on efficient expansion to meet the needs of large-scale language models and generative AI models; each chip of the processor integrates 21 100Gbps Ethernet interfaces dedicated to interconnection.

In Stable Diffusion training, Intel Gaudi2 processors demonstrated a near-linear 99% scalability from 1 card to 64 cards, and a near-linear expansion of 95% from 256 accelerators to 384 accelerators.

The AI server NF5698G7, which Intel cooperates with Tide Information, is based on the Gaudi2 processor. NF5698G7 is a new generation of AI server innovatively developed for the generative AI market, which integrates eight Gaudi2 accelerator cards HL-225B.