Intel has lifted the lids on second-generation Gaudi accelerators, which significantly reduces the time it takes to train large-scale AI models.
Announced at Intel Vision 2022 in Dallas, the Gaudi 2 processor is built on a 7nm process, has 24 integrated 100GbE RoCE ports and has the largest memory on any accelerator on the market (96GB HBM2e).
The new processors are a product of Israel-based Habana Labs Perceived by Intel Back in 2019, and designed for this Servers Dedicated to in-depth learning workload.
Training AI models
In recent years, several large-scale natural language processing (NLP) and computer vision models have emerged that offer better performance than previous entries in related fields.
The problem is that training for these multi-billion parameter models is very calculated, and therefore expensive and time consuming, which is a limiting factor in technology development.
However, with the new Gaudi 2 accelerator, Intel claims that it will significantly reduce both the cost and time it takes to develop advanced new AI models.
According to the COO in Elton Medina, Habana, the price-to-performance ratio is an important factor for customers, so it was given priority during the development of second-generation accelerators.
Benchmarks provided by Intel Visions indicate that Gaudi 2 processors deliver approximately 2x of training output over popular NLP and Vision workloads (BERT and Restnet-50) compared to Nvidia’s A100 GPU.
At the same time, it is said that the new Gaudi chips save almost 40% of the cost of both workload types compared to the A100 GPUs.
“The data center with Intel AI and Habana Accelerator is developing value for customers that are the perfect solution for servers dedicated to deep learning,” Medina said. “We believe this category is very important.”
Gaudi 2 processors are readily available to customers and may also be underpinned Cloud Further down examples from AWS, like With the previous generation.