Habana Labs, which is owned by Intel, has unveiled its second-generation Gaudi accelerators.
Intel has unveiled a new generation of Gaudi accelerators that could drastically shorten the time it takes to train large-scale AI models.
The Gaudi 2 processors, which were announced at Intel Vision 2022 in Dallas, are constructed on a 7nm process, have 24 integrated 100GbE RoCE ports, and have the most RAM of any accelerator on the market (96GB HBM2e).
The new CPUs are meant for servers specialised to deep learning applications and were developed by Israel-based Habana Labs, which was acquired by Intel in 2019.
AI model training
A variety of large-scale natural language processing (NLP) and computer vision models have appeared in recent years, giving performance significantly superior than prior entries in the fields.
The issue is that training these multi-billion parameter models is extremely computationally intensive, making it expensive and time-consuming, and hence a limiting factor in the technology’s progress.
According to Intel, the new Gaudi 2 accelerators will drastically reduce the cost and time required to construct powerful new AI models.
According to Eltan Medina, Habana’s COO, the price-to-performance ratio is a critical issue for clients, so it was prioritised during the creation of the second-generation accelerators.
In comparison to Nvidia’s A100 GPU, Gaudi 2 processors give nearly 2x the training throughput across popular NLP and vision workloads (BERT and Restnet-50), according to benchmarks presented at Intel Visions.
At the same time, the new Gaudi chips are expected to save around 40% on both task categories when compared to A100 GPUs.
“Intel is expanding AI and the value for data centre clients with Habana accelerators, which are the best solution for deep learning servers,” Medina said. “We believe this is an extremely important category.”
Gaudi 2 processors are now accessible to clients and, like the previous generation, are anticipated to underlie AWS cloud instances in the future.
With the best bare-metal hosting services available, you may remove the virtualization layer.