Nvidia has announced a flurry of AI-focused enterprise products, including details about a new silicon architecture, Hopper; its first datacentre GPU built with that architecture, the H100; a new superchip, Grace CPU; and plans to build what it claims would be the fastest AI supercomputer, Eos.
The company has been among the biggest beneficiaries of the AI boom of the past decade, with its GPUs turning out to be a perfect match for data-intensive deep learning methods. With the AI sector’s growing demand for data compute, Nvidia wants to provide more firepower.
The company stressed on Transformer, a type of machine learning system. This method has powered language models such as OpenAI’s GPT-3 to medical systems such as DeepMind’s AlphaFold. These models have increased exponentially in size over the last few years.
“Training these giant models still takes months,” said Nvidia Senior Director of Product Management Paresh Kharya said.
“So you fire a job and wait for one and half months to see what happens. A key challenge to reducing this time to train is that performance gains start to decline as you increase the number of GPUs in a data centre.”
The Hopper architecture will ameliorate these difficulties. The architecture specialises in accelerating training of Transformer models on H100 GPUs by six times against previous-generation chips. The fourth-generation Nivida NVlink can connect up to 256 H100 GPUs at nine times higher bandwidth than its previous generation.
“For the training of giant Transformer models, H100 will offer up to nine times higher performance, training in days what used to take weeks,” said Kharya.
The company, which also announced the new data centre CPU — the Grace CPU Superchip — at the annual GTC conference, said it consisted of two CPUs connected directly through a new low-latency NVLink-C2C. The chip is designed to serve large-scale HPC and AI applications. It has 144 Arm cores and 1TB/s memory bandwidth.
The company also teased its new AI supercomputer, Eos, and said it would be built using the Hopper architecture and contain 4,600 H100 GPUs for 18.4 exaflops of AI performance. Nvidia will use the system for internal research only and said it would be online in a few months.
