Google announced the launch of its new A3 supercomputer virtual machines at Google I/O 2023. The company explained that the new offerings are designed to cater to the needs of the resource-intensive Large Language Models (LLMs).
Google explained that its A3 supercomputer virtual machines (VMs) are powered by Nvidia’s H100 GPUs built on the Hopper architecture. Each machine will offer 2TB of RAM and 3.6 TB/s bisectional bandwidth across the GPUs through NVLink 4.0 and NVSwitch. The A3 virtual machines also integrate with fourth-generation Intel Xeon Scalable processors to manage administrative tasks.
“Google Cloud’s A3 VMs, powered by next-generation NVIDIA H100 GPUs, will accelerate training and serving of generative AI applications,” said Ian Buck, vice president of hyperscale and high performance computing at NVIDIA. “On the heels of Google Cloud’s recently launched G2 instances, we’re proud to continue our work with Google Cloud to help transform enterprises around the world with purpose-built AI infrastructure.”
According to Google, the new A3 supercomputers can deliver up to 26 exaFlops of power. It should help to reduce the time and cost required to train machine learning models. It’s the first production-level deployment of its GPU-to-GPU interface that lets users share data at 200 Gbps. Google claims that the data transfer between the CPUs should be 10 times faster compared to A2 VMs.
Google says that customers who are interested to use A3 virtual machines will need to sign up for a preview waitlist. The company plans to integrate the new cloud offerings with its software and cloud products.
It’s important to note that Google is not the only company involved in the development of supercomputer virtual machines. Last year, Microsoft unveiled a partnership with Nvidia to build their own Azure-powered AI supercomputers. IBM also plans to develop its own AI supercomputer design called Vela for government agencies.