At the Google I/O conference, Google spent more than an hour talking about its many advances in artificial intelligence, such as the new AI for its generative AI Bard. PaLM 2 Large Language Model (LLM). It’s the foundational pillar for adding AI capabilities across Google’s portfolio, including Google Maps, Google Photos and Gmail, among others.
In addition, Google also launched a supercomputer with outstanding performance – A3 Virtual Machines. It features 4th Gen Intel Xeon Scalable processors, 2TB of DDR5-4800 memory, and eight Nvidia H100 “Hopper” GPUs for 3.6 TBps of bisection bandwidth via NVLink 4.0 and NVSwitch.
The new A3 supercomputer is “dedicated to training and serving the most demanding AI models that power today’s generative AI and large-scale language model innovations,” Google said. According to reports, this supercomputer can provide 26 exaFlops of artificial intelligence performance.
According to Google, the A3 represents the first production-grade deployment of its GPU-to-GPU data interface, allowing data to be shared at 200Gbps while bypassing the host CPU. This interface, which Google calls an Infrastructure Processing Unit (IPU), can increase the available network bandwidth of an A3 virtual machine (VM) by as much as 10 times that of an A2 virtual machine.
“Google Cloud’s A3 virtual machines, powered by next-generation Nvidia H100 GPUs, will accelerate the training and serving of generative AI applications,” said Ian Barker, vice president of hyperscale and high-performance computing at NVIDIA. Following the G2 instance, we are proud to continue our partnership with Google Cloud to help transform businesses around the world with purpose-built AI infrastructure. “
If you want to try out the A3, you can apply to join the early access program through the Google A3 preview now. But it must be reminded here that filling out the form is just an application, and it does not guarantee that you will pass the review and be eligible to use A3.